On Fri, Mar 24, 2017 at 3:44 PM, Ashutosh Sharma <ashu.coek88@gmail.com> wrote:
> Agreed. Moreover, previous approach might even change the current
> behaviour of functions like hash_page_items() and hash_page_stats()
> basically when we pass UNUSED pages to these functions. Attached is
> the newer version of patch.
This patch doesn't completely solve the problem:
pgbench -i -s 40
alter table pgbench_accounts drop constraint pgbench_accounts_pkey;
create index on pgbench_accounts using hash (aid);
insert into pgbench_accounts select * from pgbench_accounts;
select hash_page_type, min(n), max(n), count(*) from (select n,
hash_page_type(get_raw_page('pgbench_accounts_aid_idx'::text, n)) from
generate_series(0,
(pg_relation_size('pgbench_accounts_aid_idx')/8192)::integer - 1) n) x
group by 1;
ERROR: index table contains zero page
This happens because of the sparse allocation forced by
_hash_alloc_buckets. Perhaps the real fix is to have that function
initialize all of the pages instead of just the last one, but unless
and until we decide to do that, we've got to cope with zero pages in
the index. Actually, even if we did fix that I think we'd still need
to do this, because the way relation extension works in general is
that we first write a page of zeroes and then go back and fill in the
data; an intervening crash can leave behind the intermediate state.
A second problem is that the one page initialized by
_hash_alloc_buckets needs to end up with a valid special space;
otherwise, you get the same error Jeff complained about originally
when you try to use hash_page_type() on it.
I fixed those issues and committed this.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
