On Tue, Aug 29, 2017 at 8:32 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Mon, Aug 28, 2017 at 10:17 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> In the meantime, I think what we should do is commit the bug fix more or
>> less as I have it, and then work on Amit's concern about losing parallel
>> efficiency by separating the resetting of shared parallel-scan state
>> into a new plan tree traversal that's done before launching new worker
>> processes.
>>
Sounds reasonable plan to me.
>> The only real alternative is to lobotomize the existing rescan
>> optimizations, and that seems like a really poor choice from here.
>
> There's already ExecParallelReinitialize, which could be made to walk
> the nodes in addition to what it does already, but I don't understand
> exactly what else needs fixing.
>
Sure, but it is not advisable to reset state of all the nodes below
gather at that place, otherwise, it will be more or less like we are
forcing rescan of each node. I think there we can reset the shared
parallel state of parallel-aware nodes (or anything related) and then
allow rescan to reset the master backend specific state for all nodes
beneath gather.
--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com