Re: MERGE SQL Statement for PG11

On Fri, Oct 27, 2017 at 7:41 AM, Simon Riggs <simon@2ndquadrant.com> wrote:
> Good points.
>
> I didn't say it but my intention was to just throw an ERROR if no
> single unique index can be identified.

You'd also throw an error when there was no "upsert compatible" join
quals, I take it?

I don't see the point in that. That's just mapping one syntax on to another.

> It could be possible to still run MERGE in that situaton but we would
> need to take a full table lock at ShareRowExclusive. It's quite likely
> that such statements would throw duplicate update errors, so I
> wouldn't be aiming to do anything with that for PG11.

I would avoid mixing up ON CONFLICT DO UPDATE and MERGE. The
"anomalies" you describe in MERGE are not really anomalies IMV.
They're simply how the feature is supposed to operate, and how it's
possible to make MERGE use alternative join algorithms based only on
the underlying cost. You might use a merge join for a bulk load
use-case, for example.

I think an SQL MERGE feature would be compelling, but I don't think
that it should take much from ON CONFLICT. As I've said many times
[1], they're really two different features (Teradata had features
similar to each, for example). I suggest that you come up with
something that has the semantics that the standard requires, and
therefore makes none of the ON CONFLICT guarantees about an outcome
under concurrency (INSERT or UPDATE). Those guarantees are basically
incompatible with how MERGE needs to work.

In case it matters, I think that the idea of varying relation
heavyweight lock strength based on subtle semantics within a DML
statement is a bad one. Frankly, I think that that's going to be a
nonstarter.

[1] https://www.postgresql.org/message-id/CAM3SWZRP0c3g6+aJ=YYDGYAcTZg0xA8-1_FCVo5Xm7hrEL34kw@mail.gmail.com
-- 
Peter Geoghegan