Minor Description
After fixing KAFKA-4785 all internal topics using built-in RecordMetadataTimestampExtractor to read timestamps.
Which doesn't seem to work correctly out of box with kafka brokers configured with log.message.timestamp.type=LogAppendTime when using custom message timestamp extractor.
Example use-case windowed grouping + aggregation on late data:
KTable<Windowed<Tuple>, Long> summaries = in
.groupBy((key, value) -> ......)
.windowedBy(TimeWindows.of(TimeUnit.HOURS.toMillis(1l)))
.count();
when processing late events:
- custom timestamp extractor will pick up timestamp in the past from message (let's say hour ago)
- re-partition topic during grouping phase will be written back to kafka using timestamp from (1)
- kafka broker will ignore provided timestamp in (2) to favor ingestion time
- streams lib will read re-partitioned topic back with RecordMetadataTimestampExtractor
- and will get ingestion timestamp (3), which usually close to "now"
- window start/end will be incorrectly set based on "now" instead of original timestamp from payload
Understand there are ways to configure per-topic timestamp type in kafka brokers to solve this, but it will be really nice if kafka-streams library can take care of it itself.
To follow "least-surprise" principle. If library relies on timestamp.type for topic it manages it should enforce it.
CC guozhang based on user group email discussion.
Attachments
Issue Links
- links to