Details

    Description

      Multi-master load hangs in seconds possible causes:

      Attachments

        Issue Links

          Activity

            seppo Seppo Jaakola added a comment -

            My performance test with a two node cluster, where sysbench load is targeted to first node only, and configuring only one replication applier thread, gives following transaction throughput rates:

            • 10.6-MDEV-24966: 318.56 per sec
            • 10.5: 287.54 per sec

            so, my figures are in par with Jan's observations for 10.6-MDEV-24966 build, but my 10.5 build shows a lot lower throughput.

            seppo Seppo Jaakola added a comment - My performance test with a two node cluster, where sysbench load is targeted to first node only, and configuring only one replication applier thread, gives following transaction throughput rates: 10.6- MDEV-24966 : 318.56 per sec 10.5: 287.54 per sec so, my figures are in par with Jan's observations for 10.6- MDEV-24966 build, but my 10.5 build shows a lot lower throughput.

            I tested normal master-slave async replication using row based replication with slave_parallel_threads=16 using same sysbench setting as previously where load is targeted to master only.

            • 10.6-MDEV-24966 : 595 per sec with avg latency 26ms
            • 10.5 : 577 per sec with avg latency 27 ms
            • I do not know how to finish performance test only when slave node has applied all replication events (this same applies to Galera at least when wsrep_sync_wait=0 i.e. default)
            • So at least on my tests I do not see issues on async replication
            jplindst Jan Lindström (Inactive) added a comment - I tested normal master-slave async replication using row based replication with slave_parallel_threads=16 using same sysbench setting as previously where load is targeted to master only. 10.6- MDEV-24966 : 595 per sec with avg latency 26ms 10.5 : 577 per sec with avg latency 27 ms I do not know how to finish performance test only when slave node has applied all replication events (this same applies to Galera at least when wsrep_sync_wait=0 i.e. default) So at least on my tests I do not see issues on async replication

            Read-only regression between 10.5 and 10.6 is caused by https://jira.mariadb.org/browse/MDEV-25404 and similar regression exist on read-write loads also as noted. Pull request solves multi-master hang and other clear issues.

            jplindst Jan Lindström (Inactive) added a comment - Read-only regression between 10.5 and 10.6 is caused by https://jira.mariadb.org/browse/MDEV-25404 and similar regression exist on read-write loads also as noted. Pull request solves multi-master hang and other clear issues.

            Buildbot review.

            jplindst Jan Lindström (Inactive) added a comment - Buildbot review.

            Removing the remaining effort since this issue is closed.

            julien.fritsch Julien Fritsch added a comment - Removing the remaining effort since this issue is closed.

            People

              seppo Seppo Jaakola
              jplindst Jan Lindström (Inactive)
              Votes:
              0 Vote for this issue
              Watchers:
              5 Start watching this issue

              Dates

                Created:
                Updated:
                Resolved:

                Git Integration

                  Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.