furquan39 commented on issue #66299:
URL: https://github.com/apache/doris/issues/66299#issuecomment-5133611053

   Adding measured evidence for this retry-loop, from a version-bloated 
reproduction table where a subset of tablets are permanently un-recompactable: 
a single already-merged rowset materializes over 4 GiB of one column 
uncompressed, so every compaction attempt fails E-3113 (the sibling issue 
#66298). That is exactly the permanent-failure class this issue is about.
   
   Under the current no-backoff gate (`tablet_sched_delay_time_ms` = 5000 ms, 
tablet_manager.cpp:770-780) these tablets retried continuously with zero 
durable progress. Sampling the cumulative E-3113 failure count on the table 
across successive ~10-minute windows, it climbed 12 -> 14 -> 26 -> 42 -> 68 
while our fix attempts made no progress: a self-sustaining failed-compaction 
retry loop burning continuous wasted I/O. Nothing throttled it; only manually 
setting `disable_auto_compaction=true` stopped it.
   
   At production scale on shared storage this is the amplifier. A single 
un-compactable tablet becomes the dominant writer and turns into a cluster-wide 
I/O storm that starved co-located services and took down unrelated workloads. 
The underlying failure cause is beside the point for this issue: a per-tablet 
consecutive-failure counter plus exponential backoff plus quarantine of 
permanently-failing tablets would bound the wasted I/O to negligible regardless 
of why a given compaction is doomed.
   
   We remain willing to submit the PR.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to