I've been looking into bug #6053, in which Regina Obe complains that
hash-based DISTINCT queries fail for type "citext". The cause is not
far to seek: the header comment for execGrouping.c states
* Note: we currently assume that equality and hashing functions are not
* collation-sensitive, so the code in this file has no support for passing
* collation settings through from callers. That may have to change someday.
and indeed the failure comes directly from the fact that citext's hash
function *does* expect a collation to be passed to it. I'm a bit
embarrassed to not have noticed that citext was a counterexample for
this assumption, especially since I already fixed one bug that should
have clued me in (commit a0b75a41a907e1582acdb8aa6ebb9cacca39d7d8).
Now, removing this assumption from execGrouping.c is already a pretty
sizable task --- for starters, at least plan node types Agg, Group,
SetOp, Unique, and WindowAgg would need collation attributes that they
don't have today. But the assumption that equality operators are not
collation-sensitive is baked into a number of other places too; for
nodeAgg.c @ line 600
indxpath.c @ line 2200
prepunion.c @ line 640
ri_triggers.c @ line 3000
and that's just places where there's a comment about it :-(.
It's worth noting also that in many of these places, paying attention to
collation is not merely going to need more coding; it will directly
translate to a performance hit, one that is entirely unnecessary for the
normal case where collation doesn't affect equality.
So this leaves us between a rock and a hard place. I think there's just
about no chance of fixing all these things without a serious fresh slip
in the 9.1 schedule. Also, I'm *not* prepared to fix these things
personally. I already regret the amount of time I put into collations
this past winter/spring, and am not willing to drop another several
weeks down that sinkhole right now.
The most workable alternative that I can see is to lobotomize citext so
that it always does lower-casing according to the database's "default"
collation, which would allow us to pretend that its notion of equality
is not collation-sensitive after all. We could hope to improve this in
future release cycles, but not till we've done the infrastructure work
outlined above. One bit of infrastructure that might be a good idea is
a flag to indicate whether an equality operator's behavior is
potentially collation-dependent, so that we could avoid taking
performance hits in the normal case.
Comments, other ideas?
regards, tom lane