If a global LOOP takes over 25 minutes between two slaves finishing their iteration, the barrier code in a global loop, spuriously interpreted the timeout as meaning 'proceed with next iteration', with the result that thormaster started the next iteration prematurely (tried to) or started an extra iteration that the slaves were not expecting.
In the case of starting an extra iteration, where the slaves have actually finished looping, the error seen is:
serializeCreateContexts - Timeout receiving from slaves
In the cast where it times out on an intermediate iteration, it would mean there were unconsumed end of loop messages from previous iterations pending, which would be consumed (early) by next iteration, causing all subsequent iterations to start early (on the master). It's likely you'd hit similar timeouts to the one above, as graphs initializing on master vs slave expect messages in short time frames, don't get them quickly enough.