Re: Hooks for session start and end, take two

Hi,

On 2019-10-03 11:19:58 -0300, Fabrízio de Royes Mello wrote:
> On Thu, Oct 3, 2019 at 1:35 AM Andres Freund <andres(at)anarazel(dot)de> wrote:
> >
> > [...]
> >
> > In this state I think this patch should be flat out rejected.
> >
>
> Ok.
>
> > I'm seriously baffled at how this stuff is being pursued aggressively,
> > quite apparently without any serious design considerations. I mean this
> > stuff had to be backed out twice, yet it's being reproposed without much
> > consideration for concerns?
> >
>
> And what if (again) for the first version of session start hook we do it
> just for:
> - client backends
> - background workers
>
> For sure we'll need new design, but for now I'm can't imagine a use case
> for sessions different than listed above.

I think what would need to be designed is something more like
RegisterXactCallback():

/*
* start- and end-of-transaction callbacks for dynamically loaded modules
*/
typedef enum
{
XACT_EVENT_COMMIT,
XACT_EVENT_PARALLEL_COMMIT,
XACT_EVENT_ABORT,
XACT_EVENT_PARALLEL_ABORT,
XACT_EVENT_PREPARE,
XACT_EVENT_PRE_COMMIT,
XACT_EVENT_PARALLEL_PRE_COMMIT,
XACT_EVENT_PRE_PREPARE
} XactEvent;

typedef void (*XactCallback) (XactEvent event, void *arg);

extern void RegisterXactCallback(XactCallback callback, void *arg);
extern void UnregisterXactCallback(XactCallback callback, void *arg);

which would be called at various parts of a processes lifetime. Maybe
something like:

1) shortly after process creation
2) shortly after shared memory attach
3) after database connection has been established
4) after database writes are possible (this excludes database-less walsenders)
5a) before closing database connection due to fatal error
(with a note saying that no writes are ever allowed)
5b) before closing database connection due to normal exit
(with a note saying that errors better be handled, and that no
transaction may escape)
6) before shared memory detach
7) before final exit

Greetings,

Andres Freund