Red Hat Enterprise Linux
MariaDB Column Store 10.4.12-6

I gave Faisal a few suggestions re SM configuration at the time; it is unclear whether they helped or to what degree.

My suspicion is that SM flushes -> S3 are causing excessive contention in this use case. DML writes 8KB blocks to the version buffer constantly in this case, and the version buffer is a large monolithic file. When SM flushes (every 10s or when the write-cache is exhausted), it currently has to lock at the file level, which will block all DML writes while it is happening. This isn't a problem when not using DML, because the 'regular' database file are not large or monolithic.

I believe there are multiple overlapping tickets for reducing lock granularity in SM and for improving DML efficiency; we should find them and make the proper associations.

Had a thought while brainstorming a problem. If we always write to a tmp location, then move a file into the proper place on success, the window for a race on the data is greatly reduced. This should allow us to reduce some critical sections substantially.