[
https://issues.apache.org/jira/browse/CASSANDRA-19776?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17954519#comment-17954519
]
Branimir Lambov commented on CASSANDRA-19776:
---------------------------------------------
{quote}Update the affected metrics to avoid using {{{}selectAndReference{}}}.
This aligns metric behavior across the board and directly addresses the burning
issue outlined in this ticket.
{quote}
This is a very bad idea. Using an sstable without referencing it is a bug that
can cause very serious problems, including crashes due to accessing released
memory. You could get away with doing {{getSSTables}} instead of
{{selectAndReference}} if you only reference the sstable reader's statistics;
some metrics do that and should be safe.
{{{}estimatedRowCount{}}}/{{{}estimatedPartitionCount{}}} need to read the
indexes and _will_ cause crashes if they are not protecting their sstable
access.
Compaction itself takes references of the input sstables for the duration of
the compaction for the same reason. This needn't include expired sstables,
because they are not used as input. If we do need to have a quick fix, changing
the compaction code to reference the whole transaction set instead of just the
actually compacting ones would work.
Doing it properly is not more complex, though. We can do this is by marking all
expired sstables as obsoleted at the beginning of the compaction to avoid the
possibility of [prematurely releasing them during early-open
checkpoints|https://github.com/blambov/cassandra/blob/trunk/src/java/org/apache/cassandra/db/lifecycle/LifecycleTransaction.java#L377].
> 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: 20m
> 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
