umgeher/leveled
0
Fork 0
Code Issues Pull requests Releases Activity Actions
leveled/src/leveled_pmem.erl

452 lines
16 KiB
Erlang
Raw Normal View History

Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
%% -------- PENCILLER MEMORY ---------
%%
%% Module that provides functions for maintaining the L0 memory of the
%% Penciller.
%%
%% It is desirable that the L0Mem can efficiently handle the push of new trees
%% whilst maintaining the capability to quickly snapshot the memory for clones
%% of the Penciller.
%%
Minor quibbles Move legacy CDB code used only in unit tests into test area. Fix column width in pmem and comment out the unused case statement (in healthy tests) from the penciller test code
2016-11-03 16:46:25 +00:00
%% ETS tables are not used due to complications with managing their mutability,
%% as the database is snapshotted.
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
%%
%% An attempt was made to merge all trees into a single tree on push (in a
%% spawned process), but this proved to have an expensive impact as the tree
%% got larger.
%%
%% This approach is to keep a list of trees which have been received in the
%% order which they were received. There is then a fixed-size array of hashes
%% used to either point lookups at the right tree in the list, or inform the
%% requestor it is not present avoiding any lookups.
%%
L0 cache size counter improved This is now accurate save for hash collisions. It may now be a small under-estimate, whereas previously it could be a large over-estimate.
2016-11-03 16:22:51 +00:00
%% The trade-off taken with the approach is that the size of the L0Cache is
Update comments to reflect changes
2017-01-06 10:09:15 +00:00
%% uncertain. The Size count is incremented based on the inbound size and so
%% does not necessarily reflect the size once the lists are merged (reflecting
%% rotating objects)
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
-module(leveled_pmem).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
-include("include/leveled.hrl").
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
-export([
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
prepare_for_index/2,
Mas i389 rebuildledger (#390) * Protect penciller from empty ledger cache updates which may occur when loading the ledger from the journal, after the ledger has been cleared. * Score caching and randomisation The test allkeydelta_journal_multicompact can occasionally fail when a compaction doesn't happen, but then does the next loop. Suspect this is as a result of score caching, randomisation of key grabs for scoring, plus jitter on size boundaries. Modified test for predictability. Plus formatting changes * Avoid small batches Avoid small batches due to large SQN gaps * Rationalise tests Two tests overlaps with the new, much broader, replace_everything/1 test. Ported over any remaining checks of interest and dropped two tests.
2023-01-18 11:44:02 +00:00
add_to_cache/5,
Handle gen_server:cast slowness There was some unpredictable performance in tests, that was related to the amount of time it took the sft gen_server to accept a cast whihc passed the levelzero_cache. The response time looked to be broadly proportional to the size of the cache - so it appeared to be an issue with passing the large object to the process queue. To avoid this, the penciller now instructs the SFT gen_server to callback to the server for each tree in the cache in turn as it is building the list from the cache. Each of these requests should be reltaively short, and the processing in-between should space out the requests so the Pencille ris not blocked from answering queries when pompting a L0 write.
2016-10-31 01:33:33 +00:00
to_list/2,
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
check_levelzero/3,
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
check_levelzero/4,
Re-introduce ETS Index Add ETS Index back in to avoid having to check each skip list in turn. Also this helps keep a lower skip list size.
2016-12-11 05:23:24 +00:00
merge_trees/4,
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
add_to_index/3,
Re-introduce ETS Index Add ETS Index back in to avoid having to check each skip list in turn. Also this helps keep a lower skip list size.
2016-12-11 05:23:24 +00:00
new_index/0,
Handle L0 cache being full A test thta will cause leveled to crash due to a low cache size being set - but protect against this (as well as the general scenario of the cache being full). There could be a potential case where a L0 file present (post pending) without work backlog being set. In this case we want to roll the level zero to memory, but don't accept the cache update if the L0 cache is already full.
2019-01-14 12:27:51 +00:00
check_index/2,
cache_full/1
Initial support for OTP 26 (#395) * Initial support for OTP 26 * Extend timeout in test
2023-03-14 16:27:08 +00:00
]).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
Align cache with default Stop L0 from growing too large i.e. 127 * 4K is more than 10 times bigger than the default.
2019-02-25 23:35:12 +00:00
-define(MAX_CACHE_LINES, 31). % Must be less than 128
Mas d31 nhskv16sst (#428) * Add performance/profiling test Add test to perf_SUITE to do performance tests and also profile different activities in leveled. This can then be used to highlight functions with unexpectedly high execution times, and prove the impact of changes. Switch between riak_ctperf and riak_fullperf to change from standard test (with profile option) to full-scale performance test * Change shape of default perfTest * Refactor SST Compare and contrast profile for guess, before and after refactor: pre ``` lists:map_1/2 313370 2.33 32379 [ 0.10] lists:foldl_1/3 956590 4.81 66992 [ 0.07] leveled_sst:'-expand_list_by_pointer/5-fun-0-'/4 925020 6.13 85318 [ 0.09] erlang:binary_to_term/1 3881 8.55 119012 [ 30.67] erlang:'++'/2 974322 11.55 160724 [ 0.16] lists:member/2 4000180 15.00 208697 [ 0.05] leveled_sst:find_pos/4 4029220 21.01 292347 [ 0.07] leveled_sst:member_check/2 4000000 21.17 294601 [ 0.07] -------------------------------------------------- -------- ------- ------- [----------] Total: 16894665 100.00% 1391759 [ 0.08] ``` post ``` lists:map_1/2 63800 0.79 6795 [ 0.11] erlang:term_to_binary/1 15726 0.81 6950 [ 0.44] lists:keyfind/3 180967 0.92 7884 [ 0.04] erlang:spawn_link/3 15717 1.08 9327 [ 0.59] leveled_sst:'-read_slots/5-fun-1-'/8 31270 1.15 9895 [ 0.32] gen:do_call/4 7881 1.31 11243 [ 1.43] leveled_penciller:find_nextkey/8 180936 2.01 17293 [ 0.10] prim_file:pread_nif/3 15717 3.89 33437 [ 2.13] leveled_sst:find_pos/4 4028940 17.85 153554 [ 0.04] erlang:binary_to_term/1 15717 51.97 447048 [ 28.44] -------------------------------------------------- ------- ------- ------ [----------] Total: 6704100 100.00% 860233 [ 0.13] ``` * Update leveled_penciller.erl * Mas d31 nhskv16sstpcl (#426) Performance updates to leveled: - Refactoring of pointer expansion when fetching from leveled_sst files to avoid expensive list concatenation. - Refactoring of leveled_ebloom to make more flexible, reduce code, and improve check time. - Refactoring of querying within leveled_sst to reduce the number of blocks that need to be de-serialised per query. - Refactoring of the leveled_penciller's query key comparator, to make use of maps and simplify the filtering. - General speed-up of frequently called functions.
2024-01-22 21:22:54 +00:00
-type index_array() :: list(array:array())|[]|none.
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
Fix unexported types file:location not exported?
2018-06-04 10:57:37 +01:00
-export_type([index_array/0]).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
%%%============================================================================
%%% API
%%%============================================================================
Handle L0 cache being full A test thta will cause leveled to crash due to a low cache size being set - but protect against this (as well as the general scenario of the cache being full). There could be a potential case where a L0 file present (post pending) without work backlog being set. In this case we want to roll the level zero to memory, but don't accept the cache update if the L0 cache is already full.
2019-01-14 12:27:51 +00:00
-spec cache_full(list()) -> boolean().
%% @doc
Align cache with default Stop L0 from growing too large i.e. 127 * 4K is more than 10 times bigger than the default.
2019-02-25 23:35:12 +00:00
%% If there are already 31 entries in the cache then the cache is full
Handle L0 cache being full A test thta will cause leveled to crash due to a low cache size being set - but protect against this (as well as the general scenario of the cache being full). There could be a potential case where a L0 file present (post pending) without work backlog being set. In this case we want to roll the level zero to memory, but don't accept the cache update if the L0 cache is already full.
2019-01-14 12:27:51 +00:00
cache_full(L0Cache) ->
Align cache with default Stop L0 from growing too large i.e. 127 * 4K is more than 10 times bigger than the default.
2019-02-25 23:35:12 +00:00
length(L0Cache) == ?MAX_CACHE_LINES.
Handle L0 cache being full A test thta will cause leveled to crash due to a low cache size being set - but protect against this (as well as the general scenario of the cache being full). There could be a potential case where a L0 file present (post pending) without work backlog being set. In this case we want to roll the level zero to memory, but don't accept the cache update if the L0 cache is already full.
2019-01-14 12:27:51 +00:00
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
-spec prepare_for_index(
array:array(), leveled_codec:segment_hash()) -> array:array().
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
%% @doc
%% Add the hash of a key to the index. This is 'prepared' in the sense that
%% this index is not use until it is loaded into the main index.
%%
%% prepare_for_index is called from the Bookie when been added to the ledger
%% cache, but the index is not used until that ledger cache is in the
%% penciller L0 memory
Handle no-lookup hash in retain strategy
2017-01-05 22:42:59 +00:00
prepare_for_index(IndexArray, no_lookup) ->
IndexArray;
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
prepare_for_index(IndexArray, Hash) ->
{Slot, H0} = split_hash(Hash),
Bin = array:get(Slot, IndexArray),
Mas d31 nhskv16sst (#428) * Add performance/profiling test Add test to perf_SUITE to do performance tests and also profile different activities in leveled. This can then be used to highlight functions with unexpectedly high execution times, and prove the impact of changes. Switch between riak_ctperf and riak_fullperf to change from standard test (with profile option) to full-scale performance test * Change shape of default perfTest * Refactor SST Compare and contrast profile for guess, before and after refactor: pre ``` lists:map_1/2 313370 2.33 32379 [ 0.10] lists:foldl_1/3 956590 4.81 66992 [ 0.07] leveled_sst:'-expand_list_by_pointer/5-fun-0-'/4 925020 6.13 85318 [ 0.09] erlang:binary_to_term/1 3881 8.55 119012 [ 30.67] erlang:'++'/2 974322 11.55 160724 [ 0.16] lists:member/2 4000180 15.00 208697 [ 0.05] leveled_sst:find_pos/4 4029220 21.01 292347 [ 0.07] leveled_sst:member_check/2 4000000 21.17 294601 [ 0.07] -------------------------------------------------- -------- ------- ------- [----------] Total: 16894665 100.00% 1391759 [ 0.08] ``` post ``` lists:map_1/2 63800 0.79 6795 [ 0.11] erlang:term_to_binary/1 15726 0.81 6950 [ 0.44] lists:keyfind/3 180967 0.92 7884 [ 0.04] erlang:spawn_link/3 15717 1.08 9327 [ 0.59] leveled_sst:'-read_slots/5-fun-1-'/8 31270 1.15 9895 [ 0.32] gen:do_call/4 7881 1.31 11243 [ 1.43] leveled_penciller:find_nextkey/8 180936 2.01 17293 [ 0.10] prim_file:pread_nif/3 15717 3.89 33437 [ 2.13] leveled_sst:find_pos/4 4028940 17.85 153554 [ 0.04] erlang:binary_to_term/1 15717 51.97 447048 [ 28.44] -------------------------------------------------- ------- ------- ------ [----------] Total: 6704100 100.00% 860233 [ 0.13] ``` * Update leveled_penciller.erl * Mas d31 nhskv16sstpcl (#426) Performance updates to leveled: - Refactoring of pointer expansion when fetching from leveled_sst files to avoid expensive list concatenation. - Refactoring of leveled_ebloom to make more flexible, reduce code, and improve check time. - Refactoring of querying within leveled_sst to reduce the number of blocks that need to be de-serialised per query. - Refactoring of the leveled_penciller's query key comparator, to make use of maps and simplify the filtering. - General speed-up of frequently called functions.
2024-01-22 21:22:54 +00:00
array:set(Slot, <<Bin/binary, H0:24/integer>>, IndexArray).
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
-spec add_to_index(array:array(), index_array(), integer()) -> index_array().
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
%% @doc
%% Expand the penciller's current index array with the details from a new
%% ledger cache tree sent from the Bookie. The tree will have a cache slot
%% which is the index of this ledger_cache in the list of the ledger_caches
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
add_to_index(LM1Array, L0Index, CacheSlot) when CacheSlot < 128 ->
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
[LM1Array|L0Index].
Re-introduce ETS Index Add ETS Index back in to avoid having to check each skip list in turn. Also this helps keep a lower skip list size.
2016-12-11 05:23:24 +00:00
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
-spec new_index() -> array:array().
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
%% @doc
%% Create a new index array
Re-introduce ETS Index Add ETS Index back in to avoid having to check each skip list in turn. Also this helps keep a lower skip list size.
2016-12-11 05:23:24 +00:00
new_index() ->
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
array:new([{size, 256}, {default, <<>>}]).
Re-introduce ETS Index Add ETS Index back in to avoid having to check each skip list in turn. Also this helps keep a lower skip list size.
2016-12-11 05:23:24 +00:00
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
-spec check_index(leveled_codec:segment_hash(), index_array())
-> list(non_neg_integer()).
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
%% @doc
%% return a list of positions in the list of cache arrays that may contain the
%% key associated with the hash being checked
Re-introduce ETS Index Add ETS Index back in to avoid having to check each skip list in turn. Also this helps keep a lower skip list size.
2016-12-11 05:23:24 +00:00
check_index(Hash, L0Index) ->
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
{Slot, H0} = split_hash(Hash),
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
{_L, Positions} =
lists:foldl(
fun(A, {SlotC, PosList}) ->
B = array:get(Slot, A),
case find_pos(B, H0) of
true -> {SlotC + 1, [SlotC|PosList]};
false -> {SlotC + 1, PosList}
end
end,
{1, []},
L0Index),
lists:reverse(Positions).
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
Mas i389 rebuildledger (#390) * Protect penciller from empty ledger cache updates which may occur when loading the ledger from the journal, after the ledger has been cleared. * Score caching and randomisation The test allkeydelta_journal_multicompact can occasionally fail when a compaction doesn't happen, but then does the next loop. Suspect this is as a result of score caching, randomisation of key grabs for scoring, plus jitter on size boundaries. Modified test for predictability. Plus formatting changes * Avoid small batches Avoid small batches due to large SQN gaps * Rationalise tests Two tests overlaps with the new, much broader, replace_everything/1 test. Ported over any remaining checks of interest and dropped two tests.
2023-01-18 11:44:02 +00:00
-spec add_to_cache(
integer(),
{tuple(), integer(), integer()},
integer(),
list(),
boolean()) -> {integer(), integer(), list()}|empty_push.
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
%% @doc
%% The penciller's cache is a list of leveled_trees, this adds a new tree to
%% that cache, providing an update to the approximate size of the cache and
%% the Ledger's SQN.
Mas i389 rebuildledger (#390) * Protect penciller from empty ledger cache updates which may occur when loading the ledger from the journal, after the ledger has been cleared. * Score caching and randomisation The test allkeydelta_journal_multicompact can occasionally fail when a compaction doesn't happen, but then does the next loop. Suspect this is as a result of score caching, randomisation of key grabs for scoring, plus jitter on size boundaries. Modified test for predictability. Plus formatting changes * Avoid small batches Avoid small batches due to large SQN gaps * Rationalise tests Two tests overlaps with the new, much broader, replace_everything/1 test. Ported over any remaining checks of interest and dropped two tests.
2023-01-18 11:44:02 +00:00
%% Updates to cache must set Writable to true if the update could generate a
%% Level 0 file - as this must guard against empty entries (which may lead to
%% an attempt to write an empty L0 file)
add_to_cache(L0Size, {LM1, MinSQN, MaxSQN}, LedgerSQN, TreeList, Writeable) ->
case {Writeable, leveled_tree:tsize(LM1)} of
{true, 0} ->
empty_push;
{_, LM1Size} ->
if
MinSQN >= LedgerSQN ->
{MaxSQN,
L0Size + LM1Size,
[LM1|TreeList]}
end
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
end.
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
-spec to_list(
integer(), fun((pos_integer()) -> leveled_tree:leveled_tree())) -> list().
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
%% @doc
%% The cache is a list of leveled_trees of length Slots. This will fetch
%% each tree in turn by slot ID and then produce a merged/sorted output of
%% Keys and Values (to load into a SST file).
%%
%% Each slot is requested in turn to avoid halting the penciller whilst it
%% does a large object copy of the whole cache.
Handle gen_server:cast slowness There was some unpredictable performance in tests, that was related to the amount of time it took the sft gen_server to accept a cast whihc passed the levelzero_cache. The response time looked to be broadly proportional to the size of the cache - so it appeared to be an issue with passing the large object to the process queue. To avoid this, the penciller now instructs the SFT gen_server to callback to the server for each tree in the cache in turn as it is building the list from the cache. Each of these requests should be reltaively short, and the processing in-between should space out the requests so the Pencille ris not blocked from answering queries when pompting a L0 write.
2016-10-31 01:33:33 +00:00
to_list(Slots, FetchFun) ->
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
SW = os:timestamp(),
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
SlotList = lists:seq(1, Slots),
Handle gen_server:cast slowness There was some unpredictable performance in tests, that was related to the amount of time it took the sft gen_server to accept a cast whihc passed the levelzero_cache. The response time looked to be broadly proportional to the size of the cache - so it appeared to be an issue with passing the large object to the process queue. To avoid this, the penciller now instructs the SFT gen_server to callback to the server for each tree in the cache in turn as it is building the list from the cache. Each of these requests should be reltaively short, and the processing in-between should space out the requests so the Pencille ris not blocked from answering queries when pompting a L0 write.
2016-10-31 01:33:33 +00:00
FullList = lists:foldl(fun(Slot, Acc) ->
Tree = FetchFun(Slot),
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
L = leveled_tree:to_list(Tree),
Handle gen_server:cast slowness There was some unpredictable performance in tests, that was related to the amount of time it took the sft gen_server to accept a cast whihc passed the levelzero_cache. The response time looked to be broadly proportional to the size of the cache - so it appeared to be an issue with passing the large object to the process queue. To avoid this, the penciller now instructs the SFT gen_server to callback to the server for each tree in the cache in turn as it is building the list from the cache. Each of these requests should be reltaively short, and the processing in-between should space out the requests so the Pencille ris not blocked from answering queries when pompting a L0 write.
2016-10-31 01:33:33 +00:00
lists:ukeymerge(1, Acc, L)
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
end,
[],
Handle gen_server:cast slowness There was some unpredictable performance in tests, that was related to the amount of time it took the sft gen_server to accept a cast whihc passed the levelzero_cache. The response time looked to be broadly proportional to the size of the cache - so it appeared to be an issue with passing the large object to the process queue. To avoid this, the penciller now instructs the SFT gen_server to callback to the server for each tree in the cache in turn as it is building the list from the cache. Each of these requests should be reltaively short, and the processing in-between should space out the requests so the Pencille ris not blocked from answering queries when pompting a L0 write.
2016-10-31 01:33:33 +00:00
SlotList),
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
leveled_log:log_timer(pm002, [length(FullList)], SW),
Handle gen_server:cast slowness There was some unpredictable performance in tests, that was related to the amount of time it took the sft gen_server to accept a cast whihc passed the levelzero_cache. The response time looked to be broadly proportional to the size of the cache - so it appeared to be an issue with passing the large object to the process queue. To avoid this, the penciller now instructs the SFT gen_server to callback to the server for each tree in the cache in turn as it is building the list from the cache. Each of these requests should be reltaively short, and the processing in-between should space out the requests so the Pencille ris not blocked from answering queries when pompting a L0 write.
2016-10-31 01:33:33 +00:00
FullList.
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
-spec check_levelzero(tuple(), list(integer()), list())
-> {boolean(), tuple|not_found}.
%% @doc
%% Check for the presence of a given Key in the Level Zero cache, with the
%% index array having been checked first for a list of potential positions
%% in the list of ledger caches - and then each potential ledger_cache being
%% checked (with the most recently received cache being checked first) until a
%% match is found.
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
check_levelzero(Key, PosList, TreeList) ->
Change hash algorithm for penciller Switch from magic hash to md5 - to hopefully remove the need for some of the artificial jumps required to get expected fall positive ratios. Also split the hash into two 16-bit integers. We assume that SegmentID (from the perspective of AAE merkle/tictac trees) will always be at least 16 bits. the idea is that hashes should be used in blooms and indexes such that some advantage can be gained from just knowing the segmentID - in particular when folding over all the keys in a bucket. Performance testing has been difficult so far - I think due to “cloud” mysteries.
2017-10-20 23:04:29 +01:00
check_levelzero(Key, leveled_codec:segment_hash(Key), PosList, TreeList).
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
Change hash algorithm for penciller Switch from magic hash to md5 - to hopefully remove the need for some of the artificial jumps required to get expected fall positive ratios. Also split the hash into two 16-bit integers. We assume that SegmentID (from the perspective of AAE merkle/tictac trees) will always be at least 16 bits. the idea is that hashes should be used in blooms and indexes such that some advantage can be gained from just knowing the segmentID - in particular when folding over all the keys in a bucket. Performance testing has been difficult so far - I think due to “cloud” mysteries.
2017-10-20 23:04:29 +01:00
-spec check_levelzero(tuple(), {integer(), integer()}, list(integer()), list())
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
-> {boolean(), tuple|not_found}.
%% @doc
%% Check for the presence of a given Key in the Level Zero cache, with the
%% index array having been checked first for a list of potential positions
%% in the list of ledger caches - and then each potential ledger_cache being
%% checked (with the most recently received cache being checked first) until a
%% match is found.
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
check_levelzero(_Key, _Hash, _PosList, []) ->
{false, not_found};
check_levelzero(_Key, _Hash, [], _TreeList) ->
Switch to using ets set as index of L0 cache Hope is that this will cause less garbage collection, and also will be slightly faster. Note that snapshots don't now get an index - they get the special index 'snap'. However, the SkipLists have bloom protection, and most snapshots are iterators not fetchers.
2016-12-10 14:15:35 +00:00
{false, not_found};
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
check_levelzero(Key, Hash, PosList, TreeList) ->
check_slotlist(Key, Hash, PosList, TreeList).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
Dialyzer helper Add specs and docs to external API
2017-05-30 15:36:50 +01:00
-spec merge_trees(tuple(), tuple(), list(tuple()), tuple()) -> list().
%% @doc
%% Return a list of keys and values across the level zero cache (and the
%% currently unmerged bookie's ledger cache) that are between StartKey
%% and EndKey (inclusive).
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
merge_trees(StartKey, EndKey, TreeList, LevelMinus1) ->
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
lists:foldl(
fun(Tree, Acc) ->
R = leveled_tree:match_range(StartKey, EndKey, Tree),
lists:ukeymerge(1, Acc, R) end,
[],
[LevelMinus1|TreeList]).
Initial Commit for SkipLists To speed check on alternate platform
2016-11-22 23:21:47 +00:00
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
%%%============================================================================
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
%%% Internal Functions
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
%%%============================================================================
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
find_pos(<<>>, _Hash) ->
false;
Mas d31 nhskv16sst (#428) * Add performance/profiling test Add test to perf_SUITE to do performance tests and also profile different activities in leveled. This can then be used to highlight functions with unexpectedly high execution times, and prove the impact of changes. Switch between riak_ctperf and riak_fullperf to change from standard test (with profile option) to full-scale performance test * Change shape of default perfTest * Refactor SST Compare and contrast profile for guess, before and after refactor: pre ``` lists:map_1/2 313370 2.33 32379 [ 0.10] lists:foldl_1/3 956590 4.81 66992 [ 0.07] leveled_sst:'-expand_list_by_pointer/5-fun-0-'/4 925020 6.13 85318 [ 0.09] erlang:binary_to_term/1 3881 8.55 119012 [ 30.67] erlang:'++'/2 974322 11.55 160724 [ 0.16] lists:member/2 4000180 15.00 208697 [ 0.05] leveled_sst:find_pos/4 4029220 21.01 292347 [ 0.07] leveled_sst:member_check/2 4000000 21.17 294601 [ 0.07] -------------------------------------------------- -------- ------- ------- [----------] Total: 16894665 100.00% 1391759 [ 0.08] ``` post ``` lists:map_1/2 63800 0.79 6795 [ 0.11] erlang:term_to_binary/1 15726 0.81 6950 [ 0.44] lists:keyfind/3 180967 0.92 7884 [ 0.04] erlang:spawn_link/3 15717 1.08 9327 [ 0.59] leveled_sst:'-read_slots/5-fun-1-'/8 31270 1.15 9895 [ 0.32] gen:do_call/4 7881 1.31 11243 [ 1.43] leveled_penciller:find_nextkey/8 180936 2.01 17293 [ 0.10] prim_file:pread_nif/3 15717 3.89 33437 [ 2.13] leveled_sst:find_pos/4 4028940 17.85 153554 [ 0.04] erlang:binary_to_term/1 15717 51.97 447048 [ 28.44] -------------------------------------------------- ------- ------- ------ [----------] Total: 6704100 100.00% 860233 [ 0.13] ``` * Update leveled_penciller.erl * Mas d31 nhskv16sstpcl (#426) Performance updates to leveled: - Refactoring of pointer expansion when fetching from leveled_sst files to avoid expensive list concatenation. - Refactoring of leveled_ebloom to make more flexible, reduce code, and improve check time. - Refactoring of querying within leveled_sst to reduce the number of blocks that need to be de-serialised per query. - Refactoring of the leveled_penciller's query key comparator, to make use of maps and simplify the filtering. - General speed-up of frequently called functions.
2024-01-22 21:22:54 +00:00
find_pos(<<Hash:24/integer, _T/binary>>, Hash) ->
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
true;
Mas d31 nhskv16sst (#428) * Add performance/profiling test Add test to perf_SUITE to do performance tests and also profile different activities in leveled. This can then be used to highlight functions with unexpectedly high execution times, and prove the impact of changes. Switch between riak_ctperf and riak_fullperf to change from standard test (with profile option) to full-scale performance test * Change shape of default perfTest * Refactor SST Compare and contrast profile for guess, before and after refactor: pre ``` lists:map_1/2 313370 2.33 32379 [ 0.10] lists:foldl_1/3 956590 4.81 66992 [ 0.07] leveled_sst:'-expand_list_by_pointer/5-fun-0-'/4 925020 6.13 85318 [ 0.09] erlang:binary_to_term/1 3881 8.55 119012 [ 30.67] erlang:'++'/2 974322 11.55 160724 [ 0.16] lists:member/2 4000180 15.00 208697 [ 0.05] leveled_sst:find_pos/4 4029220 21.01 292347 [ 0.07] leveled_sst:member_check/2 4000000 21.17 294601 [ 0.07] -------------------------------------------------- -------- ------- ------- [----------] Total: 16894665 100.00% 1391759 [ 0.08] ``` post ``` lists:map_1/2 63800 0.79 6795 [ 0.11] erlang:term_to_binary/1 15726 0.81 6950 [ 0.44] lists:keyfind/3 180967 0.92 7884 [ 0.04] erlang:spawn_link/3 15717 1.08 9327 [ 0.59] leveled_sst:'-read_slots/5-fun-1-'/8 31270 1.15 9895 [ 0.32] gen:do_call/4 7881 1.31 11243 [ 1.43] leveled_penciller:find_nextkey/8 180936 2.01 17293 [ 0.10] prim_file:pread_nif/3 15717 3.89 33437 [ 2.13] leveled_sst:find_pos/4 4028940 17.85 153554 [ 0.04] erlang:binary_to_term/1 15717 51.97 447048 [ 28.44] -------------------------------------------------- ------- ------- ------ [----------] Total: 6704100 100.00% 860233 [ 0.13] ``` * Update leveled_penciller.erl * Mas d31 nhskv16sstpcl (#426) Performance updates to leveled: - Refactoring of pointer expansion when fetching from leveled_sst files to avoid expensive list concatenation. - Refactoring of leveled_ebloom to make more flexible, reduce code, and improve check time. - Refactoring of querying within leveled_sst to reduce the number of blocks that need to be de-serialised per query. - Refactoring of the leveled_penciller's query key comparator, to make use of maps and simplify the filtering. - General speed-up of frequently called functions.
2024-01-22 21:22:54 +00:00
find_pos(<<_Miss:24/integer, T/binary>>, Hash) ->
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
find_pos(T, Hash).
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
Change hash algorithm for penciller Switch from magic hash to md5 - to hopefully remove the need for some of the artificial jumps required to get expected fall positive ratios. Also split the hash into two 16-bit integers. We assume that SegmentID (from the perspective of AAE merkle/tictac trees) will always be at least 16 bits. the idea is that hashes should be used in blooms and indexes such that some advantage can be gained from just knowing the segmentID - in particular when folding over all the keys in a bucket. Performance testing has been difficult so far - I think due to “cloud” mysteries.
2017-10-20 23:04:29 +01:00
split_hash({SegmentID, ExtraHash}) ->
Slot = SegmentID band 255,
H0 = (SegmentID bsr 8) bor (ExtraHash bsl 8),
Mas d31 nhskv16sst (#428) * Add performance/profiling test Add test to perf_SUITE to do performance tests and also profile different activities in leveled. This can then be used to highlight functions with unexpectedly high execution times, and prove the impact of changes. Switch between riak_ctperf and riak_fullperf to change from standard test (with profile option) to full-scale performance test * Change shape of default perfTest * Refactor SST Compare and contrast profile for guess, before and after refactor: pre ``` lists:map_1/2 313370 2.33 32379 [ 0.10] lists:foldl_1/3 956590 4.81 66992 [ 0.07] leveled_sst:'-expand_list_by_pointer/5-fun-0-'/4 925020 6.13 85318 [ 0.09] erlang:binary_to_term/1 3881 8.55 119012 [ 30.67] erlang:'++'/2 974322 11.55 160724 [ 0.16] lists:member/2 4000180 15.00 208697 [ 0.05] leveled_sst:find_pos/4 4029220 21.01 292347 [ 0.07] leveled_sst:member_check/2 4000000 21.17 294601 [ 0.07] -------------------------------------------------- -------- ------- ------- [----------] Total: 16894665 100.00% 1391759 [ 0.08] ``` post ``` lists:map_1/2 63800 0.79 6795 [ 0.11] erlang:term_to_binary/1 15726 0.81 6950 [ 0.44] lists:keyfind/3 180967 0.92 7884 [ 0.04] erlang:spawn_link/3 15717 1.08 9327 [ 0.59] leveled_sst:'-read_slots/5-fun-1-'/8 31270 1.15 9895 [ 0.32] gen:do_call/4 7881 1.31 11243 [ 1.43] leveled_penciller:find_nextkey/8 180936 2.01 17293 [ 0.10] prim_file:pread_nif/3 15717 3.89 33437 [ 2.13] leveled_sst:find_pos/4 4028940 17.85 153554 [ 0.04] erlang:binary_to_term/1 15717 51.97 447048 [ 28.44] -------------------------------------------------- ------- ------- ------ [----------] Total: 6704100 100.00% 860233 [ 0.13] ``` * Update leveled_penciller.erl * Mas d31 nhskv16sstpcl (#426) Performance updates to leveled: - Refactoring of pointer expansion when fetching from leveled_sst files to avoid expensive list concatenation. - Refactoring of leveled_ebloom to make more flexible, reduce code, and improve check time. - Refactoring of querying within leveled_sst to reduce the number of blocks that need to be de-serialised per query. - Refactoring of the leveled_penciller's query key comparator, to make use of maps and simplify the filtering. - General speed-up of frequently called functions.
2024-01-22 21:22:54 +00:00
{Slot, H0 band 16#FFFFFF}.
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
check_slotlist(Key, _Hash, CheckList, TreeList) ->
Switch to using ets set as index of L0 cache Hope is that this will cause less garbage collection, and also will be slightly faster. Note that snapshots don't now get an index - they get the special index 'snap'. However, the SkipLists have bloom protection, and most snapshots are iterators not fetchers.
2016-12-10 14:15:35 +00:00
SlotCheckFun =
Add support for with_lookup head_only head_only mode cna be run with_lookup - but there is no L0 index created in this case. So the L0 index wasn't returning a potition list and the L0 cache wasn't being checked. Code now checks every position in the L0 cache, when a lookup is attempted in head_only mode.
2018-06-23 15:15:49 +01:00
fun(SlotToCheck, {Found, KV}) ->
case Found of
true ->
{Found, KV};
false ->
CheckTree = lists:nth(SlotToCheck, TreeList),
case leveled_tree:match(Key, CheckTree) of
none ->
Switch to using ets set as index of L0 cache Hope is that this will cause less garbage collection, and also will be slightly faster. Note that snapshots don't now get an index - they get the special index 'snap'. However, the SkipLists have bloom protection, and most snapshots are iterators not fetchers.
2016-12-10 14:15:35 +00:00
{Found, KV};
Add support for with_lookup head_only head_only mode cna be run with_lookup - but there is no L0 index created in this case. So the L0 index wasn't returning a potition list and the L0 cache wasn't being checked. Code now checks every position in the L0 cache, when a lookup is attempted in head_only mode.
2018-06-23 15:15:49 +01:00
{value, Value} ->
{true, {Key, Value}}
Switch to using ets set as index of L0 cache Hope is that this will cause less garbage collection, and also will be slightly faster. Note that snapshots don't now get an index - they get the special index 'snap'. However, the SkipLists have bloom protection, and most snapshots are iterators not fetchers.
2016-12-10 14:15:35 +00:00
end
Add support for with_lookup head_only head_only mode cna be run with_lookup - but there is no L0 index created in this case. So the L0 index wasn't returning a potition list and the L0 cache wasn't being checked. Code now checks every position in the L0 cache, when a lookup is attempted in head_only mode.
2018-06-23 15:15:49 +01:00
end
end,
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
lists:foldl(SlotCheckFun, {false, not_found}, CheckList).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
%%%============================================================================
%%% Test
%%%============================================================================
-ifdef(TEST).
Initial support for OTP 26 (#395) * Initial support for OTP 26 * Extend timeout in test
2023-03-14 16:27:08 +00:00
-include_lib("eunit/include/eunit.hrl").
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
generate_randomkeys_aslist(Seqn, Count, BucketRangeLow, BucketRangeHigh) ->
lists:ukeysort(1,
generate_randomkeys(Seqn,
Count,
[],
BucketRangeLow,
BucketRangeHigh)).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
generate_randomkeys(Seqn, Count, BucketRangeLow, BucketRangeHigh) ->
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
KVL = generate_randomkeys(Seqn,
Count,
[],
BucketRangeLow,
BucketRangeHigh),
Change so that type of cache is parameterised The Tree doesn't seem to be better than the skiplist. Get ready to make this switchable
2017-01-21 11:38:26 +00:00
leveled_tree:from_orderedlist(lists:ukeysort(1, KVL), ?CACHE_TYPE).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
generate_randomkeys(_Seqn, 0, Acc, _BucketLow, _BucketHigh) ->
Acc;
generate_randomkeys(Seqn, Count, Acc, BucketLow, BRange) ->
Remove use of string rather than straddle OTP version string functions were used in unit tetss only, and were replaceable with io_lib:format
2018-12-11 15:44:37 +00:00
BNumber =
lists:flatten(
Resolve coverage gaps logs should always at least have a log_level defined which is in the ?LOG_LEVELS - so the final case of `-is_active_level([] , _, _) -> false.` is unnecessary (even if a bad log level is passed in). correct Bucket and Key definition to hit expected eunit test cases in leveled_pmem
2018-12-11 22:25:37 +00:00
io_lib:format("~4..0B",
Remove use of string rather than straddle OTP version string functions were used in unit tetss only, and were replaceable with io_lib:format
2018-12-11 15:44:37 +00:00
[BucketLow + leveled_rand:uniform(BRange)])),
KNumber =
Resolve coverage gaps logs should always at least have a log_level defined which is in the ?LOG_LEVELS - so the final case of `-is_active_level([] , _, _) -> false.` is unnecessary (even if a bad log level is passed in). correct Bucket and Key definition to hit expected eunit test cases in leveled_pmem
2018-12-11 22:25:37 +00:00
lists:flatten(io_lib:format("~4..0B", [leveled_rand:uniform(1000)])),
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
{K, V} = {{o, "Bucket" ++ BNumber, "Key" ++ KNumber, null},
{Seqn, {active, infinity}, null}},
generate_randomkeys(Seqn + 1,
Count - 1,
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
[{K, V}|Acc],
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
BucketLow,
BRange).
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
compare_method_test() ->
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
R = lists:foldl(fun(_X, {LedgerSQN, L0Size, L0TreeList}) ->
Minor quibbles Move legacy CDB code used only in unit tests into test area. Fix column width in pmem and comment out the unused case statement (in healthy tests) from the penciller test code
2016-11-03 16:46:25 +00:00
LM1 = generate_randomkeys(LedgerSQN + 1,
2000, 1, 500),
Mas i389 rebuildledger (#390) * Protect penciller from empty ledger cache updates which may occur when loading the ledger from the journal, after the ledger has been cleared. * Score caching and randomisation The test allkeydelta_journal_multicompact can occasionally fail when a compaction doesn't happen, but then does the next loop. Suspect this is as a result of score caching, randomisation of key grabs for scoring, plus jitter on size boundaries. Modified test for predictability. Plus formatting changes * Avoid small batches Avoid small batches due to large SQN gaps * Rationalise tests Two tests overlaps with the new, much broader, replace_everything/1 test. Ported over any remaining checks of interest and dropped two tests.
2023-01-18 11:44:02 +00:00
add_to_cache(
L0Size,
{LM1, LedgerSQN + 1, LedgerSQN + 2000},
LedgerSQN,
L0TreeList,
true)
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
end,
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
{0, 0, []},
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
lists:seq(1, 16)),
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
{SQN, Size, TreeList} = R,
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
?assertMatch(32000, SQN),
L0 cache size counter improved This is now accurate save for hash collisions. It may now be a small under-estimate, whereas previously it could be a large over-estimate.
2016-11-03 16:22:51 +00:00
?assertMatch(true, Size =< 32000),
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
TestList = leveled_tree:to_list(generate_randomkeys(1, 2000, 1, 800)),
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
FindKeyFun =
fun(Key) ->
fun(Tree, {Found, KV}) ->
case Found of
true ->
{true, KV};
false ->
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
L0 = leveled_tree:match(Key, Tree),
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
case L0 of
none ->
{false, not_found};
{value, Value} ->
{true, {Key, Value}}
Replace ledger-side gb_trees Try to make minimal change to replace gb_trees with gb_tree API-like skiplists
2016-11-25 14:50:13 +00:00
end
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
end
end
end,
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
S0 = lists:foldl(fun({Key, _V}, Acc) ->
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
R0 = lists:foldl(FindKeyFun(Key),
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
{false, not_found},
TreeList),
[R0|Acc] end,
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
[],
TestList),
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
PosList = lists:seq(1, length(TreeList)),
Magic Hash - and no L0 Index Move to using the DJ Bernstein Magic Hash consistently, and trying to make sure we only hash once for each operation (as the hash is more expensive than phash2). The improved lookup time for missing keys should allow for the L0 index to be removed, and hence speed up the completion time for push_mem operations. It is expected there will be a second stage of creating a tinybloom as part of the SFT creation process, and then adding that tinybloom to the manifest. This will then reduce the message passing required for a GET not in the cache or higher levels
2016-12-11 01:02:56 +00:00
S1 = lists:foldl(fun({Key, _V}, Acc) ->
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
R0 = check_levelzero(Key, PosList, TreeList),
Switch to using ets set as index of L0 cache Hope is that this will cause less garbage collection, and also will be slightly faster. Note that snapshots don't now get an index - they get the special index 'snap'. However, the SkipLists have bloom protection, and most snapshots are iterators not fetchers.
2016-12-10 14:15:35 +00:00
[R0|Acc]
end,
[],
TestList),
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
?assertMatch(S0, S1),
StartKey = {o, "Bucket0100", null, null},
EndKey = {o, "Bucket0200", null, null},
SWa = os:timestamp(),
Handle gen_server:cast slowness There was some unpredictable performance in tests, that was related to the amount of time it took the sft gen_server to accept a cast whihc passed the levelzero_cache. The response time looked to be broadly proportional to the size of the cache - so it appeared to be an issue with passing the large object to the process queue. To avoid this, the penciller now instructs the SFT gen_server to callback to the server for each tree in the cache in turn as it is building the list from the cache. Each of these requests should be reltaively short, and the processing in-between should space out the requests so the Pencille ris not blocked from answering queries when pompting a L0 write.
2016-10-31 01:33:33 +00:00
FetchFun = fun(Slot) -> lists:nth(Slot, TreeList) end,
DumpList = to_list(length(TreeList), FetchFun),
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
Q0 = lists:foldl(fun({K, V}, Acc) ->
P = leveled_codec:endkey_passed(EndKey, K),
case {K, P} of
{K, false} when K >= StartKey ->
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
[{K, V}|Acc];
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
_ ->
Acc
end
end,
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
[],
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
DumpList),
Change so that type of cache is parameterised The Tree doesn't seem to be better than the skiplist. Get ready to make this switchable
2017-01-21 11:38:26 +00:00
Tree = leveled_tree:from_orderedlist(lists:ukeysort(1, Q0), ?CACHE_TYPE),
Switch the LM1 cache to be a tree Use a tree of lists not a skiplist
2017-01-20 16:36:20 +00:00
Sz0 = leveled_tree:tsize(Tree),
Replace ledger-side gb_trees Try to make minimal change to replace gb_trees with gb_tree API-like skiplists
2016-11-25 14:50:13 +00:00
io:format("Crude method took ~w microseconds resulting in tree of " ++
"size ~w~n",
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
[timer:now_diff(os:timestamp(), SWa), Sz0]),
SWb = os:timestamp(),
Change so that type of cache is parameterised The Tree doesn't seem to be better than the skiplist. Get ready to make this switchable
2017-01-21 11:38:26 +00:00
Q1 = merge_trees(StartKey, EndKey, TreeList, leveled_tree:empty(?CACHE_TYPE)),
Replace ledger-side gb_trees Try to make minimal change to replace gb_trees with gb_tree API-like skiplists
2016-11-25 14:50:13 +00:00
Sz1 = length(Q1),
io:format("Merge method took ~w microseconds resulting in tree of " ++
"size ~w~n",
Refactor of L0 memory Not yet integrated, but there is now unit-tested module for the new way of managing L0 memory cache in the Penciller. This mechansim is considerably more efficient than previous efforts and should allow for further simplification of the code.
2016-10-29 13:27:21 +01:00
[timer:now_diff(os:timestamp(), SWb), Sz1]),
?assertMatch(Sz0, Sz1).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
Test timeouts So that coverage testing will run.
2017-07-02 22:23:02 +01:00
with_index_test_() ->
% Otherwise this test may timeout when run with coverage enabled
{timeout, 60, fun with_index_test2/0}.
with_index_test2() ->
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
IndexPrepareFun =
fun({K, _V}, Acc) ->
Change hash algorithm for penciller Switch from magic hash to md5 - to hopefully remove the need for some of the artificial jumps required to get expected fall positive ratios. Also split the hash into two 16-bit integers. We assume that SegmentID (from the perspective of AAE merkle/tictac trees) will always be at least 16 bits. the idea is that hashes should be used in blooms and indexes such that some advantage can be gained from just knowing the segmentID - in particular when folding over all the keys in a bucket. Performance testing has been difficult so far - I think due to “cloud” mysteries.
2017-10-20 23:04:29 +01:00
H = leveled_codec:segment_hash(K),
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
prepare_for_index(Acc, H)
end,
LoadFun =
fun(_X, {{LedgerSQN, L0Size, L0TreeList}, L0Idx, SrcList}) ->
LM1 = generate_randomkeys_aslist(LedgerSQN + 1, 2000, 1, 500),
LM1Array = lists:foldl(IndexPrepareFun, new_index(), LM1),
Change so that type of cache is parameterised The Tree doesn't seem to be better than the skiplist. Get ready to make this switchable
2017-01-21 11:38:26 +00:00
LM1SL = leveled_tree:from_orderedlist(lists:ukeysort(1, LM1), ?CACHE_TYPE),
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
UpdL0Index = add_to_index(LM1Array, L0Idx, length(L0TreeList) + 1),
Mas i389 rebuildledger (#390) * Protect penciller from empty ledger cache updates which may occur when loading the ledger from the journal, after the ledger has been cleared. * Score caching and randomisation The test allkeydelta_journal_multicompact can occasionally fail when a compaction doesn't happen, but then does the next loop. Suspect this is as a result of score caching, randomisation of key grabs for scoring, plus jitter on size boundaries. Modified test for predictability. Plus formatting changes * Avoid small batches Avoid small batches due to large SQN gaps * Rationalise tests Two tests overlaps with the new, much broader, replace_everything/1 test. Ported over any remaining checks of interest and dropped two tests.
2023-01-18 11:44:02 +00:00
R = add_to_cache(
L0Size,
{LM1SL, LedgerSQN + 1, LedgerSQN + 2000},
LedgerSQN,
L0TreeList,
true),
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
{R, UpdL0Index, lists:ukeymerge(1, LM1, SrcList)}
end,
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
R0 = lists:foldl(LoadFun, {{0, 0, []}, [], []}, lists:seq(1, 16)),
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
{{SQN, Size, TreeList}, L0Index, SrcKVL} = R0,
?assertMatch(32000, SQN),
?assertMatch(true, Size =< 32000),
Ongoing work on skip lists Initial rough implementation with some timing tests
2016-11-24 20:16:41 +00:00
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
CheckFun =
fun({K, V}, {L0Idx, L0Cache}) ->
Change hash algorithm for penciller Switch from magic hash to md5 - to hopefully remove the need for some of the artificial jumps required to get expected fall positive ratios. Also split the hash into two 16-bit integers. We assume that SegmentID (from the perspective of AAE merkle/tictac trees) will always be at least 16 bits. the idea is that hashes should be used in blooms and indexes such that some advantage can be gained from just knowing the segmentID - in particular when folding over all the keys in a bucket. Performance testing has been difficult so far - I think due to “cloud” mysteries.
2017-10-20 23:04:29 +01:00
H = leveled_codec:segment_hash(K),
Switch to binary index in pmem Remove the ets index in pmem and use a binary index instead. This may be slower, but avoids the bulk upload to ets, and means that matches know of position (so only skiplists with a match need be tried). Also stops the discrepancy between snapshots and non-snapshots - as previously the snapshots were always slowed by not having access to the ETS table.
2017-01-05 21:58:33 +00:00
PosList = check_index(H, L0Idx),
?assertMatch({true, {K, V}},
check_slotlist(K, H, PosList, L0Cache)),
{L0Idx, L0Cache}
end,
_R1 = lists:foldl(CheckFun, {L0Index, TreeList}, SrcKVL).
Pneciller Memory Test The current penciller memory setup is inefficient. Is there an alternative which is still relatively simple and but more efficient?
2016-10-29 01:06:00 +01:00
Develop 3.1 d30update (#386) * Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
2022-12-18 20:18:03 +00:00
index_performance_test() ->
LM1 = generate_randomkeys_aslist(1, 2000, 1, 500),
LM2 = generate_randomkeys_aslist(2001, 2000, 1, 500),
HL1 = lists:map(fun({K, _V}) -> leveled_codec:segment_hash(K) end, LM1),
HL2 = lists:map(fun({K, _V}) -> leveled_codec:segment_hash(K) end, LM2),
SWP = os:timestamp(),
A1 =
lists:foldl(
fun(H, A) -> prepare_for_index(A, H) end,
new_index(),
HL1),
io:format(
user,
"~nPrepare single index takes ~w microsec~n",
[timer:now_diff(os:timestamp(), SWP)]),
SWL = os:timestamp(),
PMI1 =
lists:foldl(
fun(I, Idx) -> add_to_index(A1, Idx, I) end, [], lists:seq(1, 8)),
io:format(
user,
"Appending to array takes ~w microsec~n",
[timer:now_diff(os:timestamp(), SWL)]),
SWC1 = os:timestamp(),
R0 = lists:seq(1, 8),
lists:foreach(fun(H) -> ?assertMatch(R0, check_index(H, PMI1)) end, HL1),
io:format(
user,
"Checking 2000 matches in array at each level takes ~w microsec~n",
[timer:now_diff(os:timestamp(), SWC1)]),
SWC2 = os:timestamp(),
FPT =
lists:foldl(
fun(H, FPC) -> FPC + length(check_index(H, PMI1)) end,
0,
HL2),
io:format(
user,
"Checking 2000 misses in array at each level takes ~w microsec " ++
"with ~w false positives~n",
[timer:now_diff(os:timestamp(), SWC2), FPT]).
Add compatibility for old and new random / rand functions
2017-07-31 20:20:39 +02:00
-endif.
Reference in a new issue Copy permalink