[
https://issues.apache.org/jira/browse/NIFI-2395?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15395657#comment-15395657
]
Brian Davis commented on NIFI-2395:
-----------------------------------
My configuration for the provenance should be very close to the default. The
only difference is I am using 2 separate disks to do provenance and only
provenance. Here is the config:
# Provenance Repository Properties
nifi.provenance.repository.implementation=org.apache.nifi.provenance.PersistentProvenanceRepository
# Persistent Provenance repository Properties
nifi.provenance.repository.directory.prov1=/data/04/nifi/provenance_repository
nifi.provenance.repository.directory.prov2=/data/05/nifi/provenance_repository
nifi.provenance.repository.max.storage.time=24 hours
nifi.provenance.repository.max.storage.size=1 GB
nifi.provenance.repository.rollover.time=30 secs
nifi.provenance.repository.rollover.size=100 MB
nifi.provenance.repository.query.threads=2
nifi.provenance.repository.indexing.threads=1
nifi.provenance.repository.compress.on.rollover=true
nifi.provenance.repository.always.sync=false
nifi.provenance.repository.journal.count=16
# Comma-separated list of fields. Fields that are not indexed will not be
searchable. Valid fields are:
# EventType, FlowFileUUID, Filename, TransitURI, ProcessorID,
AlternateIdentifierURI, Relationship, Details
nifi.provenance.repository.indexed.fields=EventType, FlowFileUUID, Filename,
ProcessorID
# FlowFile Attributes that should be indexed and made searchable
nifi.provenance.repository.indexed.attributes=twitter.msg, language
# Large values for the shard size will result in more Java heap usage when
searching the Provenance Repository
# but should provide better performance
nifi.provenance.repository.index.shard.size=500 MB
# Indicates the maximum length that a FlowFile attribute can be when retrieving
a Provenance Event form
# the repository. If the length of any attribute exceeds this value, it will
be truncated when the event is retrieved.
nifi.provenance.repository.max.attribute.length=65536
# Volatile Provenance Repository Properties
nifi.provenance.repository.buffer.size=100000
Now that I have gone through these properties with a little more detail, I see
some places that need to change. :)
> 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
>
> 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)
