Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
Impala 0.6
-
None
-
None
Description
We've noticed a small (~5%) perf regression on some queries after a change to impose a max in flight row batch limit to control memory usage in pathological cases. This required a bit of lock refactoring in the scan node. Setting the limit to a very high value (one that is never hit by the query) still has the perf issue so it's likely the lock refactoring is the problem.
We should add thread instrumentation using getrusage to help diagnose the issue.