Re: Non-replayable WAL records through overflows and >MaxAllocSize lengths

Hi,

On 2022-07-15 11:25:54 +0200, Matthias van de Meent wrote:
> On Mon, 14 Mar 2022 at 18:14, Andres Freund <andres(at)anarazel(dot)de> wrote:
> >
> > A random thought I had while thinking about the size limits: We could use the
> > low bits of the length and xl_prev to store XLR_SPECIAL_REL_UPDATE |
> > XLR_CHECK_CONSISTENCY and give rmgrs the full 8 bit of xl_info. Which would
> > allow us to e.g. get away from needing Heap2. Which would aestethically be
> > pleasing.
>
> I just remembered your comment while going through the xlog code and
> thought this about the same issue: We still have 2 bytes of padding in
> XLogRecord, between xl_rmid and xl_crc. Can't we instead use that
> space for rmgr-specific flags, as opposed to stealing bits from
> xl_info?

Sounds like a good idea to me. I'm not sure who is stealing bits from what
right now, but it clearly seems worthwhile to separate "flags" from "record
type within rmgr".

I think we should split it at least into three things:

1) generic per-record flags for xlog machinery (ie. XLR_SPECIAL_REL_UPDATE, XLR_CHECK_CONSISTENCY)
2) rmgr record type identifier (e.g. XLOG_HEAP_*)
2) rmgr specific flags (e.g. XLOG_HEAP_INIT_PAGE)

Greetings,

Andres Freund