The main motivation for a shorter commit interval for EOS is
end-to-end-latency. A Topology could consist of multiple sub-topologies
and the end-to-end-latency for the EOS case is roughly commit-interval
times number-of-subtopologies.
For regular rebalances/restarts, a longer commit interval has no impact,
because for a regular rebalance/restart, offsets would be committed
right away to guarantee a clean hand-off. Only in case of failure, a
longer commit interval can lead to larger amount of duplicates (of
course only for at-least-once guarantees).
For at-least-once, you would still get output continuously, depending on
throughput and producer configs. Only offsets are committed each 30
seconds by default. This continuous output is also the reason why there
is not latency impact for at-least-once using a longer commit interval.
Beside an impact on latency, there is also a throughput impact. Using a
longer commit interval provides higher throughput.
-Matthias
On 10/4/21 7:31 AM, Pushkar Deole wrote:
Hi All,
I am looking into the commit.interval.ms in kafka streams which says that
it is the time interval at which streams would commit offsets to source
topics.
However for exactly-once guarantee, default of this time is 100ms whereas
for at-least-once it is 30000ms (i.e. 30sec)
Why is there such a huge time difference between the 2 guarantees and what
does it mean to have this interval as high as 30 seconds, does it also
cause more probability of higher no. of duplicates in case of application
restarts or partition rebalance ?
Does it mean that the streams application would also publish events to
destination topic only at this interval which means delay in publishing
events to destinations topic ?
