In WT-8392 I added debugging to detect a rare missing record in test_timestamp_abort. I started running a 10x parallel job of the test to try to reproduce it. Finally, after eight days and almost 24000 iterations (x10 for nearly a quarter million successful runs) I hit the problem.

The debugging added does NOT indicate a file system bug. Instead there is a rare path where the log record is not flushed to the log in the operating system. My debugging in pwrite does not appear in the debugging output.

The thing that is unique about the local record in timestamp_abort is that it uses implicit transactions instead of explicit ones (like random_abort). So I speculate there is some path with the implicit transaction that misses that we need to flush the log record.