[
https://issues.apache.org/jira/browse/KAFKA-20902?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18103582#comment-18103582
]
Suman Pal commented on KAFKA-20902:
-----------------------------------
Hi [~jess], I’ve locally reproduced this issue using UT and am working on a fix
(total comparator with a per-iterator sequence id, as suggested in the ticket).
Assigning this to myself, will open a PR soon for review.
> Streams metered stores: oldest-iterator-open-since-ms can be wrong when two
> iterators open in the same millisecond
> ------------------------------------------------------------------------------------------------------------------
>
> Key: KAFKA-20902
> URL: https://issues.apache.org/jira/browse/KAFKA-20902
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Reporter: Jess Jin
> Priority: Minor
>
> {{openIterators}} is a {{ConcurrentSkipListSet}} keyed only by
> {{startTimestamp()}} (ms resolution). Two iterators opened in the same
> millisecond compare equal, so the second {{add()}} is a silent no-op and a
> {{remove()}} on close can evict the still-open one — the
> {{oldest-iterator-open-since-ms}} gauge then under-reports (may skip a
> still-open iterator or return 0). {{num-open-iterators}} is unaffected
> (separate {{LongAdder}}); iterators still close correctly, so it's a
> metric-accuracy issue, not a leak. Affects all metered store wrappers. *Fix:*
> make the comparator total (tie-break on a per-iterator sequence id). Found in
> PR [#22975|https://github.com/apache/kafka/pull/22975] review
> [{{r3730260330}}|https://github.com/apache/kafka/pull/22975#discussion_r3730260330].
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
