On 15/10/2025 02:54, Michael Paquier wrote:
On Tue, Oct 14, 2025 at 03:20:16PM +0300, Heikki Linnakangas wrote:
I'm not sure I agree with the premise that we should try to get rid of
RM_HEAP2_ID. There's nothing wrong with that scheme as such. As an
alternative, we could easily teach e.g pg_waldump to treat RM_HEAP_ID and
RM_HEAP2_ID the same for statistics purposes.
Yeah, I'd rather keep heap2 as well. As long as there is more room
for the record IDs, we're still going to need it in the long run.
IMO if we do expand xl_info so that we have a full byte for
RMGR-specific info, then it does make sense to get rid of RM_HEAP2_ID
like this patch does. But if that was the only reason for changing the
WAL record header format, I don't think it's worth it.
This patch consumes one of the padding bytes. That's not entirely free, as
there is an opportunity cost: we could squeeze out the padding bytes and
save 2 bytes on every WAL record instead.
Do you recall an alternative where it would have been possible to save
2 bytes for each record by removing the padding, and still have the
full byte of xl_info be usable freely by each RMGR? I cannot recall
any magic based on how XLogRecord is designed now, but perhaps I
have missed an argument.
I don't know how to do that. My point was that if we *don't* consume a
full byte for RMGR-specific info, we could remove the two padding bytes
instead and make every WAL record 2 bytes smaller.
- Heikki
