Re: 9.2.1 & index-only scans : abnormal heap fetches after VACUUM FULL

On 02/16/2014 10:19 PM, Jim Nasby wrote:
> On 1/24/14, 3:52 PM, Jaime Casanova wrote:
>> On Tue, Dec 3, 2013 at 11:25 AM, Bruce Momjian<bruce(at)momjian(dot)us> wrote:
>>>>
>>>> Is everyone else OK with this approach? Updated patch attached.
>>>>
>> Hi,
>>
>> I started to look at this patch and i found that it fails an assertion
>> as soon as you run a VACUUM FULL after a lazy VACUUM even if those are
>> on unrelated relations. For example in an assert-enabled build with
>> the regression database run:
>>
>> VACUUM customer;
>> [... insert here whatever commands you like or nothing at all ...]
>> VACUUM FULL tenk1;
>
> Is anyone else confused/concerned that regression testing didn't pick this up?

I wouldn't expect that to be explicitly tested - it's pretty unexpected
that they work independently but not when run one after another. But
it's a bit surprising that we don't happen to do that combination in any
of the tests by pure chance.

> The vacuum.sql test does not test lazy vacuum at all, and I can't seem to find any other tests that test lazy vacuum either...

There are several lazy vacuums in the regression suite:

sql/alter_table.sql:vacuum analyze atacc1(a);
sql/alter_table.sql:vacuum analyze atacc1("........pg.dropped.1........");
sql/hs_standby_disallowed.sql:VACUUM hs2;
sql/indirect_toast.sql:VACUUM FREEZE toasttest;
sql/indirect_toast.sql:VACUUM FREEZE toasttest;
sql/matview.sql:VACUUM ANALYZE hogeview;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_add;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_sub;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_div;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_mul;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_sqrt;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_ln;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_log10;
sql/numeric_big.sql:VACUUM ANALYZE num_exp_power_10_ln;
sql/numeric.sql:VACUUM ANALYZE num_exp_add;
sql/numeric.sql:VACUUM ANALYZE num_exp_sub;
sql/numeric.sql:VACUUM ANALYZE num_exp_div;
sql/numeric.sql:VACUUM ANALYZE num_exp_mul;
sql/numeric.sql:VACUUM ANALYZE num_exp_sqrt;
sql/numeric.sql:VACUUM ANALYZE num_exp_ln;
sql/numeric.sql:VACUUM ANALYZE num_exp_log10;
sql/numeric.sql:VACUUM ANALYZE num_exp_power_10_ln;
sql/sanity_check.sql:VACUUM;
sql/without_oid.sql:VACUUM ANALYZE wi;
sql/without_oid.sql:VACUUM ANALYZE wo;

Most of those commands are there to analyze, rather than vacuum, but
lazy vacuum is definitely exercised by the regression tests. I agree
it's quite surprising that vacuum.sql doesn't actually perform any lazy
vacuums.

- Heikki