Botond Dénes
Botond Dénes
@nyh if you have a reproducer, can you try to move the `_querier_cache.stop()` to *before* `_user_sstables_manager->close();` and see if it solves the problem?
Maybe we also need to move the semaphore stops as well, the semaphores can also have inactive reads lingering, holding on to sstables.
> I think you're right about the semaphore thing, but it will need a different scenario to reproduce a problem (a workload with high concurrency during shutdown leading to one...
> > I think you're right about the semaphore thing, but it will need a different scenario to reproduce a problem (a workload with high concurrency during shutdown leading to...
> > > I gave it some more thought, and I don't think this semaphore closure order can cause a real problem. If we have some requests hung on the...
> I've seen elsewhere a discussion that we cannot ignore commit log. @denesb ? If we flush *all* memtables just prior to compaction, then we can ignore the commitlog. If...
> I am bit worried about ignoring uncompacting sstables. Please make sure that all sstables are included by major in this new mode, there can be concurrent compactions. I, myself,...
We have https://github.com/scylladb/scylladb/pull/17969 to drop JMX from packaging.
> it's hard coded in the test, to use the old version of the tool. > > and when changing it, we can get into: #17412 > > i.e. /proc/sys/fs/aio-max-nr...
I thought we already removed jmx from the offline installer, as part of https://github.com/scylladb/scylladb/pull/18487.