[
https://issues.apache.org/jira/browse/IGNITE-17380?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kirill Gusakov updated IGNITE-17380:
------------------------------------
Description:
On the leader election RAFT stores meta data about new term to separate file.
This procedure takes significant amount of time according to profiler dump
(attached).
This issue slowdowns:
- leader reelection in general
- start of partitions with 1 replicas (in this case, start will consists of
syncrhonous electSelf call)
We should investigate the ways to smooth out this effect:
- switch off fsync for this operation (it needs to investigate, that
correctness will not suffer)
- store this information in rocksdb instead of file
- <find another way>
was:
On the leader election raft store meta data about new term to separate file.
This procedure takes significant amount of time according to profiler dump
(attached).
This issue slowdowns:
- leader reelection in general
- start of partitions with 1 replicas (in this case, start will consists of
syncrhonous electSelf call)
We should investigate the ways to smooth out this effect:
- switch off fsync for this operation (it needs to investigate, that
correctness will not suffer)
- store this information in rocksdb instead of file
- <find another way>
> Slow atomicMove+fsync on leader election
> ----------------------------------------
>
> Key: IGNITE-17380
> URL: https://issues.apache.org/jira/browse/IGNITE-17380
> Project: Ignite
> Issue Type: Task
> Reporter: Kirill Gusakov
> Priority: Major
> Attachments: new-cpu.html
>
>
> On the leader election RAFT stores meta data about new term to separate file.
> This procedure takes significant amount of time according to profiler dump
> (attached).
> This issue slowdowns:
> - leader reelection in general
> - start of partitions with 1 replicas (in this case, start will consists of
> syncrhonous electSelf call)
> We should investigate the ways to smooth out this effect:
> - switch off fsync for this operation (it needs to investigate, that
> correctness will not suffer)
> - store this information in rocksdb instead of file
> - <find another way>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
