On Mon, Oct 27, 2014 at 5:15 PM, Peter Geoghegan <pg@heroku.com> wrote:
>> Let's see if we can link these two thoughts.
>>
>> 1. You think the biggest problem is the lack of attention to the design.
>>
>> 2. I keep asking you to put the docs in a readable form.
>>
>> If you can't understand the link between those two things, I am at a loss.
>
> You've read the docs. Please be clearer. In what sense are they not
> readable? The main description of the feature appears on the INSERT
> reference page:
>
> http://postgres-benchmarks.s3-website-us-east-1.amazonaws.com/on-conflict-docs/sql-insert.html
I've updated that reference page. I did a fair amount of copy-editing,
but also updated the docs to describe the latest (unpublished)
refinements to the syntax. Which is, as you and Robert requested, that
the target and rejected-for-insertion tuples may be referenced with
magical aliases in the style of OLD.* and NEW.*. I've spelt these
aliases as TARGET.* and EXCLUDED.*, since OLD.* and NEW.* didn't seem
to make much sense here. This requires some special processing during
rewriting (which, as you probably know, is true of DML statements in
general), and is certainly more invasive than what I had before, but
all told isn't too bad. Basically, there is still an ExcludedExpr, but
it only appears in the post-rewrite query tree, and is never created
by the raw grammar or processed during parse analysis.
I attach the doc patch with the relevant changes, in case you'd like a
quick reference to where things are changed.
I have already implemented the two things that you and Robert asked
for most recently: A costing model for unique index inference, and the
above syntax. I've also added IGNORE support to postgres_fdw (so you
can IGNORE if and only if a unique index inference specification is
omitted, just as with updatable views since V1.3).
Currently, I'm working on fixing an issue with RLS that I describe in
detail here:
https://wiki.postgresql.org/wiki/UPSERT#RLS
Once I fix that (provided it doesn't take too long), I'll publish a
V1.4. AFAICT, that'll close out all of the current open issues.
I hope this goes some way towards addressing your concerns.
--
Peter Geoghegan
