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.
Plugged the ne wpencille rmemory into the Penciller, and took advantage
of the increased speed to simplify the callbacks involved.
The outcome is much simpler code
Stop the penciller from writing an empty file, when shutting down and
the L0 Cache is empty.
Also parameter fiddle to see impact of the Penciller changes.
Removed o(100) lines of code by refactoring the Penciller to no longer
use ETS tables. The code is less confusing, and probably not an awful
lot slower.
The unit tests for the Penciller couldn't cope with the returned status
- and so would intermittently fail (after tightening the timeout on sft
check_ready.
There was a test that failed to close down a bookie and that caused some
issues. The issues are double-reoslved, the close down was tidied as
well as the forgotten close being added back in.
There is some generla tidy around in anticipation of TTL support.
The SFT shutdown process ahs become a series of casts to-and-from
between Penciller and SFT to stop the two processes syncronously making
requests on each other
The test confirming that deleting sft files wer eheld open whilst
snapshots were registered was actually broken. This test has now been
fixed, as well as the logic in registring snapshots which had used
ledger_sqn mistakenly rather than manifest_sqn.
The Penciller had two problems in previous commits:
- If it had a push_mem soon after a L0 file had been created, the
push_mem would stall waiting for the L0 file to complete - and this
count take 100-200ms
- The penciller's clerk favoured L0 work, but was lazy about asking for
other work in-between, so often the L1 layer was bursting over capacity
and the clerk was doing nothing but merging more L0 files in (with those
merges getting more and more expensive as they had to cover more and
more files)
There are some partial resolutions to this. There is now an aggressive
timeout when checking whther the L0 file is ready on a push_mem, and if
the timeout is breached the error is caught and a 'returned' message
goes back to the Bookie. the Bookie doesn't now empty its cache, it
carrie son filling it, but on some probability it will keep trying to
push_mem on future pushes. This increases Jitter around the expensive
operation and split out the L0 delay into defined chunks.
The penciller's clerk is now more aggressive in asking for work. There
is also some simplification of the relationship between clerk timeouts
and penciller back-pressure.
Also resolved is an issue of inconcistency between the loader and the on
startup (replaying the transaction log) and the standard push_mem
process. The loader was not correctly de-duplicating by adding first
(in order) to a tree before outputting the list from the tree.
Some thought will be given later as to whether non-L0 work can be safely
prioritised if the merge process still keeps getting behind.
The penciller had the concept of a manifest_lock - but it wasn't clear
what the purpose of it was.
The updating of the manifest has now been updated to reduce the code and
make the process cleaner and more obvious. Now the committed manifest
only covers non-L0 levels. A clerk can work concurrently on a manifest
change whilst the Penciller is accepting a new L0 file.
On startup the manifets is opened as well as any L0 file. There is a
possible race condition with killing process where there may be a L0
file which is merged but undeleted - and this is believed to be inert.
There is some outstanding work still. Currently the whole store is
paused if a push_mem is received by the Penciller, and the writing of a
L0 sft file has not been completed. The creation of a L0 file appears
to take about 300ms, so if the ledger_cache fills in this period a pause
will occurr (perhaps due to objects with lots of index entries). It
would be preferable to pause more elegantly in this situation. Perhaps
there should be a harsh timeout on the call to check the SFT complete,
and catching it should cause a refused response. The next PUT will then
wait, but a any queued GETs can progress.
To try and improve performance index entries had been removed from the
Ledger Cache, and a shadow list of the LedgerCache (in SQN order) was
kept to avoid gb_trees:to_list on push_mem.
This did not go well. The issue was that ets does not deal with
duplicate keys in the list when inserting (it will only insert one, but
it is not clear which one).
This has been reverted back out.
The ETS parameters have been changed to [set, private]. It is not used
as an iterator, and is no longer passed out of the process (the
memtable_copy is sent instead). This also avoids the tab2list function
being called.
The 2i work now has tests for removals as well as regex etc.
Some initial refactoring work has also been tried - to try and take some
tasks of the critical path of push_mem. The primary change has been to
avoid putting index keys into the gb_tree, and building the KeyChanges
list in parallel to the gb_tree (now known as ObjectTree) within the
Ledger Cache.
Some initial experiments done as to changing the ETS table in the
Penciller now that it will now be used for iterating - but that has been
reverted for now.
Added basic support for 2i query. This involved some refactoring of the
test code to share functions between suites.
There is sill a need for a Part 2 as no tests currently cover removal of
index entries.
Some additional tests following previous refactoring for abstraction,
primarily to make manifest print safer an dprove co-existence of Riak
and non-Riak objects.
This test exposed two bugs:
- Yet another set of off-by-one errors (really stupidly scanning the
Manifest from Level 1 not Level 0)
- The return of an old issue related to scanning the journal on load
whereby we fail to go back to the previous file before the current SQN
Add iterator support, used initially only for retrieving bucket
statistics.
The iterator is supported by exporting a function, and when the function
is claled it will take a snapshot of the ledger, run the iterator and
hten close the snapshot.
This required a numbe rof underlying changes, in particular to get key
comparison to work as "expected". The code had previously misunderstood
how comparison worked between Erlang terms, and in particular did not
account for tuple length being compared first by size of the tuple (and
not just by each element in order).
Reviewing code to update comments revealed a weakness in the sequence of
events between penciller and clerk committing a manifest change wherby
an ill-timed crash could lead to files being deleted without the
manifest changing.
A different, and safer pattern now used between theses two actors.
An attempt to refactor out more complex code.
The Penciller clerk and Penciller have been re-shaped so that there
relationship is much simpler, and also to make sure that they shut down
much more neatly when the clerk is busy to avoid crashdumps in ct tests.
The CDB now has a binary_mode - so that we don't do binary_to_term twice
... although this may have made things slower ??!!? Perhaps the
is_binary check now required on read is an overhead. Perhaps it is some
other mystery.
There is now a more effiicient fetching of the size on pcl_load now as
well.
This exposed another off-by-one error on startup.
This commit also includes an unsafe change to reply early from a rolling
CDB file (with lots of objects writing the hash table can take too
long). This is bad, but will be resolved through a refactor of the
manifest writing: essentially we deferred writing of the manifest
update which was an unnecessary performance optimisation. If instead we
wait on this, the process is made substantially simpler, and it is safer
to perform the roll of the complete CDB journal asynchronously. If the
manifest update takes too long, an append-only log may be used instead.
Add some initial system tests. This highlighted issues:
- That files deleted by compaction would be left orphaned and not close,
and would not in fact delete (now deleted by closure only)
- There was an issue on stratup that the first few keys in each journal
would not be re-loaded into the ledger
Some initial work to get snapshots going.
Changes required, as need to snapshot through the Bookie to ensure that
there is no race between extracting the Bookie's in-memory view and the
Penciller's view if a push_to_mem has occurred inbetween.
A lot still outstanding, especially around Inker snapshots, and handling
timeouts
Two aspects of pushing to the penciller have been refactored:
1 - Allow the penciller to respond before the ETS table has been updated
to unlock the Bookie sooner.
2 - Change the way the copy of the memtable is stored to work more
effectively with snapshots wihtout locking the Penciller any further on
a snapshot or push request
CDB did many "bitty" reads/writes when scanning or writing hash tables -
change these to bult reads and writes to speed up.
CDB also added capabilities to fetch positions and get keys by position
to help with iclerk role.
Make scanning over a CDB file generic rather than specific to read-in of
active nursery log - open to be called as an external function to
support other scanning behaviour.
An attempt to get a first inker that can build a ledger from a manifest
as well as support simple get and put operations. Basic tests surround
the building of manifests only at this stage - more work required for
get and put.
Two issues looked at
- There shouldn't be a remainder after writing the L0 file, as this
could have overlapping sequence numbers which will be missed on restart
- There should be a safety-check to stop the Clerk from doing a fake
push too soon after a background L0 file ahs been written (as the fake
push would lock the ledger waiting for the L0 file write to finish)
Added support for startup and shutdown of a Ledger. As aprt of this
will now start tracking the highest sequence number. This also adds a
safety check on pcl_pushmem to make sure that only keys with a higher
sequenc enumber are being pushed in - and hence we can happily insert
into the in-memory view without checking the sequence number.
Allow for the clerk to merge continuously is no activity for the
penciller to prompt.
The penciller now must also correctly lock the manifest - to stop races
between the creation of ne wL0 files and the completion of work by the
clerk
Standardise on record definitions between modules to make easier - then
add functionality to pushing to penciller as bookie would do. Some
initial manual testing of this seems OK.
Add further descriptions of roles following name changes. Attempt to
simplify manifest management in the Penciller by assuming there is only
one Penciller's Clerk active - and so only one piece of work can be
ongoing
