dbnode stops ticking when a node is removed from the placement
BertHartm opened this issue · comments
might be related to #3933
On version 1.3.0, we're seeing that when we remove a node from a placement in a cluster, many or most of the nodes stop ticking and seem to freeze up. They're still running, but the writes and reads stop and the data movement required can't happen.
We do see that database_tick_duration{quantile="0.99"}
becomes NaN on these hosts. Restarting the affected nodes seems to resolve the issue.
This cluster is fairly quiet at this point, which I believe why we aren't seeing the large memory spike in 3933.
I'm trying to come up with a smaller scale reproduction to share.
I'm still working on a test case, but as an additional datapoint, I used the placement/set api endpoint to bulk remove 24 nodes (of 32) in an isolation group (same cluster, same ig), and did not hit the issue, so something is different there.