Re: Is Recovery actually paused?

On Mon, Jan 18, 2021 at 9:42 PM Yugo NAGATA <nagata(at)sraoss(dot)co(dot)jp> wrote:
> If it is acceptable that pg_is_wal_replay_paused() makes users wait,
> I'm ok for the current interface. I don't feel the need of
> pg_is_wal_replay_paluse_requeseted().

Another idea could be that pg_is_wal_replay_paused() could be changed
to text, and the string could be either 'paused' or 'pause requested'
or 'not paused'. That way we'd be returning a direct representation of
the state we're keeping in memory. Some of the complexity in this
discussion seems to come from trying to squeeze 3 possibilities into a
Boolean.

Let's also consider that we don't really know whether the client wants
us to wait or not, and different clients may want different things, or
maybe not, but we don't really know at this point. If we provide an
interface that waits, and the client doesn't want to wait but just
know the current state, they don't necessarily have any great options.
If we provide an interface that doesn't wait, and the client wants to
wait, it can poll until it gets the answer it wants. Polling can be
inefficient, but anybody who is writing a tool that uses this should
be able to manage an algorithm with some reasonable back-off behavior
(e.g. try after 10ms, 20ms, keep doubling, max of 5s, or something of
that sort), so I'm not sure there's actually any real problem in
practice. So to me it seems more likely that an interface that is
based on waiting will cause difficulty for tool-writers than one that
does not.

Other people may feel differently, of course...

--
Robert Haas
EDB: http://www.enterprisedb.com