[IMPALA-341] Remote profiles may be ignored by coordinator if query has a limit

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: Impala 1.0, Impala 2.3.0
• Fix Version/s: Impala 3.2.0
• Component/s: Backend
• Labels: query-lifecycle supportability
• Target Version: Product Backlog
• Assignee: Thomas Tauber-Marshall
• Reporter: Henry Robinson
• Created: 06/May/13 22:43
• Updated: 14/Nov/18 17:49
• Resolved: 14/Nov/18 17:49

Description

For a query with a LIMIT, the coordinator cancels remote fragments once the limit has been reached. This can happen before the backends have successfully sent a profile update, and any subsequent updates are ignored because the query is cancelled, so the complete profile is never fleshed out.

When the query is an INSERT, it is possible that a race in the coordinator code handling the remote profiles leads to a crash. This only seems to happen occasionally with large clusters (as reported), and we haven't been able to reproduce this internally yet.

We should probably wait until all backends have completed one way or the other before tearing down a query completely.

A workaround is to remove the LIMIT.

Issue Links

• depends upon: IMPALA-4063 Make fragment instance reports per-query (or per-host) instead of per-fragment instance.
• is related to: IMPALA-7148 test_profile_fragment_instances is flaky
• relates to: IMPALA-6338 Tests fail due to runtime profile for query with limit missing pieces
• relates to: IMPALA-5783 Investigate exec summary behavior