Hacker News new | ask | show | jobs
by Alexss 3780 days ago
It's mentioned, albeit in the second-last line of the whole presentation, as a possible reaction to GPU-based sorting in the future.
1 comments
Ono-Sendai 3780 days ago
Ok. Because you can radix sort on CPU and it's a lot faster than e.g. quicksort in my experience.
link
zkzkz 3779 days ago
Postgres has an extensible type system so Radix sort isn't really suitable, it really needs a comparison sort.

For SSE/MMX/AVX/etc or GPU sorts it would have to use integer or floating point surrogate keys. It would probably be the same abbreviated keys mentioned in the last few slides. That could be a radix sort but generally I think bitonic sort is preferred.

link
Jweb_Guru 3777 days ago
> Postgres has an extensible type system so Radix sort isn't really suitable, it really needs a comparison sort.

That's not quite true. It supports string tries, for example (using SP-GIST). Such operations just have to be built for specific types. I think it would be more accurate to say that (AFAICT) Postgres doesn't currently have any way to extend the optimizer with custom plans (though it's been a very popular feature suggestion).

link
anarazel 3777 days ago
> I think it would be more accurate to say that (AFAICT) Postgres doesn't currently have any way to extend the optimizer with custom plans (though it's been a very popular feature suggestion).

There's http://www.postgresql.org/docs/devel/static/custom-scan.html

link