Let’s say it’s spurious and seemingly resource-dependent; data shape issues etc. would be apparent from the result data and maybe not need that much debugging.

Let’s say it’s spurious and seemingly resource-dependent; data shape issues etc. would be apparent from the result data and maybe not need that much debugging. Speaking from experience, a migration is also a pretty buffered operation (especially over a network), and stopping the migration at a breakpoint before the failure occurs will almost certainly give all layers enough time to flush this and that so that the problem magically disappears.