Hi!
Here is a new version of the patchset with following changes:
- change unit test of page_conversion to address rare cfbot fails (the reason of cfbot to time out was the unit test not written accurate enough);
- fix pg_upgarde on 32-bit systems;
- switch to XidList type (introduced by f10a025cfe97c1a34) for logical replication, abandoning previously used Int64List type;
- use cat version instead of major version as a boundary from 32 to 64 bit xids in pg_upgrade;
- this cat version boundary temporary set to 999999999 for pg_upgrade testing purpose;
- also rebased to the actual master branch.
On 32-bit arch we have noticed pg_upgrade from 32 to 64 bit xids fails due to different TOAST_MAX_CHUNK_SIZE. On a 64-bit xid page we have less available space due to adding a heap page special. This leads to recalc of TOAST_MAX_CHUNK_SIZE.
This was not a problem on 64 bit architectures, as padding bytes on 32-bit xids on TOAST pages were enough to accommodate heap page special of 64-bit xids with TOAST_MAX_CHUNK_SIZE unchanged. On 32 bits architectures padding bytes were not enough and this needed TOAST_MAX_CHUNK_SIZE to be of different size (on 64-bit xids version).
Changes of TOAST_MAX_CHUNK_SIZE lead to being unable to pg_upgrade onto 64-bit xids. This was a real problem, since TOAST of relation requires all chunks to be the same size. In other words, we can not mix TOAST chunks of previous (32-bit xid TOAST pages) with the new one with 64-bit xid TOAST pages.
The solution was to use different specials for TOAST and heap pages. Since, TOAST tuples can not have multixacts and does not need pd_multi_base on page.
Thus, v39 is improved relative to v38 a lot.
As always, feel free to review and share your thoughts on a subject.
--
Best regards,
Maxim Orlov.
