Re: Extension Templates S03E11

On Wed, Dec 11, 2013 at 2:49 PM, Dimitri Fontaine
<dimitri@2ndquadrant.fr> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>>> You've got that backwards.  We do have the luxury of rejecting new
>>> features until people are generally satisfied that the basic design is
>>> right.  There's no overlord decreeing that this must be in 9.4.
>>
>> I strongly agree.  PostgreSQL has succeeded because we try not to do
>> things at all until we're sure we know how to do them right.
>
> I still agree to the principle, or I wouldn't even try. Not in details,
> because the current design passed all the usual criteria a year ago.
>
>   http://www.postgresql.org/message-id/6466.1354817682@sss.pgh.pa.us
>
>> I can certainly understand Dimitri's frustration, in that he's written
>> several versions of this patch and none have been accepted.  But what
>
> The design was accepted, last year. It took a year to review it, which
> is fair enough, only to find new problems again. Circles at their best.
> You just said on another thread that perfect is the enemy of good. What
> about applying the same line of thoughts to this patch?

Sure.  For every patch that gets submitted, we have to decide whether
it represents an improvement over where we are today, or not.  For the
record:

1. The patch you're talking about is 2 or 3 orders of magnitude less
complicated than this one, and it is pretty easy to see that it will
not paint us into a corner.  It also happens to fix what I've long
considered a deficiency in PostgreSQL.  I think it is legitimate for
me to have more confidence in that patch than this one.

2. I explicitly did not review this patch for the precise reason that
I thought it needed a fresh pair of eyes.  You and I have not always
seen eye to eye on this and other extension-related patches, and I
don't really want to be the guy who shoots down your patches every
year.  I was prepared to sit still for this if Stephen felt it was OK.But after both Stephen and Heikki expressed
concernsabout the
 
approach, I decided to say that I found those concerns valid also.

I happen to think that Stephen did a very good job of explaining why
blobs in the catalog could be a very bad plan.  Specifically, I
believe he mentioned that it creates a dump/restore hazard.  If a new
keyword gets added in a new server version, a logical dump of the
extension made by a new pg_dump against the old server version will
restore properly on the new version.  Dumping and restoring the blobs
and re-execute on the new version may fail.  I had not thought of this
issue when this was last discussed, or at least I don't remember
having thought of it, and based on Tom's having endorsed the previous
design, I'm guessing he didn't think of it at the time, either.

I think that Stephen's other points about duplicating data, etc. are
somewhat valid as well, but I am particularly sensitive about dump and
restore hazards.  I don't think you will find any time in the history
of this project when I endorsed any change that would create more of
them or declined to endorse fixing them, and if you do it was probably
in my first year of involvement when I did not understand so well as I
do now how much pain such problems create.  Users are remarkably
skilled at finding these bugs; it's been less then two weeks since we
fixed the most recent one; and they cause a lot of pan.  The only
saving grace is that, up until now, we've pretty much always been able
to patch them by changing pg_dump(all).  The problems that this patch
would create can't be fixed that way, though: you'd have to manually
hack up the blobs stored in the catalog, or manually edit the
dumpfile.  That's not good.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company