PostgreSQL in 的优化器处理以及如何优化
在使用数据库的过程中，经常会遇到需要匹配多个值的情况。
通常的写法包括：
-- select * from table where id = any(array);
-- select * from table where id in (values);
-- select * from table where id=x or id=x or ....;
-- select * from table where id in (query);
-- select * from table where id in ( values query );
-- select * from table t1 join (query or values query) t2 on t1.id=t2.id; 每种写法会产生多种执行计划的可能，如下：
-- select * from table where id = any(array);
优化器可以使用index scan, bitmap scan, seq scan.
-- select * from table where id in (values);
优化器可以使用index scan, bitmap scan, seq scan.
-- select * from table where id=x or id=x or ....;
优化器可以使用bitmap scan + BitmapOr, seq scan.
-- select * from table where id in (query);
优化器可以使用join (merge,hash,nest).
-- select * from table where id in ( values query );
优化器可以使用join (merge,hash,nest).
-- select * from table t1 join (query or values query) t2 on t1.id=t2.id;
优化器可以使用join (merge,hash,nest). SQL优化策略是尽量减少CPU的运算以及page的扫描数量。
下面针对每种SQL，看看对应的可能的执行计划有什么差别，（使用开关来控制执行计划的选择, 如set enable_indexscan=off）
支持的开关如下：
enable_bitmapscan enable_hashjoin enable_indexscan enable_mergejoin enable_seqscan enable_tidscan
enable_hashagg enable_indexonlyscan enable_material enable_nestloop enable_sort 开始测试，使用auto_explain输出执行计划：
load 'auto_explain';
set auto_explain.log_analyze =true;
set auto_explain.log_buffers =true;
set auto_explain.log_nested_statements=true;
set auto_explain.log_timing=true;
set auto_explain.log_triggers=true;
set auto_explain.log_verbose=true;
set auto_explain.log_min_duration=0;
set client_min_messages ='log';
set work_mem='8GB'; 测试SQL写法1：
-- select * from table where id = any(array);
do language plpgsql
$$
declare
v_id int[];
begin
select array_agg(trunc(random()*100000)) into v_id from generate_series(1,200) t(id);
perform * from t_in_test where id = any (v_id);
end;
$$
; 优化器选择1 (index scan)：
离散扫描，适合小的扫描集。
LOG: duration: 2.312 ms plan:
Query Text: SELECT * from t_in_test where id = any (v_id)
Index Scan using t_in_test_pkey on public.t_in_test (cost=0.43..895.50 rows=200 width=37) (actual time=0.039..2.266 rows=200 loops=1)
Output: id, info
Index Cond: (t_in_test.id = ANY ('{50836,73414,41071,45604,...省略部分...,76236}'::integer[]))
Buffers: shared hit=776
CONTEXT: SQL statement "SELECT * from t_in_test where id = any (v_id)"
PL/pgSQL function inline_code_block line 6 at PERFORM 优化器选择2 (bitmap scan)：
比index scan多了Recheck的开销，以及按照ctid排序的开销。
适合大的扫描集，排序的目的是减少离散扫描，还可以用到块设备的prefetch。
LOG: duration: 1.602 ms plan:
Query Text: SELECT * from t_in_test where id = any (v_id)
Bitmap Heap Scan on public.t_in_test (cost=888.55..1711.16 rows=200 width=37) (actual time=0.880..1.563 rows=200 loops=1)
Output: id, info
Recheck Cond: (t_in_test.id = ANY ('{32635,31123,6282,59640,...省略部分...,87705}'::integer[]))
Heap Blocks: exact=184
Buffers: shared hit=784
-> Bitmap Index Scan on t_in_test_pkey (cost=0.00..888.50 rows=200 width=0) (actual time=0.846..0.846 rows=200 loops=1)
Index Cond: (t_in_test.id = ANY ('{32635,31123,6282,59640,...省略部分...,87705}'::integer[]))
Buffers: shared hit=600
CONTEXT: SQL statement "SELECT * from t_in_test where id = any (v_id)"
PL/pgSQL function inline_code_block line 6 at PERFORM 优化器选择3 (seq scan)：
适合非常庞大的扫描集。
LOG: duration: 19940.394 ms plan:
Query Text: SELECT * from t_in_test where id = any (v_id)
Seq Scan on public.t_in_test (cost=0.00..2683354.80 rows=200 width=37) (actual time=4.237..19940.330 rows=199 loops=1)
Output: id, info
Filter: (t_in_test.id = ANY ('{45867,72450,95153,86233,63073,11016,56010,47158,...省略部分...,90444}'::integer[]))
Rows Removed by Filter: 9999801
Buffers: shared hit=83334
CONTEXT: SQL statement "SELECT * from t_in_test where id = any (v_id)"
PL/pgSQL function inline_code_block line 6 at PERFORM 测试SQL写法2：
-- select * from table where id in (values);
do language plpgsql
$$
declare
v_where text;
begin
select string_agg(id::text,',') into v_where from (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t;
execute 'select * from t_in_test where id in ('||v_where||')';
end;
$$
; 优化器选择1 (index scan)：
LOG: duration: 0.919 ms plan:
Query Text: select * from t_in_test where id in (8826,2038,72163,29843,76886,37893,5279,64308,...省略部分...,48126,44868)
Index Scan using t_in_test_pkey on public.t_in_test (cost=0.43..895.50 rows=200 width=37) (actual time=0.017..0.894 rows=200 loops=1)
Output: id, info
Index Cond: (t_in_test.id = ANY ('{8826,2038,72163,29843,76886,37893,5279,64308,7370,80216,...省略部分...,48126,44868}'::integer[]))
Buffers: shared hit=779
CONTEXT: SQL statement "select * from t_in_test where id in (8826,2038,72163,29843,76886,37893,5279,64308,7370,80216,...省略部分...,73366,48126,44868)"
PL/pgSQL function inline_code_block line 6 at EXECUTE 优化器选择2 (bitmap scan)：
LOG: duration: 1.012 ms plan:
Query Text: select * from t_in_test where id in (17424,80517,35148,38245,93037,...省略部分...,14997,34639,10646)
Bitmap Heap Scan on public.t_in_test (cost=888.55..1711.16 rows=200 width=37) (actual time=0.657..0.978 rows=200 loops=1)
Output: id, info
Recheck Cond: (t_in_test.id = ANY ('{17424,80517,35148,38245,93037,4516,...省略部分...,14997,34639,10646}'::integer[]))
Heap Blocks: exact=177
Buffers: shared hit=779
-> Bitmap Index Scan on t_in_test_pkey (cost=0.00..888.50 rows=200 width=0) (actual time=0.629..0.629 rows=200 loops=1)
Index Cond: (t_in_test.id = ANY ('{17424,80517,35148,38245,93037,4516,27690,...省略部分...,34639,10646}'::integer[]))
Buffers: shared hit=602
CONTEXT: SQL statement "select * from t_in_test where id in (17424,80517,35148,38245,93037,4516,27690,48978,11902,...省略部分...,34639,10646)"
PL/pgSQL function inline_code_block line 6 at EXECUTE 优化器选择3 (seq scan)：
LOG: duration: 19678.014 ms plan:
Query Text: select * from t_in_test where id in (77056,1340,73056,42536,6862,44702,64810,42774,...省略部分...,24083,11322)
Seq Scan on public.t_in_test (cost=0.00..2683354.80 rows=200 width=37) (actual time=2.045..19677.975 rows=200 loops=1)
Output: id, info
Filter: (t_in_test.id = ANY ('{77056,1340,73056,42536,6862,...省略部分...,24083,11322}'::integer[]))
Rows Removed by Filter: 9999800
Buffers: shared hit=83334
CONTEXT: SQL statement "select * from t_in_test where id in (77056,1340,73056,42536,6862,44702,...省略部分...,24083,11322)"
PL/pgSQL function inline_code_block line 6 at EXECUTE 测试SQL写法3：
-- select * from table where id=x or id=x or ....;
do language plpgsql
$$
declare
v_where text := 'id=';
v int;
begin
for v in select trunc(random()*100000)::int from generate_series(1,200) t(id)
loop
v_where := ' '|| v_where ||v||' or id=';
end loop;
v_where := rtrim(v_where,'or id=');
execute 'select * from t_in_test where '||v_where;
end;
$$
; 优化器选择1 (bitmapindex scan + bitmapor)：
使用or的写法，只能选择bitmap index scan，所以不如使用IN的写法。
LOG: duration: 1.085 ms plan:
Query Text: select * from t_in_test where id=29207 or id=69918 or id=4044 or ...省略部分... or id=53009 or id=28015 or id=11763
Bitmap Heap Scan on public.t_in_test (cost=898.50..1771.11 rows=200 width=37) (actual time=0.754..1.043 rows=200 loops=1)
Output: id, info
Recheck Cond: ((t_in_test.id = 29207) OR (t_in_test.id = 69918) OR (t_in_test.id = 4044) OR (t_in_test.id = 65838) OR ...省略部分... OR (t_in_test.id = 28015) OR (t_in_test.id = 11763))
Heap Blocks: exact=180
Buffers: shared hit=781
-> BitmapOr (cost=898.50..898.50 rows=200 width=0) (actual time=0.725..0.725 rows=0 loops=1)
Buffers: shared hit=601
-> Bitmap Index Scan on t_in_test_pkey (cost=0.00..4.44 rows=1 width=0) (actual time=0.020..0.020 rows=1 loops=1)
Index Cond: (t_in_test.id = 29207)
Buffers: shared hit=3
-> Bitmap Index Scan on t_in_test_pkey (cost=0.00..4.44 rows=1 width=0) (actual time=0.011..0.011 rows=1 loops=1)
Index Cond: (t_in_test.id = 69918)
Buffers: shared hit=3
.....省略部分
-> Bitmap Index Scan on t_in_test_pkey (cost=0.00..4.44 rows=1 width=0) (actual time=0.004..0.004 rows=1 loops=1)
Index Cond: (t_in_test.id = 11763)
Buffers: shared hit=3
CONTEXT: SQL statement "select * from t_in_test where id=29207 or id=69918 or id=4044 or ...省略部分... or id=28015 or id=11763"
PL/pgSQL function inline_code_block line 11 at EXECUTE 优化器选择2 (seq scan)：
LOG: duration: 107484.074 ms plan:
Query Text: select * from t_in_test where id=51946 or id=17129 or id=90027 or ...省略部分... or id=22127 or id=62334 or id=11722
Seq Scan on public.t_in_test (cost=0.00..5183374.80 rows=200 width=37) (actual time=17.394..107483.942 rows=199 loops=1)
Output: id, info
Filter: ((t_in_test.id = 51946) OR (t_in_test.id = 17129) OR (t_in_test.id = 90027) OR ...省略部分... OR (t_in_test.id = 62334) OR (t_in_test.id = 11722))
Rows Removed by Filter: 9999801
Buffers: shared hit=83334
CONTEXT: SQL statement "select * from t_in_test where id=51946 or id=17129 or id=90027 or ...省略部分... or id=62334 or id=11722"
PL/pgSQL function inline_code_block line 11 at EXECUTE 测试SQL写法4：
-- select * from table where id in (query);
do language plpgsql
$$
declare
begin
perform * from t_in_test where id in (select trunc(random()*100000)::int as id from generate_series(1,200) t(id));
end;
$$
; 优化器选择1 (nestloop join)：
适合小的扫描集，并且其中有一个表的JOIN列是带有主键或唯一约束的。
LOG: duration: 1.314 ms plan:
Query Text: SELECT * from t_in_test where id in (select trunc(random()*100000)::int as id from generate_series(1,200) t(id))
Nested Loop (cost=32.94..1727.00 rows=5000040 width=37) (actual time=0.166..1.226 rows=200 loops=1)
Output: t_in_test.id, t_in_test.info
Buffers: shared hit=800
-> HashAggregate (cost=32.50..34.50 rows=200 width=4) (actual time=0.149..0.189 rows=200 loops=1)
Output: ((trunc((random() * '100000'::double precision)))::integer)
Group Key: (trunc((random() * '100000'::double precision)))::integer
-> Function Scan on pg_catalog.generate_series t (cost=0.00..20.00 rows=1000 width=0) (actual time=0.033..0.089 rows=200 loops=1)
Output: (trunc((random() * '100000'::double precision)))::integer
Function Call: generate_series(1, 200)
-> Index Scan using t_in_test_pkey on public.t_in_test (cost=0.43..8.45 rows=1 width=37) (actual time=0.004..0.005 rows=1 loops=200)
Output: t_in_test.id, t_in_test.info
Index Cond: (t_in_test.id = ((trunc((random() * '100000'::double precision)))::integer))
Buffers: shared hit=800
CONTEXT: SQL statement "SELECT * from t_in_test where id in (select trunc(random()*100000)::int as id from generate_series(1,200) t(id))"
PL/pgSQL function inline_code_block line 4 at PERFORM 优化器选择2 (hash join)：
适合大的扫描集，同时两个表的JOIN列上面都没有索引的情况。
LOG: duration: 2454.400 ms plan:
Query Text: SELECT * from t_in_test where id in (select trunc(random()*100000)::int as id from generate_series(1,200) t(id))
Hash Join (cost=37.00..220874.10 rows=5000040 width=37) (actual time=0.413..2454.343 rows=200 loops=1)
Output: t_in_test.id, t_in_test.info
Hash Cond: (t_in_test.id = ((trunc((random() * '100000'::double precision)))::integer))
Buffers: shared hit=83334
-> Seq Scan on public.t_in_test (cost=0.00..183334.80 rows=10000080 width=37) (actual time=0.010..1182.626 rows=10000000 loops=1)
Output: t_in_test.id, t_in_test.info
Buffers: shared hit=83334
-> Hash (cost=34.50..34.50 rows=200 width=4) (actual time=0.221..0.221 rows=200 loops=1)
Output: ((trunc((random() * '100000'::double precision)))::integer)
Buckets: 1024 Batches: 1 Memory Usage: 16kB
-> HashAggregate (cost=32.50..34.50 rows=200 width=4) (actual time=0.149..0.177 rows=200 loops=1)
Output: ((trunc((random() * '100000'::double precision)))::integer)
Group Key: (trunc((random() * '100000'::double precision)))::integer
-> Function Scan on pg_catalog.generate_series t (cost=0.00..20.00 rows=1000 width=0) (actual time=0.033..0.088 rows=200 loops=1)
Output: (trunc((random() * '100000'::double precision)))::integer
Function Call: generate_series(1, 200)
CONTEXT: SQL statement "SELECT * from t_in_test where id in (select trunc(random()*100000)::int as id from generate_series(1,200) t(id))"
PL/pgSQL function inline_code_block line 4 at PERFORM 优化器选择3 (merge join)：
适合大的扫描集，并且两个表的JOIN列都有索引。
LOG: duration: 32.551 ms plan:
Query Text: SELECT * from t_in_test where id in (select trunc(random()*100000)::int as id from generate_series(1,200) t(id))
Merge Join (cost=42.58..368067.98 rows=5000040 width=37) (actual time=0.561..32.497 rows=200 loops=1)
Output: t_in_test.id, t_in_test.info
Merge Cond: (t_in_test.id = ((trunc((random() * '100000'::double precision)))::integer))
Buffers: shared hit=1112
-> Index Scan using t_in_test_pkey on public.t_in_test (cost=0.43..343022.64 rows=10000080 width=37) (actual time=0.016..20.499 rows=99905 loops=1)
Output: t_in_test.id, t_in_test.info
Buffers: shared hit=1108
-> Sort (cost=42.15..42.65 rows=200 width=4) (actual time=0.268..0.296 rows=200 loops=1)
Output: ((trunc((random() * '100000'::double precision)))::integer)
Sort Key: ((trunc((random() * '100000'::double precision)))::integer)
Sort Method: quicksort Memory: 34kB
Buffers: shared hit=4
-> HashAggregate (cost=32.50..34.50 rows=200 width=4) (actual time=0.148..0.181 rows=200 loops=1)
Output: ((trunc((random() * '100000'::double precision)))::integer)
Group Key: (trunc((random() * '100000'::double precision)))::integer
-> Function Scan on pg_catalog.generate_series t (cost=0.00..20.00 rows=1000 width=0) (actual time=0.035..0.078 rows=200 loops=1)
Output: (trunc((random() * '100000'::double precision)))::integer
Function Call: generate_series(1, 200)
CONTEXT: SQL statement "SELECT * from t_in_test where id in (select trunc(random()*100000)::int as id from generate_series(1,200) t(id))"
PL/pgSQL function inline_code_block line 4 at PERFORM 测试SQL写法5：
-- select * from table where id in ( values query );
do language plpgsql
$$
declare
v_values text := 'values ( ';
v int;
begin
for v in select trunc(random()*100000)::int from generate_series(1,200) t(id)
loop
v_values := v_values ||v||'),(';
end loop;
v_values := rtrim( v_values,',(' );
execute 'select * from t_in_test where id in ( select * from ('||v_values||') as t(id))';
end;
$$
; 优化器选择1 (nestloop join)：
LOG: duration: 1.272 ms plan:
Query Text: select * from t_in_test where id in ( select * from (values ( 96474),(39030),(12481),(60519),...省略部分...,(23783),(9253)) as t(id))
Nested Loop (cost=3.44..1697.50 rows=5000040 width=37) (actual time=0.130..1.195 rows=200 loops=1)
Output: t_in_test.id, t_in_test.info
Buffers: shared hit=802
-> HashAggregate (cost=3.00..5.00 rows=200 width=4) (actual time=0.105..0.143 rows=200 loops=1)
Output: "*VALUES*".column1
Group Key: "*VALUES*".column1
-> Values Scan on "*VALUES*" (cost=0.00..2.50 rows=200 width=4) (actual time=0.001..0.040 rows=200 loops=1)
Output: "*VALUES*".column1
-> Index Scan using t_in_test_pkey on public.t_in_test (cost=0.43..8.45 rows=1 width=37) (actual time=0.004..0.005 rows=1 loops=200)
Output: t_in_test.id, t_in_test.info
Index Cond: (t_in_test.id = "*VALUES*".column1)
Buffers: shared hit=802
CONTEXT: SQL statement "select * from t_in_test where id in ( select * from (values ( 96474),(39030),(12481),(60519),(70354),(33117),...省略部分...,(15818),(23783),(9253)) as t(id))"
PL/pgSQL function inline_code_block line 11 at EXECUTE 优化器选择2 (hash join)：
LOG: duration: 2444.648 ms plan:
Query Text: select * from t_in_test where id in ( select * from (values ( 95286),(76612),(56400),(99838),(2155),...省略部分...,(29527),(99252)) as t(id))
Hash Join (cost=7.50..220844.60 rows=5000040 width=37) (actual time=0.222..2444.573 rows=200 loops=1)
Output: t_in_test.id, t_in_test.info
Hash Cond: (t_in_test.id = "*VALUES*".column1)
Buffers: shared hit=83334
-> Seq Scan on public.t_in_test (cost=0.00..183334.80 rows=10000080 width=37) (actual time=0.009..1174.724 rows=10000000 loops=1)
Output: t_in_test.id, t_in_test.info
Buffers: shared hit=83334
-> Hash (cost=5.00..5.00 rows=200 width=4) (actual time=0.173..0.173 rows=200 loops=1)
Output: "*VALUES*".column1
Buckets: 1024 Batches: 1 Memory Usage: 16kB
-> HashAggregate (cost=3.00..5.00 rows=200 width=4) (actual time=0.101..0.135 rows=200 loops=1)
Output: "*VALUES*".column1
Group Key: "*VALUES*".column1
-> Values Scan on "*VALUES*" (cost=0.00..2.50 rows=200 width=4) (actual time=0.001..0.042 rows=200 loops=1)
Output: "*VALUES*".column1
CONTEXT: SQL statement "select * from t_in_test where id in ( select * from (values ( 95286),(76612),(56400),...省略部分...,(29527),(99252)) as t(id))"
PL/pgSQL function inline_code_block line 11 at EXECUTE 优化器选择3 (merge join)：
LOG: duration: 32.296 ms plan:
Query Text: select * from t_in_test where id in ( select * from (values ( 18704),(70725),(55056),...省略部分...,(80068),(28737)) as t(id))
Merge Semi Join (cost=10.58..368035.98 rows=5000040 width=37) (actual time=0.560..32.212 rows=200 loops=1)
Output: t_in_test.id, t_in_test.info
Merge Cond: (t_in_test.id = "*VALUES*".column1)
Buffers: shared hit=1110
-> Index Scan using t_in_test_pkey on public.t_in_test (cost=0.43..343022.64 rows=10000080 width=37) (actual time=0.023..20.733 rows=99962 loops=1)
Output: t_in_test.id, t_in_test.info
Buffers: shared hit=1110
-> Sort (cost=10.14..10.64 rows=200 width=4) (actual time=0.105..0.134 rows=200 loops=1)
Output: "*VALUES*".column1
Sort Key: "*VALUES*".column1
Sort Method: quicksort Memory: 34kB
-> Values Scan on "*VALUES*" (cost=0.00..2.50 rows=200 width=4) (actual time=0.002..0.035 rows=200 loops=1)
Output: "*VALUES*".column1
CONTEXT: SQL statement "select * from t_in_test where id in ( select * from (values ( 18704),(70725),(55056),...省略部分...,(28737)) as t(id))"
PL/pgSQL function inline_code_block line 11 at EXECUTE 测试SQL写法6：
-- select * from table t1 join (query or values query) t2 on t1.id=t2.id;
do language plpgsql
$$
declare
begin
perform * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id);
end;
$$
; 优化器选择1 (nestloop join)：
LOG: duration: 1.327 ms plan:
Query Text: SELECT * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id)
Nested Loop (cost=0.44..8404.50 rows=1000 width=41) (actual time=0.062..1.241 rows=200 loops=1)
Output: t1.id, t1.info, ((trunc((random() * '100000'::double precision)))::integer)
Buffers: shared hit=802
-> Function Scan on pg_catalog.generate_series t (cost=0.00..20.00 rows=1000 width=0) (actual time=0.034..0.131 rows=200 loops=1)
Output: (trunc((random() * '100000'::double precision)))::integer
Function Call: generate_series(1, 200)
-> Index Scan using t_in_test_pkey on public.t_in_test t1 (cost=0.43..8.36 rows=1 width=37) (actual time=0.005..0.005 rows=1 loops=200)
Output: t1.id, t1.info
Index Cond: (t1.id = ((trunc((random() * '100000'::double precision)))::integer))
Buffers: shared hit=802
CONTEXT: SQL statement "SELECT * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id)"
PL/pgSQL function inline_code_block line 4 at PERFORM 优化器选择2 (hash join)：
LOG: duration: 4883.088 ms plan:
Query Text: SELECT * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id)
Hash Join (cost=308335.80..308390.80 rows=1000 width=41) (actual time=4882.749..4883.023 rows=200 loops=1)
Output: t1.id, t1.info, ((trunc((random() * '100000'::double precision)))::integer)
Hash Cond: (((trunc((random() * '100000'::double precision)))::integer) = t1.id)
Buffers: shared hit=83334
-> Function Scan on pg_catalog.generate_series t (cost=0.00..20.00 rows=1000 width=0) (actual time=0.033..0.125 rows=200 loops=1)
Output: (trunc((random() * '100000'::double precision)))::integer
Function Call: generate_series(1, 200)
-> Hash (cost=183334.80..183334.80 rows=10000080 width=37) (actual time=4767.895..4767.895 rows=10000000 loops=1)
Output: t1.id, t1.info
Buckets: 16777216 Batches: 1 Memory Usage: 804901kB
Buffers: shared hit=83334
-> Seq Scan on public.t_in_test t1 (cost=0.00..183334.80 rows=10000080 width=37) (actual time=0.014..1325.338 rows=10000000 loops=1)
Output: t1.id, t1.info
Buffers: shared hit=83334
CONTEXT: SQL statement "SELECT * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id)"
PL/pgSQL function inline_code_block line 4 at PERFORM 优化器选择3 (merge join)：
LOG: duration: 32.505 ms plan:
Query Text: SELECT * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id)
Merge Join (cost=80.27..368117.67 rows=1000 width=41) (actual time=0.182..32.429 rows=200 loops=1)
Output: t1.id, t1.info, ((trunc((random() * '100000'::double precision)))::integer)
Merge Cond: (t1.id = ((trunc((random() * '100000'::double precision)))::integer))
Buffers: shared hit=1102
-> Index Scan using t_in_test_pkey on public.t_in_test t1 (cost=0.43..343022.64 rows=10000080 width=37) (actual time=0.022..20.782 rows=99360 loops=1)
Output: t1.id, t1.info
Buffers: shared hit=1102
-> Sort (cost=79.83..82.33 rows=1000 width=4) (actual time=0.154..0.180 rows=200 loops=1)
Output: ((trunc((random() * '100000'::double precision)))::integer)
Sort Key: ((trunc((random() * '100000'::double precision)))::integer)
Sort Method: quicksort Memory: 34kB
-> Function Scan on pg_catalog.generate_series t (cost=0.00..20.00 rows=1000 width=0) (actual time=0.036..0.099 rows=200 loops=1)
Output: (trunc((random() * '100000'::double precision)))::integer
Function Call: generate_series(1, 200)
CONTEXT: SQL statement "SELECT * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id)"
PL/pgSQL function inline_code_block line 4 at PERFORM 优化选择建议如下，可以根据需求来写SQL，最不建议写的是OR：
index scan：
离散扫描，适合小的扫描集。
bitmap scan：
比index scan多了Recheck的开销，以及按照ctid排序的开销。
适合大的扫描集，排序的目的是减少离散扫描，还可以用到块设备的prefetch。
seq scan：
适合非常庞大的扫描集。
bitmapindex scan + bitmapor|And：
使用or的写法，只能选择bitmap index scan，所以不如使用IN的写法。
nestloop join：
适合小的扫描集，并且其中有一个表的JOIN列是带有主键或唯一约束的。
hash join：
适合大的扫描集，同时两个表的JOIN列上面都没有索引的情况。
merge join：
适合大的扫描集，并且两个表的JOIN列都有索引。
如果你发现写好的SQL没有选择最优的执行计划，可以通过设置优化器开关，或者使用hint plan这个插件来指定优化器使用对应的scan或join method.
http://pghintplan.osdn.jp/pg_hint_plan.html
git clone git://git.osdn.jp/gitroot/pghintplan/pg_hint_plan.git
mv pg_hint_plan postgresql-9.5.0/contrib/
cd postgresql-9.5.0/contrib/pg_hint_plan
export PATH=/home/digoal/pgsql9.5.0/bin:$PATH
make
make install
psql
postgres=# create extension pg_hint_plan;
CREATE EXTENSION
postgres=# LOAD 'pg_hint_plan';
postgres=# /*+ NestLoop(t1 t2) */ explain select * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id);
LOG: duration: 0.000 ms plan:
Query Text: /*+ NestLoop(t1 t2) */ explain select * from t_in_test t1 join (select trunc(random()*100000)::int as id from generate_series(1,200) t(id)) t2 on (t1.id=t2.id);
Nested Loop (cost=0.44..8404.50 rows=1000 width=41)
Output: t1.id, t1.info, ((trunc((random() * '100000'::double precision)))::integer)
-> Function Scan on pg_catalog.generate_series t (cost=0.00..20.00 rows=1000 width=0)
Output: (trunc((random() * '100000'::double precision)))::integer
Function Call: generate_series(1, 200)
-> Index Scan using t_in_test_pkey on public.t_in_test t1 (cost=0.43..8.36 rows=1 width=37)
Output: t1.id, t1.info
Index Cond: (t1.id = ((trunc((random() * '100000'::double precision)))::integer))
QUERY PLAN
------------------------------------------------------------------------------------------
Nested Loop (cost=0.44..8404.50 rows=1000 width=41)
-> Function Scan on generate_series t (cost=0.00..20.00 rows=1000 width=0)
-> Index Scan using t_in_test_pkey on t_in_test t1 (cost=0.43..8.36 rows=1 width=37)
Index Cond: (id = ((trunc((random() * '100000'::double precision)))::integer))
(4 rows) 
