Sahil Takiar created IMPALA-9301:
------------------------------------
Summary: Aux error info should detect multiple RPC failures
Key: IMPALA-9301
URL: https://issues.apache.org/jira/browse/IMPALA-9301
Project: IMPALA
Issue Type: Sub-task
Components: Backend
Reporter: Sahil Takiar
Suggested during the review of
([IMPALA-9296|http://issues.cloudera.org/browse/IMPALA-9296])
[https://gerrit.cloudera.org/#/c/15046/]
{quote}
I'm not sure that this is the right wa[y] to do it, since it means that if a
backend sees multiple rpc failures in a single query only one will ever be
reported to the coordinator.
Of course, I've been advocating for being aggressive about blacklisting.
Suppose there were two rpc failures, then there are two cases here - either
both rpcs were to the same other executor, in which case the fact that there
were two failures makes us more confident something is going on with that
executor and we might actually want to blacklist the executor twice (which will
just extend the amount of time that it stays blacklisted for), or the two rpcs
were to different executors, in which case if we only blacklist one of them if
we then retry the query it may very well fail again.
And even if we do want to stay more conservative about blacklisting, you've
suggested before (and I agree) that its generally preferable to report as much
info about errors as we've got, and then centralize the logic for deciding how
to act on those errors in the coordinator.
{quote}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
