Re: Increasing test coverage of WAL redo functions

On 11/19/2014 05:01 PM, Andres Freund wrote:
> On 2014-11-19 11:54:47 -0300, Alvaro Herrera wrote:
>> Heikki Linnakangas wrote:
>>> Schema | Name | Type | Owner | Size | Description
>>> --------+------------------+-------+--------+---------+-------------
>>> public | btree_tall_tbl | table | heikki | 24 kB |
>>> public | btree_test_tbl | table | heikki | 392 kB |
>>> public | gin_test_tbl | table | heikki | 588 MB |
>>> public | gist_point_tbl | table | heikki | 1056 kB |
>>> public | spgist_point_tbl | table | heikki | 1056 kB |
>>> public | spgist_text_tbl | table | heikki | 1472 kB |
>>> (6 rows)
>>
>> I think it's good to have these tests, though Tom was complaining
>> earlier about the size of the regression test database. Would it work
>> to have this in a separate test suite, like the numeric_big stuff?
>> We can have it run optionally, and perhaps set up a few buildfarm
>> members to exercise them on a regular basis.
>
> I think the tests except the gin one are resonably sized - I'd much
> rather run them all the time. We shouldn't make the buildfarm
> configuration unnecessarily complex.

Agreed.

Committed, except for the large GIN test. I kept a smaller version of
the GIN test, which exercises vacuum and page deletion, but not the
internal page split. I also slimmed down the other tests down a little bit.

This grew "pg_dumpall | wc -c" from 5505689 to 6926066 bytes. The size
of the regression database grew, according to psql's "\l+" command grew
from 45 MB to 57 MB. The amount of WAL generated by "make installcheck"
grew from 75 MB to 104 MB.

That's a quite big increase in all those measures, but I think it's
still acceptable, and I would really like these to run as part of the
regular regression suite.

- Heikki