2024-11-25T04:02:15.219Z

<Sridhar Seshasayee> @Matan Breizman your assessment w.r.t `osd_mclock_override_recovery_settings` is correct.  It's unclear if the cluster is HDD based. If it's HDD based and according to the symptoms, the user is likely hitting <https://tracker.ceph.com/issues/66289>. The fix is in main and will make to Reef and Squid upstream soon. A more complete fix is in the works and would need changes across the dmClock submodule and in Ceph OSD/mClockScheduler. The fix in dmClock is almost completed and the PR would soon be out for review.

2024-11-25T04:11:55.600Z

<Sridhar Seshasayee> The issue mainly boils down to multiple mClock queues per OSD (i.e. multiple op queue shards) which currently work independently resulting in incorrect QoS calculations. The fix essentially ties the multiple queues together from the tag calculation standpoint resulting in accurate scheduling while still reaping the benefit of multiple shard and thread per shard configuration.

2024-11-25T06:12:44.171Z

<Æmerson> It's an artifact of Asio's co_compose. Put a pragma in or disable the warning.

2024-11-25T07:30:53.993Z

<Ronen Friedman> @Sridhar Seshasayee - can I quote the following from your response:
"It's unclear if the cluster is HDD based. If it's HDD based and according to the symptoms, the user is likely hitting <https://tracker.ceph.com/issues/66289>. The fix is in main and will make to Reef and Squid upstream soon."
as a comment in that Reddit thread?

2024-11-25T07:33:53.438Z

<Sridhar Seshasayee> @Ronen Friedman Sure, that's fine. In addition, if the user wants to resolve it, the shard configuration may be modified as per the associated PR and the OSDs restarted for the change to take effect. The fix essentially does the same.

2024-11-25T09:56:07.793Z

<Igor Golikov> How this code passes CI? Do they disable the warning during the build?

2024-11-25T17:14:39.809Z

<Æmerson> It's specific to GCC.

2024-11-25T17:14:44.323Z

<Æmerson> Clang doesn't have it.

2024-11-25T17:19:13.528Z

<Æmerson> That is, Clang may have the warning, but only GCC has the false positive.

2024-11-25T20:43:06.335Z

<Igor Gomon> Hello, I have a question about Rocks DB compaction in OSDs, specifically about what configuration options are available to control the process. Also how resource-intensive is the compaction process? Would it significantly degrade OSDs capability to handle client requests? Thanks.