Re: [PoC] Improve dead tuple storage for lazy vacuum

On Wed, Jan 17, 2024 at 9:20 AM John Naylor <johncnaylorls(at)gmail(dot)com> wrote:
>
> On Tue, Jan 16, 2024 at 1:18 PM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> > Just changing "items" to be the local tidstore struct could make the
> > code tricky a bit, since max_bytes and num_items are on the shared
> > memory while "items" is a local pointer to the shared tidstore.
>
> Thanks for trying it this way! I like the overall simplification but
> this aspect is not great.
> Hmm, I wonder if that's a side-effect of the "create" functions doing
> their own allocations and returning a pointer. Would it be less tricky
> if the structs were declared where we need them and passed to "init"
> functions?

Seems worth trying. The current RT_CREATE() API is also convenient as
other data structure such as simplehash.h and dshash.c supports a
similar

>
> That may be a good idea for other reasons. It's awkward that the
> create function is declared like this:
>
> #ifdef RT_SHMEM
> RT_SCOPE RT_RADIX_TREE *RT_CREATE(MemoryContext ctx, Size max_bytes,
> dsa_area *dsa,
> int tranche_id);
> #else
> RT_SCOPE RT_RADIX_TREE *RT_CREATE(MemoryContext ctx, Size max_bytes);
> #endif
>
> An init function wouldn't need these parameters: it could look at the
> passed struct to know what to do.

But the init function would initialize leaf_ctx etc,no? Initializing
leaf_ctx needs max_bytes that is not stored in RT_RADIX_TREE. The same
is true for dsa. I imagined that an init function would allocate a DSA
memory for the control object. So I imagine we will end up still
requiring some of them.

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com