Re: Inefficiency in parallel pg_restore with many tables

On Fri, Sep 01, 2023 at 01:52:48PM -0700, Nathan Bossart wrote:
> On Fri, Sep 01, 2023 at 04:00:44PM -0400, Robert Haas wrote:
>> In hindsight, I think that making binaryheap depend on Datum was a bad
>> idea. I think that was my idea, and I think it wasn't very smart.
>> Considering that people have coded to that decision up until now, it
>> might not be too easy to change at this point. But in principle I
>> guess you'd want to be able to make a heap out of any C data type,
>> rather than just Datum, or just Datum in the backend and just void *
>> in the frontend.
>
> Yeah, something similar to simplehash for binary heaps could be nice. That
> being said, I don't know if there's a strong reason to specialize the
> implementation for a given C data type in most cases. I suspect many
> callers are just fine with dealing with pointers (e.g., I wouldn't store an
> entire TocEntry in the array), and smaller types like integers are already
> stored directly in the array thanks to the use of Datum. However, it
> _would_ allow us to abandon this frontend/backend void */Datum kludge,
> which is something.

I ended up hacking together a (nowhere near committable) patch to see how
hard it would be to allow using any type with binaryheap. It doesn't seem
too bad.

--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com