On Wed, Oct 11, 2017 at 11:51 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > I'm inclined to believe Robert's hypothesis that there is some race > condition there. > > The question is whether that still constitutes a serializability > problem since we haven't done anything with the data, just passed it > upwards to be modified.
Well, the question is whether passing it upwards constitutes reading it. I kind of suspect it does. The plan tree isn't just blindly propagating values upward but stuff with them. There could be quite a bit between the ModifyTable node and the underlying scan if DELETE .. FROM or UPDATE .. USING is in use.
Right. In general we can't skip SIReadLock just basing on the fact that we're under ModifyTable node.
It seems still possible for me to skip SIReadLock in some very limited (but probably typical) cases.
But before analyzing this deeper, it would be nice to estimate possible benefits.
We can try some broken version which skip all SIReadLock's under ModifyTable node and benchmark it.
If it wouldn't have significant performance benefit, then there is no reason to do something further...
