Details
-
Bug
-
Resolution: Fixed
-
Major
-
7.6.0, 7.6.1
-
Untriaged
-
0
-
Yes
-
Build Team 2024 Sprint 3
Description
In 7.1.x and 7.2.x we started using a patched version of jemalloc which allows for more tcaches.
This was necessary to avoid https://issues.couchbase.com/browse/MB-51608 while simultaneously fixing https://issues.couchbase.com/browse/MB-55268.
Our fork is under https://github.com/couchbasedeps/jemalloc. Versions of jemalloc 5.2.1 (used in 7.1.x and 7.2.x) and jemalloc 5.3.0 used in 7.6.0 have been patched, but 5.3.0 was never rebuild with our patch on top. As a result, 7.6.0 and 7.6.1 contain a jemalloc which doesn't contain our change, and will be affected by https://issues.couchbase.com/browse/MB-51608 again.
Customers can see crashes when (num-bucket x num-cpus) is large and they are using Disk IO Optimized thread settings. For example, under a 64 core machine with 20 Magma buckets.
See this comment: https://issues.couchbase.com/browse/MB-51608?focusedId=592355&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-592355
Fix
We need to rebuild jemalloc for 7.6.2 from the 5.3.0-couchbase branch.
Workaround
The number of IO threads can be reduced to avoid a crash.