Arseniy Mukhin <arseniy.mukhin.dev@gmail.com> writes:
> Good point, thanks for the explanation. I forgot that there can be
> many attributes. And I agree, the more determinism in the system, the
> easier it is to work with it and the less room for bugs. OTOH it seems
> from the performance POV we want to have the stricter keys to be the
> first so we do less work and fail fast on the first keys. It looks
> like these two rules (excludeOnly keys LAST and more restrictive keys
> FIRST) are kind of in conflict with each other. I tried to do some
> experiments and it's seems GIN quite sensitive to it, at least in this
> artificial example:
Yeah, it is. I recall seeing some comments to the effect that
optimizing the order of scan keys would be a good thing, but if there
is any code in there that tries to do so, I'm not seeing where.
Seems like a fertile area for future research.
> With applying patch both queries show the same time (second one). So
> currently the user can tune the query by defining more restrictive
> keys first. With the proposed fix it looks like users will have less
> freedom here.
I think most people would consider it a bug if they have to tune the
order of the WHERE clauses manually. The original statement of the
current bug was basically that: it worked in one order and not the
other.
regards, tom lane
