Re: [HACKERS] make async slave to wait for lsn to be replayed

Hello.

I looked this briefly but not tested.

At Fri, 06 Mar 2020 00:24:01 +0300, Kartyshov Ivan <i(dot)kartyshov(at)postgrespro(dot)ru> wrote in
> On 2018-03-06 14:50, Simon Riggs wrote:
> > On 6 March 2018 at 11:24, Dmitry Ivanov <d(dot)ivanov(at)postgrespro(dot)ru>
> > wrote:
> >>> In PG11, I propose the following command, sticking mostly to Ants'
> >>> syntax, and allowing to wait for multiple events before it returns. It
> >>> doesn't hold snapshot and will not get cancelled by Hot Standby.
> >>> WAIT FOR event [, event ...] options
> >>> event is
> >>> LSN value
> >>> TIMESTAMP value
> >>> options
> >>> TIMEOUT delay
> >>> UNTIL TIMESTAMP timestamp
> >>> (we have both, so people don't need to do math, they can use whichever
> >>> they have)
> >> I have a (possibly) dumb question: if we have specified several
> >> events,
> >> should WAIT finish if only one of them triggered? It's not immediately
> >> obvious if we're waiting for ALL of them to trigger, or just one will
> >> suffice (ANY). IMO the syntax could be extended to something like:
> >> WAIT FOR [ANY | ALL] event [, event ...] options,
> >> with ANY being the default variant.
> > +1
>
> Here I made new patch of feature, discussed above.
>
> WAIT FOR - wait statement to pause beneath statements
> ==========
>
> Synopsis
> ==========
> WAIT FOR [ANY | SOME | ALL] event [, event ...] options
> and event is:
> LSN value
> TIMESTAMP value
>
> and options is:
> TIMEOUT delay
> UNTIL TIMESTAMP timestamp

The syntax seems getting confused. What happens if we typed in the
command "WAIT FOR TIMESTAMP '...' UNTIL TIMESTAMP '....'"? It seems
to me the options is useles. Couldn't the TIMEOUT option be a part of
event? I know gram.y doesn't accept that syntax but it is not
apparent from the description above.

As I read through the previous thread, one of the reason for this
feature implemented as a syntax is it was inteded to be combined into
BEGIN statement. If there is not any use case for the feature midst
of a transaction, why don't you turn it into a part of BEGIN command?

> Description
> ==========
> WAIT FOR - make to wait statements (that are beneath) on sleep until
> event happens (Don’t process new queries until an event happens).
...
> Notice: WAIT FOR will release on PostmasterDeath or Interruption
> events
> if they come earlier then LSN or timeout.

I think interrupts ought to result in ERROR.

wait.c adds a fair amount of code and uses proc-array based
approach. But Thomas suggested queue-based approach and I also think
it is better. We already have a queue-based mechanism that behaves
almost the same with this feature in the comit code on master-side. It
avoids spurious backend wakeups. Couldn't we extend SyncRepWaitForLSN
or share a part of the code/infrastructures so that this feature can
share the code?

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center