On Wed, Nov 2, 2016 at 12:49 PM, Tomas Vondra
<tomas.vondra@2ndquadrant.com> wrote:
> On 11/01/2016 08:32 PM, Robert Haas wrote:
>> On Tue, Nov 1, 2016 at 10:58 AM, Tomas Vondra
>> <tomas.vondra@2ndquadrant.com> wrote:
>>>
>>> Damn! You're right of course. Who'd guess I need more coffee this early?
>>>
>>> Attached is a fix replacing the flag with an array of flags, indexed by
>>> ParallelMasterBackendId. Hopefully that makes it work with multiple
>>> concurrent parallel queries ... still, I'm not sure this is the right
>>> solution.
>>
>> I feel like it isn't. I feel like this ought to go in the DSM for
>> that parallel query, not the main shared memory segment, but I'm not
>> sure how to accomplish that offhand. Also, if we do solve it this
>> way, surely we don't need the locking. The flag's only set before any
>> workers have started and never changes thereafter.
>
> I'm not sure what you mean by "DSM for that parallel query" - I thought the
> segments are created for Gather nodes, no? Or is there a DSM for the whole
> query that we could use?
Sure, the Gather node creates it. There's generally only one per
query, though, and that's how most information is communicated from
leader to workers.
> Another thing is that maybe we don't really want to give extensions access
> to any of those segments - my impression was those segments are considered
> internal (is there RequestAddinShmemSpace for them?). And hacking something
> just for auto_explain seems a big ugly.
Yeah. I thought that there wouldn't be any reason for third-party
code to need to get into these segments, but maybe that was
short-sighted of me. If we fix this without that, then we've got to
force pg_stat_statements to be loaded through
shared_preload_libarries, as you mentioned, and that doesn't seem nice
either.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company