[
https://issues.apache.org/jira/browse/KAFKA-17184?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17926014#comment-17926014
]
Jorge Esteban Quilcate Otoya commented on KAFKA-17184:
------------------------------------------------------
Interesting edge case. The entry seems to change state between `getIndexEntry`
and `lookupPosition` causing this exception.
I see a couple of options here:
* Use a writeLock when deleting: this would ensure the deletion and fetching
will happen at separate times. I guess this would be the cleaner option, but
writeLock was changed to readLock some time back:
[https://github.com/apache/kafka/pull/14283]
* Re-fetch on the `lookupPosition` when entry is marked for cleanup: if
marked, then cleanup and fetch again from remote.
cc [~divijvaidya] [~kamalcph]
> Remote index cache noisy logging
> --------------------------------
>
> Key: KAFKA-17184
> URL: https://issues.apache.org/jira/browse/KAFKA-17184
> Project: Kafka
> Issue Type: Bug
> Reporter: Francois Visconte
> Priority: Critical
>
> We have a tiered storage cluster where some consumers are constantly lagging
> behind.
> On this cluster, we get a ton of error logs and fail fetches with the
> following symptom:
> {code:java}
> java.lang.IllegalStateException: This entry is marked for cleanup
> at
> org.apache.kafka.storage.internals.log.RemoteIndexCache$Entry.lookupOffset(RemoteIndexCache.java:569)
> at
> org.apache.kafka.storage.internals.log.RemoteIndexCache.lookupOffset(RemoteIndexCache.java:446)
> at
> kafka.log.remote.RemoteLogManager.lookupPositionForOffset(RemoteLogManager.java:1445)
> at kafka.log.remote.RemoteLogManager.read(RemoteLogManager.java:1391)
> at kafka.log.remote.RemoteLogReader.call(RemoteLogReader.java:62)
> at kafka.log.remote.RemoteLogReader.call(RemoteLogReader.java:31)
> at java.base/java.util.concurrent.FutureTask.run(FutureTask.java:264)
> at
> java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1136)
> at
> java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:635)
> at java.base/java.lang.Thread.run(Thread.java:840)
> {code}
> I believe this should be handled differently by either
> * Set log level to warn or info
> * reload the index entry synchronously when an offset is requested and the
> entry is marked for cleanup.
> We do use the default setting for
> {{remote.log.index.file.cache.total.size.bytes}} (1GiB).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
