greigwise <greigwise(at)comcast(dot)net> writes:
> I do still have the 9.6.2 instance.
> Here is the explain analyze results from the 9.6.2 instance:
Oh ... you have parallel query enabled, eh? In that case I'm less
surprised, there were multiple parallel-query fixes since 9.6.2.
Although those still don't look like they would result in actually
wrong query answers. I wouldn't be surprised if whatever you're
seeing is still there, and you're just escaping it by luck in 9.6.3.
If you run the query multiple times on 9.6.2, exactly where do the
"actual" row counts change? Can you extract a test case you wouldn't
mind publishing?
regards, tom lane