Re: Collation versioning - Mailing list pgsql-hackers

From Julien Rouhaud
Subject Re: Collation versioning
Date
Msg-id 20200908201334.GB57691@nol
Whole thread Raw
In response to Re: Collation versioning  (Christoph Berg <myon@debian.org>)
List pgsql-hackers
On Tue, Sep 08, 2020 at 09:33:26PM +0200, Christoph Berg wrote:
> Re: Julien Rouhaud
> > Here's the rationale for this new flag, extracted from the patch's commit
> > message:
> > 
> > Add a new amnostablecollorder flag in IndexAmRoutine.
> > 
> > This flag indicates if the access method does not rely on a stable collation
> > ordering for deterministic collation, i.e. would not be corrupted if the
> > underlying collation library changes its ordering.  This is done this way to
> > make sure that if any external access method isn't updated to correctly setup
> > this flag it will be assumed to rely on a stable collation ordering as this
> > should be the case for the majority of access methods.
> > 
> > This flag will be useful for an upcoming commit that will add detection of
> > possibly corrupted index due to changed collation library version.
> 
> Hmm. Does that flag gain us much? What about non-collation locale
> changes that might still affect indexes like lower() and the citext
> extension? That still depends on locale changes, but that flag
> wouldn't be able to help with "this index is (not) affected by this
> locale change".
> 
> IOW, I think we should aim at simply tracking the version, and leave
> it to the admin (with the help of supplied SQL queries) to either
> rebuild indexes or waive them.
> 
> Or maybe I misunderstood the problem.
> 

I see your point, and indeed this isn't really clear how the flag will be used
given this description.

I guess that my idea is that how exactly an index depends on a stable
collation ordering isn't part of this flag definition, as it should be the same
for any access method.

In the commit that uses the infrastructure, the lack of stable ordering
requirement is only used for index key column, but not for any
expression, whether on index key or qual, because as you mention there's no
guarantee that the expression itself depends on a stable ordering or not.

There could be some improvements done for some simple case (like maybe md5() is
used often enough in index keys that it could be worth to detect), but nothing
in done to attempt that for now.



pgsql-hackers by date:

Previous
From: Tom Lane
Date:
Subject: Re: Issue with cancel_before_shmem_exit while searching to remove a particular registered exit callbacks
Next
From: Alvaro Herrera
Date:
Subject: Re: default partition and concurrent attach partition