Andrew Lawrenson wrote:
I'm not sure if it's the exact same issue, but I have observed the replication failing under heavy load where the replication buffers would fill faster than the logs can be shipped to the slave (even with fast servers & fast network). I've previously tweaked a copy of derby to allow up to 100 buffers (rather than the current max of 10), and optimized the various thresholds, but still observed over 80 buffers being used under high load (this was with 10.4.2.0 on Solaris).
I think it is the same issue - I saw an errors in the master log that mentioned something about the buffers filling up. The thing that seems to trigger it is doing a number of whole-table deletes, presumably because while that's being processed on the slave the master is still chugging along writing more log records.
What I don't understand is why the master isn't being throttled if the slave can't keep up.
-- Alan Burlison --
