Re: Proposal: scan key push down to heap [WIP]

Dilip Kumar <dilipbalaut@gmail.com> writes:
> On Thu, Oct 13, 2016 at 6:05 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>> I seriously doubt that this should EVER be supported for anything
>> other than "var op const", and even then only for very simple
>> operators.

> Yes, with existing key push down infrastructure only "var op const",
> but If we extend that I think we can cover many other simple
> expressions, i.e

I think it is a mistake to let this idea drive any additional
complication of the ScanKey data structure.  That will have negative
impacts on indexscan performance, not to mention require touching
quite a lot of unrelated code.  And as Robert points out, we do not
want to execute anything expensive while holding the buffer LWLock.

>> Part of the trick if we want to make this work is going to be figuring
>> out how we'll identify which operators are safe.

> Yes, I agree that's the difficult part. Can't we process full qual
> list and decide decide the operator are safe or not based on their
> datatype ?

Possibly restricting it to builtin, immutable functions on non-toastable
data types would do.  Or for more safety, only allow pass-by-value data
types.  But I have a feeling that there might still be counterexamples.
        regards, tom lane