[
https://issues.apache.org/jira/browse/RATIS-2390?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ivan Andika updated RATIS-2390:
-------------------------------
Description:
Currently, SegmentedRaftLog uses RaftLogBase.lock for all its operations even
those that only operates on certain segments (e.g. purge, appends). It might be
possible to reduce lock contention by having a unique lock per segment
(following the segment lifecycle).
For example, operations that need to operate on segments only need to hold read
lock on the top level RaftLogBase.lock (in case top level attributes like
commitIndex, snapshotIndex, etc are updated) and subsequently try to hold
SegmentedRaftLog segment locks. We can implement something like
writeLock(TermIndex fromIndex, TermIndex toIndex) and SegmentedRaftLog should
only pick the segments locks for that range.
was:
Currently, SegmentedRaftLog uses RaftLogBase.lock for all its operations even
those that only operates on certain segments (e.g. purge, appends). It might be
possible to reduce lock contention by having a unique lock per segment.
For example, operations that need to operate on segments only need to hold read
lock on the top level RaftLogBase.lock (in case top level attributes like
commitIndex, snapshotIndex, etc are updated) and subsequently try to hold
SegmentedRaftLog segment locks. We can implement something like
writeLock(TermIndex fromIndex, TermIndex toIndex) and SegmentedRaftLog should
only pick the segments locks for that range.
> Granular segment lock for SegmentedRaftLog
> ------------------------------------------
>
> Key: RATIS-2390
> URL: https://issues.apache.org/jira/browse/RATIS-2390
> Project: Ratis
> Issue Type: Improvement
> Components: RaftLog, server
> Reporter: Ivan Andika
> Priority: Major
> Labels: performance
>
> Currently, SegmentedRaftLog uses RaftLogBase.lock for all its operations even
> those that only operates on certain segments (e.g. purge, appends). It might
> be possible to reduce lock contention by having a unique lock per segment
> (following the segment lifecycle).
> For example, operations that need to operate on segments only need to hold
> read lock on the top level RaftLogBase.lock (in case top level attributes
> like commitIndex, snapshotIndex, etc are updated) and subsequently try to
> hold SegmentedRaftLog segment locks. We can implement something like
> writeLock(TermIndex fromIndex, TermIndex toIndex) and SegmentedRaftLog should
> only pick the segments locks for that range.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
