GH-4107 fix unclosed iteration fedx #4108
Conversation
…y finished and does not need the results anymore
|
|
||
| /** | ||
| * Optional implementation to close this task on a best effort basis. | ||
| */ | ||
| default void close() { | ||
|
|
||
| } |
There was a problem hiding this comment.
I've added a new method to the interface to close a task. Closing a task is similar to cancelling a task except that if a closed task is run then it just becomes a no-op instead of throwing an exception like it does if it is cancelled.
|
|
||
| @Override | ||
| public void handleClose() throws QueryEvaluationException { | ||
| try { | ||
| for (ParallelTask<T> task : tasks) { | ||
| task.close(); | ||
| } | ||
| } finally { | ||
| super.handleClose(); | ||
| } | ||
| } |
There was a problem hiding this comment.
We now try to close all the tasks, since it could be that we have been closed before all the tasks were completed (or even started). There is obviously no guarantee that started tasks can be stopped, so it's all best effort.
…was already cancelled or done
|
|
||
| @Override | ||
| public void handleClose() throws QueryEvaluationException { | ||
| try { | ||
| super.handleClose(); | ||
| } finally { | ||
| // signal the phaser to close (if currently being blocked) | ||
| phaser.forceTermination(); | ||
| } | ||
| } | ||
|
|
There was a problem hiding this comment.
This was the initial issue. The union() method would deadlock because the phaser was never terminated if the close() method had been called before all the tasks had completed.
There are three other classes that use the same approach with phasers and all those have this exact code.
| if (ex instanceof InterruptedException) { | ||
| Thread.currentThread().interrupt(); | ||
| } |
There was a problem hiding this comment.
I've tried to improve interrupt handling. Other code in RDF4J will typically re-interrupt the current thread if it needs to swallow or wrap an InterruptedException. This means that other code down the line is still able to abort their current execution, which was what the interrupt was initially signalling.
hmottestad
force-pushed
the
GH-4107-unclosed-iteration-fedx
branch
from
b4a1b12
to
03a22b9
Compare
GitHub issue resolved: #4107
Briefly describe the changes proposed in this PR:
PR Author Checklist (see the contributor guidelines for more details):
mvn process-resourcesto format from the command line)