On Fri, 5 Jul 2024 at 12:52, David Rowley <dgrowleyml@gmail.com> wrote:
> I propose we change these to uint64 while causing churn in this area,
> probably as a follow-on patch. I think a uint32 isn't wide enough as
> you could exceed the limit with rescans.
I wondered how large a query it would take to cause this problem. I tried:
create table a (a int);
insert into a select x%1000 from generate_Series(1,1500000)x;
create index on a(a);
vacuum freeze analyze a;
set enable_hashjoin=0;
set enable_mergejoin=0;
set enable_indexscan=0;
set max_parallel_workers_per_gather=0;
explain (analyze, costs off, timing off, summary off)
select count(*) from a a1 inner join a a2 on a1.a=a2.a;
After about 15 mins, the trimmed output from Linux is:
Aggregate (actual rows=1 loops=1)
-> Nested Loop (actual rows=2250000000 loops=1)
-> Seq Scan on a a1 (actual rows=1500000 loops=1)
-> Bitmap Heap Scan on a a2 (actual rows=1500 loops=1500000)
Recheck Cond: (a1.a = a)
Heap Blocks: exact=2250000000
-> Bitmap Index Scan on a_a_idx (actual rows=1500 loops=1500000)
Index Cond: (a = a1.a)
Whereas, on MSVC, due to sizeof(long) == 4, it's:
Aggregate (actual rows=1 loops=1)
-> Nested Loop (actual rows=2250000000 loops=1)
-> Seq Scan on a a1 (actual rows=1500000 loops=1)
-> Bitmap Heap Scan on a a2 (actual rows=1500 loops=1500000)
Recheck Cond: (a1.a = a)
-> Bitmap Index Scan on a_a_idx (actual rows=1500 loops=1500000)
Index Cond: (a = a1.a)
Notice the "Heap Blocks: exact=2250000000" is missing on Windows.
This is because it wrapped around to a negative value and
show_tidbitmap_info() only shows > 0 values.
I feel this is a good enough justification to increase the width of
those counters to uint64, so I'll do that too.
David