Lauro Venancio

Just providing some more details about the performance regression @racevedoo mentioned. We've updated two production clusters from scylla 4.5 to 4.6. Both experienced performance regressions. For instance, in the first...

> > It seems to me that the partition index is not very useful when there is no clustering key. Maybe it should not be cached in this case. >...

@ejona86 I've debugged this problem using the provided reproducer and I could confirm that, sometimes, the function ``channelInactive`` is called before the final bytes of some streams are processed. When...

@ejona86 It doesn't seem to be just a timing issue because I am still able to consistently reproduce the issue when I add a 100ms sleep at the ``channelInactive()`` function...

> The issue at hand is "why are there active streams when the channel becomes inactive (TCP connection closes)?" And the RST I found is a smoking gun for that....

> > When I move the call to super.channelInactive(ctx) to the beginning of the function, I am not able to reproduce the problem anymore. > > I'd have to look...

@ejona86 You are right. To move the `clientStream.transportReportStatus()` call makes no difference and there is a RST packet. After debugging a little more, I found the following timeline: 1. Server...