That is a bug, in fact, I have a pretty good idea which one, we fixed it in 3.23.26 - if the query is long enough and you have high load, the slave could read a log event partially because of a mutex bug - the binlog dump thread was not acquiring the binlog mutex when reading log events - so if a log event was short enough to be written atomically, this was ok, but otherwise, the slave would get confused, drop the connection, and then re-open it - by that time the event in question was already fully in the log, so this was not fatal, just performance degrading.

If 3.23.26 does not fix this, let us know - it would be nice if you had a test case we can run, but if not, we can still try to figure something out. 3.23.26 at least has better diagnostic about partial fail log even reads, so it should be easier to debug ( although we hope we won't have to do it).