Todd Lipcon has uploaded a new change for review.
Change subject: Fix stray memory writes due to tcmalloc profiling
Fix stray memory writes due to tcmalloc profiling
This fixes an issue that has been causing frequent crashes in
JD.com's production cluster as well as various Cloudera test clusters.
The crashes would be in various different places, but the key signature
was that offset 120 in some data structure or array (eg the 16th element
of a vector) would be corrupted.
After doing a git bisect using an integration testing cluster running
an ITBLL workload, I found that this was a regression caused by the
introduction of tcmalloc contention profiling. The short explanation
is that, if we experienced contention while freeing a Trace object, we
could in some cases increment offset 120 of some other allocation
which occurred soon after the deallocation of the Trace.
The issue is described in more detail in a new comment in trace.h.
With this patch, I was unable to reproduce the issue on the test
cluster. No new test is added since this is quite timing-dependent
and not amenable to unit testing or even stress testing.
 commit f6691e744b9cb796e1bbc6e07953f21f387c9a88
Reviewed-by: David Ribeiro Alves <dral...@apache.org>
Reviewed-by: Mike Percy <mpe...@apache.org>
Tested-by: Kudu Jenkins
(cherry picked from commit 0bb59836fb3e4b92e2d88674e8b88546faac5c8b)
1 file changed, 23 insertions(+), 3 deletions(-)
git pull ssh://gerrit.cloudera.org:29418/kudu refs/changes/56/3456/1
To view, visit http://gerrit.cloudera.org:8080/3456
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings
Gerrit-Owner: Todd Lipcon <t...@apache.org>