Botond Dénes
Botond Dénes
We already have a keepalive mechanism via the `pointer_to_container` param to the snapshot reader. So I need to rethink my initial analysis.
We keep the memtable alive, but not the read section, which is passed down by ref and is not owned by (and therefore not kept alive by) the memtable.
The bug is a single missing `&` character, PR coming soon.
Fix: https://github.com/scylladb/scylladb/pull/23984
> [@denesb](https://github.com/denesb) who can look into this and [#21970](https://github.com/scylladb/scylladb/issues/21970)? [@lkshminarayanan](https://github.com/lkshminarayanan)? I'll look at it myself.
It is clear why this test is flaky: it assumes the memtable will be flushed in under 300ms, which might not happen on overloaded machines.
I see @Deexie already sent a PR earlier: https://github.com/scylladb/scylladb/pull/22162
@raphaelsc could this be the load monitor not recognizing that the DC has no more node and keeps trying to connect to already removed nodes?
@raphaelsc what is the status of this issue?
> This is fixed AFAIK. [@bhalevy](https://github.com/bhalevy) submitted a fix and backported. Why wasn't the issue closed? Can we close it now?