Re: ParallelFinish-hook of FDW/CSP (Re: Steps inside ExecEndGather)

On Fri, Feb 17, 2017 at 12:46 PM, Kouhei Kaigai <kaigai(at)ak(dot)jp(dot)nec(dot)com> wrote:
> The attached patch is revised one.
>
> Invocation of Exec(Foreign|Custom)ParallelFinish was moved to
> ExecParallelRetrieveInstrumentation() not to walk on the plan-
> state tree twice.
> One (hypothetical) downside is, FDW/CSP can retrieve its own
> run-time statistics only when query is executed under EXPLAIN
> ANALYZE.
>
> This enhancement allows FDW/CSP to collect its specific run-
> time statistics more than Instrumentation, then show them as
> output of EXPLAIN. My expected examples are GPU's kernel execution
> time, DMA transfer ratio and so on. These statistics will never
> appear in the Instrumentation structure, however, can be a hot-
> point of performance bottleneck if CustomScan works on background
> workers.

Would gather_shutdown_children_first.patch from
https://www.postgresql.org/message-id/CAFiTN-s5KuRuDrQCEpiHHzmVf7JTtbnb8eb10c-6AywJDxbWrA@mail.gmail.com
help with this problem also? Suppose we did that, and then also added
an ExecShutdownCustom method. Then you'd definitely be able to get
control before the DSM went away, either from ExecEndNode() or
ExecShutdownNode().

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company