Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
2.3.0
-
None
-
None
Description
After adding quantile collection to one of our clusters we saw much higher latency for RPCs. This was traced down to the quantile collection. Samples are being buffered and inserted in groups of 500. After the buffered samples are inserted, the entire set of samples for this time period (600 seconds at the longest for us) is "compressed".
All operations for RPC metrics are synchronized. Usually this isn't an issue but it seems that this compression operation is taking a significant amount of time.
Attachments
Attachments
Issue Links
- is related to
-
HDFS-10873 Add histograms for FSNamesystemLock Metrics
- Open
-
HDFS-10475 Adding metrics for long FSNamesystem read and write locks
- Open