Re: pg_stat_io for the startup process

At Fri, 28 Apr 2023 11:15:51 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in 
> At Thu, 27 Apr 2023 17:30:40 -0400, Melanie Plageman <melanieplageman@gmail.com> wrote in 
> > After a quick example implementation of this, I found that it seemed to
> > try and flush the stats less often on an idle standby (good) than using
> > enable_timeout_every().
> 
> Just rearming with the full-interval will work that way. Our existing
> strategy for this is seen in PostgresMain().
> 
>    stats_timeout = pgstat_report_stat(false);
>    if (stats_timeout > 0)
>    {
>       if (!get_timeout_active(BLAH_TIMEOUT))
>         enable_timeout_after(BLAH_TIMEOUT, stats_timeout);
>    }
>    else
>    {
>        if (get_timeout_active(BLAH_TIMEOUT))
>          disable_timeout(BLAH_TIMEOUT, false);
>    }
>    WaitLatch();

Im my example, I left out idle-time flushing, but I realized we don't
need the timeout mechanism here since we're already managing it. So
the following should work (assuming the timestamp updates with
GetCurrentTimestamp() in my last patch).

@@ -3889,13 +3900,23 @@ WaitForWALToBecomeAvailable(XLogRecPtr RecPtr, bool randAccess,
                     /* Update pg_stat_recovery_prefetch before sleeping. */
                     XLogPrefetcherComputeStats(xlogprefetcher);
 
+                    /*
+                     * Report stats; if not time yet, set next WaitLatch to
+                     * wake up at the next reporing time.
+                     */
+                    wait_time = pgstat_report_stat(false);
+
+                    /* if no pending stats, sleep forever */
+                    if (wait_time == 0)
+                        wait_time = -1L;
+
                     /*
                      * Wait for more WAL to arrive, when we will be woken
                      * immediately by the WAL receiver.
                      */
                     (void) WaitLatch(&XLogRecoveryCtl->recoveryWakeupLatch,
                                      WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
-                                     -1L,
+                                     wait_time,
                                      WAIT_EVENT_RECOVERY_WAL_STREAM);

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center