[
https://issues.apache.org/jira/browse/KAFKA-13289?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17473235#comment-17473235
]
Eugen Dück commented on KAFKA-13289:
------------------------------------
{quote}
bq. The app uses a 0ms join window with a 0ms
That is pretty aggressive and will result in a retention of 0ms, too. Thus,
every out-of-order record will be considered late and would be dropped (and you
would see the corresponding logging).
{quote}
In terms of joining (which is done per key), that is the behavior we want.
However it seems the dropping of messages is based on "partition time", i.e.
per partition? As we don't have monotonically increasing timestamps across
keys, that would mean we have an actual problem...
So a solution would be to increase the window size or the grace period (which I
honestly still don't know what the difference would be - a link or so to an
explanation would be awesome), at the cost of having massively increased number
of join pairs.
> Bulk processing correctly ordered input data through a join with
> kafka-streams results in `Skipping record for expired segment`
> -------------------------------------------------------------------------------------------------------------------------------
>
> Key: KAFKA-13289
> URL: https://issues.apache.org/jira/browse/KAFKA-13289
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Affects Versions: 2.8.0
> Reporter: Matthew Sheppard
> Priority: Minor
>
> When pushing bulk data through a kafka-steams app, I see it log the following
> message many times...
> {noformat}
> WARN
> org.apache.kafka.streams.state.internals.AbstractRocksDBSegmentedBytesStore -
> Skipping record for expired segment.
> {noformat}
> ...and data which I expect to have been joined through a leftJoin step
> appears to be lost.
> I've seen this in practice either when my application has been shut down for
> a while and then is brought back up, or when I've used something like the
> [app-reset-rool](https://docs.confluent.io/platform/current/streams/developer-guide/app-reset-tool.html)
> in an attempt to have the application reprocess past data.
> I was able to reproduce this behaviour in isolation by generating 1000
> messages to two topics spaced an hour apart (with the original timestamps in
> order), then having kafka streams select a key for them and try to leftJoin
> the two rekeyed streams.
> Self contained source code for that reproduction is available at
> https://github.com/mattsheppard/ins14809/blob/main/src/test/java/ins14809/Ins14809Test.java
> The actual kafka-streams topology in there looks like this.
> {code:java}
> final StreamsBuilder builder = new StreamsBuilder();
> final KStream<String, String> leftStream =
> builder.stream(leftTopic);
> final KStream<String, String> rightStream =
> builder.stream(rightTopic);
> final KStream<String, String> rekeyedLeftStream = leftStream
> .selectKey((k, v) -> v.substring(0, v.indexOf(":")));
> final KStream<String, String> rekeyedRightStream = rightStream
> .selectKey((k, v) -> v.substring(0, v.indexOf(":")));
> JoinWindows joinWindow = JoinWindows.of(Duration.ofSeconds(5));
> final KStream<String, String> joined = rekeyedLeftStream.leftJoin(
> rekeyedRightStream,
> (left, right) -> left + "/" + right,
> joinWindow
> );
> {code}
> ...and the eventual output I produce looks like this...
> {code}
> ...
> 523 [523,left/null]
> 524 [524,left/null, 524,left/524,right]
> 525 [525,left/525,right]
> 526 [526,left/null]
> 527 [527,left/null]
> 528 [528,left/528,right]
> 529 [529,left/null]
> 530 [530,left/null]
> 531 [531,left/null, 531,left/531,right]
> 532 [532,left/null]
> 533 [533,left/null]
> 534 [534,left/null, 534,left/534,right]
> 535 [535,left/null]
> 536 [536,left/null]
> 537 [537,left/null, 537,left/537,right]
> 538 [538,left/null]
> 539 [539,left/null]
> 540 [540,left/null]
> 541 [541,left/null]
> 542 [542,left/null]
> 543 [543,left/null]
> ...
> {code}
> ...where as, given the input data, I expect to see every row end with the two
> values joined, rather than the right value being null.
> Note that I understand it's expected that we initially get the left/null
> values for many values since that's the expected semantics of kafka-streams
> left join, at least until
> https://cwiki.apache.org/confluence/display/KAFKA/Kafka+Streams+Join+Semantics#KafkaStreamsJoinSemantics-ImprovedLeft/OuterStream-StreamJoin(v3.1.xandnewer)spurious
> I've noticed that if I set a very large grace value on the join window the
> problem is solved, but since the input I provide is not out of order I did
> not expect to need to do that, and I'm weary of the resource requirements
> doing so in practice on an application with a lot of volume.
> My suspicion is that something is happening such that when one partition is
> processed it causes the stream time to be pushed forward to the newest
> message in that partition, meaning when the next partition is then examined
> it is found to contain many records which are 'too old' compared to the
> stream time.
> I ran across this discussion thread which seems to cover the same issue
> http://mail-archives.apache.org/mod_mbox/kafka-users/202002.mbox/%3cCAB0tB9p_vijMS18jWXBqp7TQozL__ANoo3=h57q6z3y4hzt...@mail.gmail.com%3e
> and had a request from [~cadonna] for a reproduction case, so I'm hoping my
> example above might make the issue easier to tackle!
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
