Details
-
Bug
-
Status: Resolved
-
Normal
-
Resolution: Fixed
-
None
-
Degradation
-
Normal
-
Normal
-
Performance Regression Test
-
All
-
None
-
Description
This adds a handful of relatively minor changes that makes high key density accord workloads more performant.
Increase accord repair range splitting - given a cluster with 6 million keys, this splits up accord repairs into 600 key chunks.
Streamline table metadata fetching - removes some unnecessary abstraction from the table metadata lookup path, which we hit a lot during key deserialization.
Remote unnecessary set building when building lists of overlapping keys - we don’t need to dedupe keys and the set building was taking a decent chunk of time.
Add separate recover delay for repair and increase default recover delay - this significantly reduces the number of preemptions.
The range splitting and recovery delay changes stabilize performance today, but are kind of a stop gap - we should be able to support larger repair ranges and shorter recovery delays.