Research on perf degradation with binlog_transaction_dependency_tracking=WRITESET

The patch has been back-ported to 8.0 branch ( ) and will be available in the upcoming PS 8.0.39 release.

Unfortunately, as 5.7 does not use c++11 by default, std::unordered_map is not available there and therefore the patch won’t be back-ported to this branch.

We haven’t contributed this to upstream yet. We are still discussing this, to be honest (although the change itself is that simple that Oracle can definitely implement the same on their own).

Backporting this to 8.0 and even 5.7 seems like a good idea - I do not see any obstacles to do so.

BTW, we were also experimenting with additional container tuning, so most probably there will be more improvements in 8.4.2.

I see there is a patch for this for 8.4, will this be contributed upstream ?

Also, the binlog_transaction_dependency_tracking=WRITESET feature is in 8.0 and 5.7, but not enabled by default. Are-you considering back-porting ?

And btw, this is a very nice catch !

It seems that we found a workaround for this issue. The results are average QPS for 6 write workloads:

Probably the structure that keeps row hashes is not a hash table.