Better index version downloads for the update script

[sgrif]

The query used to grab the next batch of unprocessed version downloads
is now taking on average 9 seconds to run. It can take as much as 30
seconds. I suspect this is because the size of the processed index, or
some other measurement PG is using changed so it decided that index
wasn't the right one to use.

That index wasn't very helpful anyway (it was just the processed
column). We can do better by indexing the id with the where clause.

EXPLAIN ANALYZE results:

Before:

                                                                             QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Limit  (cost=0.09..16204.73 rows=1000 width=21) (actual time=39279.902..39298.408 rows=1000 loops=1)
   ->  Index Scan using version_downloads_pkey on version_downloads  (cost=0.09..199171.32 rows=12291 width=21) (actual time=39279.900..39298.335 rows=1000 loops=1)
         Index Cond: (id > 0)
         Filter: (NOT processed)
         Rows Removed by Filter: 8544073
 Planning time: 0.369 ms
 Execution time: 39298.459 ms
(7 rows)

After:

                                                                                QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Limit  (cost=0.06..95.89 rows=1000 width=21) (actual time=0.016..0.796 rows=1000 loops=1)
   ->  Index Scan using index_version_downloads_not_processed on version_downloads  (cost=0.06..1175.30 rows=12264 width=21) (actual time=0.016..0.732 rows=1000 loops=1)
         Index Cond: (id > 0)
 Planning time: 0.123 ms
 Execution time: 0.854 ms

[ashleygwilliams]

bors: r+

[bors-voyager]

Build succeeded

• continuous-integration/travis-ci/push