Thanks for the review!
I think that if the anayze command doesn't have the same issue, we
don't need to change it.
Good point, I've wrongly assumed that analyze was also impacted but there's no parallel analyze so the block count is correct.
(a) make lazy vacuum use BufferUsage instead of
VacuumPage{Hit,Miss,Dirty}. (backpatched down to pg13).
(b) make analyze use BufferUsage and remove VacuumPage{Hit,Miss,Dirty}
variables for consistency and simplicity (only for HEAD, if we agree).
I've isolated the fix in the first patch. I've kept the analyze change and removal of VacuumPage* variables split for clarity sake.
BTW I realized that VACUUM VERBOSE running on a temp table always
shows the number of dirtied buffers being 0, which seems to be a bug.
The patch (a) will resolve it as well.
Indeed, it visible with the following test:
SET temp_buffers='1024';
CREATE TEMPORARY TABLE vacuum_fix (aid int, bid int);
INSERT INTO vacuum_fix SELECT *, * FROM generate_series(1, 1000000);
VACUUM vacuum_fix;
UPDATE vacuum_fix SET bid = bid+1;
VACUUM (VERBOSE, INDEX_CLEANUP ON) vacuum_fix;
Pre-patch:
avg read rate: 254.751 MB/s, avg write rate: 0.029 MB/s
buffer usage: 8853 hits, 8856 misses, 1 dirtied
WAL usage: 1 records, 1 full page images, 3049 bytes
The dirtied page is from pg_class (so not a local buffer)
With the fix:
avg read rate: 250.230 MB/s, avg write rate: 250.230 MB/s
buffer usage: 8853 hits, 8856 misses, 8856 dirtied
WAL usage: 1 records, 1 full page images, 3049 bytes