Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
ghx-label-14
Description
Remote storage systems (e.g. cloud stores like S3 and ABFS) often have long tail latencies. Most I/O finishes relatively quickly, but some calls make take significantly longer. Even for HDFS, this is an issue (e.g. hedged reads were developed to help mitigate tail latencies, although no such feature exists for cloud storage connectors).
Currently, scan nodes just track the total amount of time spent reading data. It would be good to have a summary stats counter that tracks the min, avg, and max time spent reading data. This should at least allow us to identify when calls to remote storage services are taking longer than usual.
Attachments
Issue Links
- relates to
-
IMPALA-8884 Track Read(), Open() and Write() operation time per disk queue
- Resolved
-
IMPALA-9033 Log metrics for slow I/Os
- Resolved