[harikb]

I think “extension” covers any custom code (even functions that need to import a c library), not just major db features. I have had to do it recently [1] to customize partitioning function to be the way I like. It would be great to do it in a friendlier language

[1] https://harikb.github.io/posts/postgres-11-hash-partitioning...

[Hendrikto]

> It would be great to do it in a friendlier language

Whether Rust is "friendlier" than C depends a lot on your definition of what that means.

[zombodb]

You're not wrong, but barring bugs in `pgx` (of which I'm sure there are plenty right now), at least Rust gives you compile-time guarantees around not crashing.

And when running inside your database process, that's a huge win.

[zombodb]

As a follow-up. Here's what Postgres' "generate_series" function looks like with pgx: https://github.com/zombodb/pgx/blob/0803bd3aa2f8ef2bd80d2e03...

And here's what you'd have to do to implement it in C: https://github.com/postgres/postgres/blob/dad75eb4a8d5835ecc...

[zombodb]

One more follow-up...

The top one is pgx, the bottom is Postgres. So there's a little room for improvement here with pgx, but that's okay for a v0.0.3 release.

    test=# select count(*) from srf.generate_series(1, 10000000);
    Time: 1552.115 ms (00:01.552)

    test=# select count(*) from generate_series(1, 10000000);
    Time: 1406.357 ms (00:01.406)

[anarazel]

The largest part of the time executing the above query isn't inside the function, so this isn't that a material comparison. The reason for that is that SRFs in FROM to be materialized into a tuplestore, which isn't free:

    postgres[607045][1]=# SELECT count(*) FROM generate_series(1, 10000000);
    ┌──────────┐
    │  count   │
    ├──────────┤
    │ 10000000 │
    └──────────┘
    (1 row)

    Time: 1249.224 ms (00:01.249)

    postgres[607045][1]=# SELECT count(*) FROM (SELECT generate_series(1, 10000000)) s;
    ┌──────────┐
    │  count   │
    ├──────────┤
    │ 10000000 │
    └──────────┘
    (1 row)

    Time: 460.206 ms

For mostly historical reasons SRFs in the target list can use the non-materializing SRF query protocol, but SRFs in the FROM list can't.

Any chance you could show the timings for the pgx version of the second query?

[zombodb]

Sure! Top is Postgres, bottom is pgx, after running each 5 times...

    test=# SELECT count(*) FROM (SELECT generate_series(1, 10000000)) s;
      count   
    ----------
     10000000
    (1 row)

    Time: 399.630 ms
    test=# SELECT count(*) FROM (SELECT srf.generate_series(1, 10000000)) s;
      count   
    ----------
     10000000
    (1 row)

    Time: 478.194 ms

Thanks for the reply. I'm not surprised there's room for optimization in pgx, especially in the Iterator-->SRF path.

edit: PS, thanks for the idea of putting what I assume is the backend pid in psql's prompt. I need to go figure out how to do that right now!

edit edit: hmm, I guess that's not the PID.

[dwaltrip]

In the function definition of `generate_series`, what is going on with the `default!` macro?

That isn't some sort of way of actually getting real default param values is it? I thought that wasn't possible in Rust.

[db48x]

That is certainly nicer.

[chc]

Does Rust give you compile-time guarantees around not crashing? I thought any function could panic without so much as a peep from the compiler.

[zombodb]

Fair. With pgx, however, Rust "panic!"s are translated into standard Postgres "ERROR"s, such that instead of crashing, only the current transaction aborts.

So while you're pretty much correct in general, pgx handles it in the way a PG extension author would expect.

[derefr]

One might say that Rust is friendlier downstream, to the ops people running the compiled binary :)

[lmm]

Really? I struggle to imagine a case for claiming that C is friendlier.

[petergeoghegan]

The ecosystem of preexisting C extensions to generalize from, for one