[
https://issues.apache.org/jira/browse/NIFI-2395?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15401999#comment-15401999
]
Joseph Witt commented on NIFI-2395:
-----------------------------------
Awesome! This is great news for two reasons.
1) It confirms that the issue was caused by underlying exceptions happening
which could lead to a livelock situation. This JIRA fixes the 'if bad things
happen in a certain critical section we will try up to X number of times then
move on'.
2) The underlying issue is related to another provenance issue we're seeing for
very large provenance repositories where queries in Lucene can cause the entire
JVM to die. There are some really awesome and well written Lucene blogs about
this that are helping us to find out what could be the cause. One such cause
could be closing things too early. So this is confirming a couple of
suspicions.
> PersistentProvenanceRepository Deadlocks caused by a blocked journal merge
> --------------------------------------------------------------------------
>
> Key: NIFI-2395
> URL: https://issues.apache.org/jira/browse/NIFI-2395
> Project: Apache NiFi
> Issue Type: Bug
> Components: Core Framework
> Affects Versions: 0.6.0, 0.7.0
> Reporter: Brian Davis
> Assignee: Joseph Witt
> Priority: Blocker
> Fix For: 1.0.0
>
>
> I have a nifi instance that I have been running for about a week and has
> deadlocked at least 3 times during this time. When I say deadlock the whole
> nifi instance stops doing any progress on flowfiles. I looked at the stack
> trace and there are a lot of threads stuck doing tasks in the
> PersistentProvenanceRepository. Looking at the code I think this is what is
> happening:
> There is a ReadWriteLock that all the reads are waiting for a write. The
> write is in the loop:
> {code}
> while (journalFileCount > journalCountThreshold || repoSize >
> sizeThreshold) {
> // if a shutdown happens while we are in this loop, kill
> the rollover thread and break
> if (this.closed.get()) {
> if (future != null) {
> future.cancel(true);
> }
> break;
> }
> if (repoSize > sizeThreshold) {
> logger.debug("Provenance Repository has exceeded its
> size threshold; will trigger purging of oldest events");
> purgeOldEvents();
> journalFileCount = getJournalCount();
> repoSize = getSize(getLogFiles(), 0L);
> continue;
> } else {
> // if we are constrained by the number of journal
> files rather than the size of the repo,
> // then we will just sleep a bit because another
> thread is already actively merging the journals,
> // due to the runnable that we scheduled above
> try {
> Thread.sleep(100L);
> } catch (final InterruptedException ie) {
> }
> }
> logger.debug("Provenance Repository is still behind.
> Keeping flow slowed down "
> + "to accommodate. Currently, there are {}
> journal files ({} bytes) and "
> + "threshold for blocking is {} ({} bytes)",
> journalFileCount, repoSize, journalCountThreshold, sizeThreshold);
> journalFileCount = getJournalCount();
> repoSize = getSize(getLogFiles(), 0L);
> }
> logger.info("Provenance Repository has now caught up with
> rolling over journal files. Current number of "
> + "journal files to be rolled over is {}",
> journalFileCount);
> }
> {code}
> My nifi is at the sleep indefinitely. The reason my nifi cannot move forward
> is because of the thread doing the merge is stopped. The thread doing the
> merge is at:
> {code}
> accepted = eventQueue.offer(new Tuple<>(record, blockIndex), 10,
> TimeUnit.MILLISECONDS);
> {code}
> so the queue is full.
> What I believe happened is that the callables created here:
> {code}
> final Callable<Object> callable = new
> Callable<Object>() {
> @Override
> public Object call() throws IOException {
> while (!eventQueue.isEmpty() ||
> !finishedAdding.get()) {
> final
> Tuple<StandardProvenanceEventRecord, Integer> tuple;
> try {
> tuple = eventQueue.poll(10,
> TimeUnit.MILLISECONDS);
> } catch (final InterruptedException
> ie) {
> continue;
> }
> if (tuple == null) {
> continue;
> }
> indexingAction.index(tuple.getKey(),
> indexWriter, tuple.getValue());
> }
> return null;
> }
> {code}
> finish before the offer adds its first event because I do not see any Index
> Provenance Events threads. My guess is the while loop condition is wrong and
> should be && instead of ||.
> I upped the thread count for the index creation from 1 to 3 to see if that
> helps. I can tell you if that helps later this week.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
