Postgres' CTE vs Subquery Performance forskel. Hvorfor?

Som @CraigRinger sagde, skulle jeg også have tjekket analysen. Faktisk ser vi fra "forklar analyse", at den første er:

Aggregate  (cost=58406.66..58406.67 rows=1 width=8) (actual time=138191.294..138191.295 rows=1 loops=1)
 CTE subq
   ->  Bitmap Heap Scan on atlas_sezioni2 a  (cost=9.93..51052.46 rows=20900 width=236211) (actual time=2.814..308.667 rows=3705 loops=1)
         Recheck Cond: (ace = 1)
         Filter: ((pro_com = 15146::numeric) AND ((code)::text = ANY ('{11100,11210,11220,11230,11240,11300,12100,14200}'::text[])))
         Rows Removed by Filter: 4
         Heap Blocks: exact=42
         ->  Bitmap Index Scan on atlas_sezioni2_ace_idx  (cost=0.00..9.88 rows=251 width=0) (actual time=0.110..0.110 rows=251 loops=1)
               Index Cond: (ace = 1)
         SubPlan 1
           ->  Limit  (cost=240.70..240.74 rows=15 width=236190) (actual time=0.630..0.636 rows=15 loops=247)
                 ->  Sort  (cost=240.70..240.87 rows=69 width=236190) (actual time=0.627..0.630 rows=15 loops=247)
                       Sort Key: ((a.geom <-> v.geom))
                       Sort Method: top-N heapsort  Memory: 26kB
                       ->  Bitmap Heap Scan on atlas_sezioni2 v  (cost=4.56..239.01 rows=69 width=236190) (actual time=0.045..0.518 rows=73 loops=247)
                             Recheck Cond: ((code)::text = '12230'::text)
                             Filter: (((a.code)::text <> (code)::text) AND (a.pro_com = pro_com))
                             Heap Blocks: exact=6916
                             ->  Bitmap Index Scan on atlas_sezioni2_code_idx  (cost=0.00..4.55 rows=73 width=0) (actual time=0.030..0.030 rows=73 loops=247)
                                   Index Cond: ((code)::text = '12230'::text)
 ->  Unique  (cost=7247.20..7351.70 rows=200 width=72) (actual time=138190.527..138191.243 rows=247 loops=1)
       ->  Sort  (cost=7247.20..7299.45 rows=20900 width=72) (actual time=138190.526..138190.800 rows=3705 loops=1)
             Sort Key: subq.n, (_st_distance(geography(subq.g1), geography(subq.g2), 0::double precision, false))
             Sort Method: quicksort  Memory: 270kB
             ->  CTE Scan on subq  (cost=0.00..5747.50 rows=20900 width=72) (actual time=159.739..138182.891 rows=3705 loops=1)
 Planning time: 2.623 ms
 Execution time: 138217.574 ms
(27 rows)

mens underforespørgslen en er:

Aggregate  (cost=6387362.91..6387362.92 rows=1 width=8) (actual time=140208.005..140208.005 rows=1 loops=1)
 ->  Unique  (cost=6386997.16..6387101.66 rows=20900 width=236230) (actual time=140207.243..140207.947 rows=247 loops=1)
       ->  Sort  (cost=6386997.16..6387049.41 rows=20900 width=236230) (actual time=140207.241..140207.514 rows=3705 loops=1)
             Sort Key: subq.n, (_st_distance(geography(subq.g1), geography(subq.g2), 0::double precision, false))
             Sort Method: quicksort  Memory: 270kB
             ->  Subquery Scan on subq  (cost=9.93..56590.96 rows=20900 width=236230) (actual time=160.784..140199.364 rows=3705 loops=1)
                   ->  Bitmap Heap Scan on atlas_sezioni2 a  (cost=9.93..51052.46 rows=20900 width=236211) (actual time=2.384..308.517 rows=3705 loops=1)
                         Recheck Cond: (ace = 1)
                         Filter: ((pro_com = 15146::numeric) AND ((code)::text = ANY ('{11100,11210,11220,11230,11240,11300,12100,14200}'::text[])))
                         Rows Removed by Filter: 4
                         Heap Blocks: exact=42
                         ->  Bitmap Index Scan on atlas_sezioni2_ace_idx  (cost=0.00..9.88 rows=251 width=0) (actual time=0.150..0.150 rows=251 loops=1)
                               Index Cond: (ace = 1)
                         SubPlan 1
                           ->  Limit  (cost=240.70..240.74 rows=15 width=236190) (actual time=0.640..0.646 rows=15 loops=247)
                                 ->  Sort  (cost=240.70..240.87 rows=69 width=236190) (actual time=0.637..0.640 rows=15 loops=247)
                                       Sort Key: ((a.geom <-> v.geom))
                                       Sort Method: top-N heapsort  Memory: 26kB
                                       ->  Bitmap Heap Scan on atlas_sezioni2 v  (cost=4.56..239.01 rows=69 width=236190) (actual time=0.045..0.527 rows=73 loops=247)
                                             Recheck Cond: ((code)::text = '12230'::text)
                                             Filter: (((a.code)::text <> (code)::text) AND (a.pro_com = pro_com))
                                             Heap Blocks: exact=6916
                                             ->  Bitmap Index Scan on atlas_sezioni2_code_idx  (cost=0.00..4.55 rows=73 width=0) (actual time=0.031..0.031 rows=73 loops=247)
                                                   Index Cond: ((code)::text = '12230'::text)
 Planning time: 1.117 ms
 Execution time: 140208.187 ms

Så det fungerer kun bedre i forklaringen :). Virkelig ydeevne ændrer sig ikke.