Hi Martin,
That is a good point. In fact in the coming release we have made such
repartition topics really "transient" by periodically purging it with the
embedded admin client, so we can actually set its retention to -1:
https://cwiki.apache.org/confluence/display/KAFKA/KIP-220%3A+Add+AdminClien
Hi Guozhang,
Thanks very much for your reply. I am inclined to consider this a bug, since
Kafka Streams in the default configuration is likely to run into this problem
while reprocessing old messages, and in most cases the problem wouldn't be
noticed (since there is no error -- the job just pro
Hello Martin,
What you've observed is correct. More generally speaking, for various
broker-side operations that based on record timestamps and treating them as
wall-clock time, there is a mismatch between the stream records' timestamp
which is basically "event time", against the broker's system wa
Follow-up: I think we figured out what was happening. Setting the broker config
log.message.timestamp.type=LogAppendTime (instead of the default value
CreateTime) stopped the messages disappearing.
The messages in the Streams app's input topic are older than the 24 hours
default retention perio
