szetszwo commented on PR #12075:
URL: https://github.com/apache/ozone/pull/12075#issuecomment-7017212746
> ... double overhead from Raft side because we double the RaftGroup/SM ...
@amaliujia , it won't since the V3 does not uses Raft, i.e. no raft log. V3
will just add a small number of objects (probably 1 or 2 per disk). We should
share the resource such as threads and also add new confs. Below are some
comments from Claude.
Claude: Closed Raft groups: you're right that they're mostly a transient
cost.
- Removal: the group is removed, but only when SCM deletes the pipeline:
after
ozone.scm.pipeline.destroy.timeout (300s), checked every
ozone.scm.pipeline.scrub.interval (150s). Until then it lives on every
datanode of the pipeline. That wait exists for RATIS, so containers can
close through Raft before the group goes away. CHAINED containers close
directly, so a CHAINED pipeline could be deleted right after its
containers
close, without waiting for the timeout.
- Storage: fine. A CHAINED group's Raft directory holds only configuration
entries and metadata, just a few KB.
- Threads: mostly not shared. Each group's ContainerStateMachine creates
its
own pool of hdds.container.ratis.num.container.op.executors (10) threads
(ContainerStateMachine:303). Ratis adds about 3–5 more per group. Only
the
chunk executors are shared. So threads grow with the number of groups,
open
plus not yet removed, and deleting CHAINED pipelines promptly keeps that
bounded. The per-group pool could also become shared later.
--
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]