Re: Is custom MemoryContext prohibited?

On Mon, Feb 3, 2020 at 7:26 PM Andres Freund <andres(at)anarazel(dot)de> wrote:
> Hm. I kinda like the idea of still having one NodeTag identifying memory
> contexts, and then some additional field identifying the actual
> type. Being able to continue to rely on IsA() etc imo is nice. I think
> nodes.h itself only would be a problem for frontend code because we put
> a lot of other stuff too. We should just separate the actually generic
> stuff out. I think it's going to be like 2 seconds once we have memory
> contexts until we're e.g. going to want to also have pg_list.h - which
> is harder without knowing the tags.

The problem with IsA() is that it assumes that you've got all the node
tags that can ever exist in one big enum. I don't see how to make that
work once you extend the system to work with more than one program. I
think it will be really confusing if frontend code starts reusing
random backend data structures. Like, fundamental things like List,
sure, that should be exposed. But if people start creating Plan or FDW
objects in the frontend, it's just going to be chaos. And I don't
think we want new objects that people may add for frontend code to be
visible to backend code, either.

> It seems like a good idea to still have an additional identifier for
> each node type, for some cross checking. How about just frobbing the
> pointer to the MemoryContextMethod slightly, and storing that in an
> additional field? That'd be something fairly unlikely to ever be a false
> positive, and it doesn't require dereferencing any additional memory.

That would be fine as an internal sanity check, but if Tom is unhappy
with the idea of having to try to make sense of a function pointer,
he's probably going to be even less happy about trying to make sense
of a frobbed pointer. And I would actually agree with him on that
point.

I think we're all pursuing slightly different goals here. KaiGai's
main goal is to make it possible for third-party code to add new kinds
of memory contexts. My main goal is to make memory contexts not depend
on backend-only infrastructure. Tom is concerned about debuggability.
Your concern here is about sanity checking. There's some overlap
between those goals but the absolute best thing for any given one of
them might be really bad for one of the other ones; hopefully we can
find some compromise that gets everybody the things they care about
most.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company