I looked into the misbehavior reported in bug #14328,
https://www.postgresql.org/message-id/20160919160136.1348.55251@wrigleys.postgresql.org
What is happening is that during the EvalPlanQual recheck to see whether
the updated row still satisfies the SELECT's quals, we run ExecInitCteScan
and then CteScanNext in the EPQ-created copy of the plan tree. This
should result in a fresh scan of the underlying CTE ... but it does not,
because ExecInitCteScan sees the communication value
estate->es_param_exec_vals[node->cteParam] already set and decides it is
not a leader node, but rather a follower of the outer-query CTE Scan node
that it's a doppelganger of. That one's already at EOF so the
node->leader->eof_cte test fails in CteScanNext and it returns nothing.
The reason this happens is that EvalPlanQualBegin copies down all the
es_param_exec_vals values from the outer query into the EPQ execution
state. That's OK for most uses of es_param_exec_vals values, but not
at all for cteParam Params, which are used as internal communication
variables within a plan tree.
I believe that recursive unions are probably broken in EPQ rechecks in the
same way, since they use a Param for similar intra-plan communication
purposes, but haven't bothered to devise a test case to prove it.
I think the most robust way to fix this is to explicitly mark
es_param_exec_vals entries as to whether they should be copied down to
the EPQ execution state. There is room in struct ParamExecData to fit
a uint8 or uint16 field into existing pad space, so we could add a flags
field and then use one bit of it for this purpose without causing any
ABI breaks in the back branches.
A simpler fix would be to decide that we never need to copy any
es_param_exec_vals entries into the EPQ execution state. I think that
that would work, but it would provoke extra evaluations of InitPlan
subplans, and I'm hesitant to make such a significant behavioral change
without a lot more analysis.
Comments?
regards, tom lane