Hi Kirill,
Following up on my earlier note, I implemented the proposed HASH overflow page reuse enhancement.Recently freed overflow pages are recorded in
_hash_freeovflpage( ),and _hash_addovflpage( ) now prefers reusing those pages during allocation before falling back to the bitmap scan.
The change is backend-local and allocation-only,with no WAL or on-disk format changes.I verified correctness using index build/drop and VACUUM cycles,and confirmed WAL neutrality by comparing WAL generated during HASH index builds with and without this change (no observable difference beyond normal noise).
The patch is attached for review.Feedback is welcome.
Best regards,
Lakshmi
Hi Kirill,
I tested your patch on the current master and confirmed the WAL reduction during HASH index build.
While testing, I noticed a possible small follow-up improvement in HASH overflow handling. Currently, any free overflow page may be reused, which can scatter overflow chains and hurt cache locality. Reusing recently freed overflow pages first could help, without changing WAL behavior or on-disk format.
I would like to work on this as a follow-up enhancement and would welcome any suggestions.
Best regards,
Lakshmi
There exists an optimization to index creation process, when we omit
to write any WAL
for index build. It is currently supported in B Tree, GIN, GiST, spg indexes.
It works because we do not need to recover anything if index creation
fails, because if was not used by any query. So, the index can be
built on-disk, and then, just before making the index alive, we can
simply log all pages to WAL.
Hash index currently lacks this optimization.
PFA implementation.
During my testing, I checked the amount of WAL generated by index
build before and after patch applied. My script was something like:
select pg_current_wal_insert_lsn();
create index on t using hash (i);
select pg_current_wal_insert_lsn();
select pg_lsn_wal_diff(lsn1, lsn2);
Resulting numbers depend on index size, but I got 2.5-3.5 times less
WAL with this patch and 8 times less WAL with this patch +
wal_compression=on.
Index creation time, however, did not change much...
About implementation:
These are many types of record that can be generated during index build.
I know for sure these are possible (double-checked using pg_waldump):
SPLIT_COMPLETE
INSERT
SPLIT_ALLOCATE_PAGE
SPLIT_PAGE
ADD_OVFL_PAGE
SQUEEZE_PAGE
INIT_META_PAGE
INIT_BITMAP_PAGE
Looks like SPLIT_COMPLETE and VACUUM_ONE_PAGE are never generated
during index build. I'm not sure about MOVE_PAGE_CONTENTS.
So, implementation is simply pass isbuild flag everywhere something is
wal-logged. Looks like it is less invasive than alternatives.
--
Best regards,
Kirill Reshke
