Fix deadlock while cancelling a query#5033
Conversation
vonzshik
force-pushed
the
5032-user-query-cancellation-deadlock-fix
branch
from
5c0b843 to
79f235d
Compare
Is the same issue present in any of 6.0 releases or "backport" means something else? |
Yeah, it was introduced with #4907, which was backported to 7.0.2, 6.0.9 and 5.0.16. |
Is there going to be a different fix in 8.0 where the cancellation is guaranteed to happen (async continuation, or some other way)? |
That's the plan, for now I'm waiting for @roji to take a look so we can discuss further. |
|
This is high on my list, I'm finishing up some large/intensive unrelated work and will look at this soon. |
roji
left a comment
roji
left a comment
There was a problem hiding this comment.
LGTM as a targeted fix for backporting - but yeah, if possible we should maybe think about it for 8.0...
4 participants
Fixes #5032
While this pr should fix the issue, it might lead to a case where we won't cancel a running query if a cancellation token is triggered before we were able to read all of the responses for the prepended queries. The ideal fix (I think) would be to make cancellation fully async, but since we want to backport this to 7.0, 6.0 and 5.0...