[
https://issues.apache.org/jira/browse/CASSANDRA-19776?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17950858#comment-17950858
]
Branimir Lambov commented on CASSANDRA-19776:
---------------------------------------------
The {{selectAndReference}} call in {{estimatedPartitionCount}} was added
recently to fix a race that caused node failures when an sstable disappears
while it's being processed. For this (and any other use of
{{selectAndReference}}) to work correctly, it cannot drop any of the selected
sstables. Instead, the method tries to reselect the sstables, which is expected
to drop whatever the problem sstables were; this is the correct behaviour to
handle a case where the view changed between the instances of creating the list
and taking the references. This should only happen if the release of the
sstables (which is meant to coincide with removing them from the view) happens
in-between the two calls, and the next iteration should succeed. This is
expected to be more efficient than taking a lock on the view for the duration
of both methods.
It seems that something quite unexpected is happening here. Aren't sstables
released only when the transaction commits? Is it really true that sstables are
fully released at the beginning of the operation, but remain present in the
tracker for its duration? If so, this is a bigger issue that probably means we
can't really abort a compaction that includes such sstables.
> Spinning trying to capture readers
> ----------------------------------
>
> Key: CASSANDRA-19776
> URL: https://issues.apache.org/jira/browse/CASSANDRA-19776
> Project: Apache Cassandra
> Issue Type: Bug
> Components: Legacy/Core
> Reporter: Cameron Zemek
> Assignee: Stefan Miklosovic
> Priority: Normal
> Fix For: 4.0.x, 4.1.x, 5.0.x, 5.x
>
> Attachments: extract.log
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> On a handful of clusters we are noticing Spin locks occurring. I traced back
> all the calls to the EstimatedPartitionCount metric (eg.
> org.apache.cassandra.metrics:type=Table,keyspace=testks,scope=testcf,name=EstimatedPartitionCount)
> Using the following patched function:
> {code:java}
> public RefViewFragment selectAndReference(Function<View,
> Iterable<SSTableReader>> filter)
> {
> long failingSince = -1L;
> boolean first = true;
> while (true)
> {
> ViewFragment view = select(filter);
> Refs<SSTableReader> refs = Refs.tryRef(view.sstables);
> if (refs != null)
> return new RefViewFragment(view.sstables, view.memtables,
> refs);
> if (failingSince <= 0)
> {
> failingSince = System.nanoTime();
> }
> else if (System.nanoTime() - failingSince >
> TimeUnit.MILLISECONDS.toNanos(100))
> {
> List<SSTableReader> released = new ArrayList<>();
> for (SSTableReader reader : view.sstables)
> if (reader.selfRef().globalCount() == 0)
> released.add(reader);
> NoSpamLogger.log(logger, NoSpamLogger.Level.WARN, 1,
> TimeUnit.SECONDS,
> "Spinning trying to capture readers {},
> released: {}, ", view.sstables, released);
> if (first)
> {
> first = false;
> try {
> throw new RuntimeException("Spinning trying to
> capture readers");
> } catch (Exception e) {
> logger.warn("Spin lock stacktrace", e);
> }
> }
> failingSince = System.nanoTime();
> }
> }
> }
> {code}
> Digging into this code I found it will fail if any of the sstables are in
> released state (ie. reader.selfRef().globalCount() == 0).
> See the extract.log for an example of one of these spin lock occurrences.
> Sometimes these spin locks last over 5 minutes. Across the worst cluster with
> this issue, I ran a log processing script that everytime the 'Spinning trying
> to capture readers' was different to previous one it would output if the
> released tables were in Compacting state. Every single occurrence has it spin
> locking with released listing a sstable that is compacting.
> In the extract.log example its spin locking saying that nb-320533-big-Data.db
> has been released. But you can see prior to it spinning that sstable is
> involved in a compaction. The compaction completes at 01:03:36 and the
> spinning stops. nb-320533-big-Data.db is deleted at 01:03:49 along with the
> other 9 sstables involved in the compaction.
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
