Hi Magnus, Sorry for the late reply. In the specific case you mentioned of an OffsetCommitResponse with 100 partitions across 10 topics we end up with -200bytes (no error) and +200bytes (any error) in absolute terms. Precise numbers depend on topic names too, of course but I did a little test using a hypothetical version 9 response and got these numbers:
Version=8 no error size=806 Version=9 no error size=606 Version=8 any error size=806 Version=9 any error size=1006 So about 25% less in the no-error case and 25% more in the error case, compared with version 8. Kind regards, Tom On Wed, Jul 8, 2020 at 8:47 PM Magnus Edenhill <mag...@edenhill.se> wrote: > Hi Tom, > > I think it would be useful with some real world (or made up!) numbers on > how much relative/% space is saved for > the most error-dense protocol requests. > E.g., an OffsetCommitResponse with 10 topics and 100 failing partitions > would reduce the overall size by % bytes. > > Thanks, > Magnus > > > Den tis 7 juli 2020 kl 17:01 skrev Colin McCabe <cmcc...@apache.org>: > > > Hi Tom, > > > > Thanks for this. I think the tough part is probably the few messages > that > > are still using manual serialization, which can't be easily converted to > > using this. So we will probably have to upgrade them to using automatic > > generation, or accept a little inconsistency for a while until they are > > upgraded. > > > > best, > > Colin > > > > > > On Wed, Jul 1, 2020, at 09:21, Tom Bentley wrote: > > > Hi all, > > > > > > Following a suggestion from Colin in the KIP-625 discussion thread, I'd > > > like to start discussion on a much smaller KIP which proposes to make > > error > > > codes and messages tagged fields in all RPCs. > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-363%3A+Make+RPC+error+codes+and+messages+tagged+fields > > > > > > I'd be grateful for any feedback you may have. > > > > > > Kind regards, > > > > > > Tom > > > > > >
