Merge remote-tracking branch 'refs/remotes/origin/master' into mas-docs2

src/leveled_log.erl

@@ -168,7 +168,13 @@
         {info, "Saved manifest file"}},
     {"PC019",
         {debug, "After ~s level ~w is ~w"}},
-    
+    {"PC020",
+        {warn, "Empty prompt deletions at ManifestSQN=~w"}},
+    {"PC021",
+        {info, "Prompting deletions at ManifestSQN=~w"}},
+    {"PC022",
+        {info, "Storing reference to deletions at ManifestSQN=~w"}},
+        
     {"I0001",
         {info, "Unexpected failure to fetch value for Key=~w SQN=~w "
                 ++ "with reason ~w"}},

src/leveled_pclerk.erl

@@ -91,12 +91,13 @@ handle_cast(prompt, State) ->
 handle_cast({push_work, Work}, State) ->
     {ManifestSQN, Deletions} = handle_work(Work, State),
     PDs = dict:store(ManifestSQN, Deletions, State#state.pending_deletions),
+    leveled_log:log("PC022", [ManifestSQN]),
     {noreply, State#state{pending_deletions = PDs}, ?MAX_TIMEOUT};
 handle_cast({prompt_deletions, ManifestSQN}, State) ->
-    Deletions = dict:fetch(ManifestSQN, State#state.pending_deletions),
+    {Deletions, UpdD} = return_deletions(ManifestSQN,
+                                            State#state.pending_deletions),
     ok = notify_deletions(Deletions, State#state.owner),
-    UpdDeletions = dict:erase(ManifestSQN, State#state.pending_deletions),
-    {noreply, State#state{pending_deletions = UpdDeletions}, ?MIN_TIMEOUT}.
+    {noreply, State#state{pending_deletions = UpdD}, ?MIN_TIMEOUT}.
 
 handle_info(timeout, State) ->
     request_work(State),
@@ -223,6 +224,18 @@ do_merge(KL1, KL2, SinkLevel, SinkB, RP, NewSQN, MaxSQN, Additions) ->
     end.
 
 
+return_deletions(ManifestSQN, PendingDeletionD) ->
+    % The returning of deletions had been seperated out as a failure to fetch
+    % here had caased crashes of the clerk.  The root cause of the failure to
+    % fetch was the same clerk being asked to do the same work twice - and this
+    % should be blocked now by the ongoing_work boolean in the Penciller
+    % LoopData
+    %
+    % So this is now allowed to crash again
+    PendingDeletions = dict:fetch(ManifestSQN, PendingDeletionD),
+    leveled_log:log("PC021", [ManifestSQN]),
+    {PendingDeletions, dict:erase(ManifestSQN, PendingDeletionD)}.
+
 %%%============================================================================
 %%% Test
 %%%============================================================================

src/leveled_penciller.erl

@@ -511,10 +511,19 @@ handle_cast({levelzero_complete, FN, StartKey, EndKey}, State) ->
                             manifest=UpdMan,
                             persisted_sqn=State#state.ledger_sqn}};
 handle_cast(work_for_clerk, State) ->
-    case State#state.levelzero_pending of
-        true ->
-            {noreply, State};
-        false ->
+    case {State#state.levelzero_pending, State#state.work_ongoing} of
+        {false, false} ->
+            % TODO - as part of supervision tree and retry work:
+            % Need to check for work_ongoing as well as levelzero_pending as
+            % there may be a race that could lead to the clerk doing the same
+            % thing twice.
+            %
+            % This has implications though if we auto-restart the pclerk in the
+            % future, without altering this state - it may never be able to
+            % request work due to ongoing work that crashed the previous clerk
+            %
+            % Perhaps the pclerk should not be restarted because of this, and
+            % the failure should ripple up
             {WL, WC} = leveled_pmanifest:check_for_work(State#state.manifest,
                                                         ?LEVEL_SCALEFACTOR),
             case WC of
@@ -534,7 +543,9 @@ handle_cast(work_for_clerk, State) ->
                                                     {TL, State#state.manifest}),
                     {noreply,
                         State#state{work_backlog=false, work_ongoing=true}}
-            end
+            end;
+        _ ->
+            {noreply, State}
     end.