Execution locality feature (changefeed processor pinning/regional sins) is under-tested.
At the very least, add roachtest to test locality pinning
Identify and fix observability gaps (are there any locality specific metrics we need to be tracking of)?
It appears that while aggregators will be pinned to a particular region, the actual rangefeed may
consume data from a replica that might not be the best choice (even outside of the locality; we prefer
to use proximate replicas, but I don't think it's guaranteed). If the above can happen, we should at least
have observability when it does.
blathers-crl[bot]
commented
1 year ago
Execution locality feature (changefeed processor pinning/regional sins) is under-tested.
At the very least, add roachtest to test locality pinning
Identify and fix observability gaps (are there any locality specific metrics we need to be tracking of)? It appears that while aggregators will be pinned to a particular region, the actual rangefeed may consume data from a replica that might not be the best choice (even outside of the locality; we prefer to use proximate replicas, but I don't think it's guaranteed). If the above can happen, we should at least have observability when it does.
Jira issue: CRDB-31395