Hi, James,
That's a good point. KAFKA-3802 should cause the log segments to be kept
longer, instead of shorter. So, there is probably something else that's
causing this behavior. Could you try if you can reproduce this? When you do
that, one thing you could try is to set log.segment.delete.delay.m
KAFKA-3802 does seem plausible; I had to restart the brokers again after
the 0.10.1.0 upgrade to change some JVM settings; maybe that touched the
mtime on the files? Not sure why that would make them *more* likely to be
deleted, though, since their mtime should've gone into the future, not into
the
Hi, James,
Thanks for testing and reporting this. What you observed is actually not
the expected behavior in 0.10.1 based on the design. The way that retention
works in 0.10.1 is that if a log segment has at least one message with a
timestamp, we will use the largest timestamp in that segment to d
Incidentally, I'd like to note that this did *not* occur in my testing
environment (which didn't expire any unexpected segments after upgrading),
so if it is a feature, it's certainly a hit-or-miss one.
On Mon, Oct 31, 2016 at 4:14 PM, James Brown wrote:
> I just finished upgrading our main prod
I just finished upgrading our main production cluster to 0.10.1.0 (from
0.9.0.1) with an on-line rolling upgrade, and I noticed something strange —
the leader for one of our big partitions just decided to expire all of the
logs from before the upgrade. I have log.retention.hours set to 336 in my
co