PostgreSQL 10 自定义并行计算聚合函数的原理与实践

背景

PostgreSQL 9.6开始就支持并行计算了，意味着聚合、扫描、排序、JOIN等都开始支持并行计算。对于聚合操作来说，并行计算与非并行计算是有差异的。

例如avg聚合，对一张表进行计算时，一个任务中操作和多个并行任务操作，算法是不一样的。

PostgreSQL提供了一套标准的接口，可以支持聚合函数的并行操作。

自定义并行聚合的原理和例子

创建聚合函数的语法如下：

CREATE AGGREGATE name ( [ argmode ] [ argname ] arg_data_type [ , ... ] ) (  SFUNC = sfunc,  STYPE = state_data_type  [ , SSPACE = state_data_size ]  [ , FINALFUNC = ffunc ]  [ , FINALFUNC_EXTRA ]  [ , COMBINEFUNC = combinefunc ]  [ , SERIALFUNC = serialfunc ]  [ , DESERIALFUNC = deserialfunc ]  [ , INITCOND = initial_condition ]  [ , MSFUNC = msfunc ]  [ , MINVFUNC = minvfunc ]  [ , MSTYPE = mstate_data_type ]  [ , MSSPACE = mstate_data_size ]  [ , MFINALFUNC = mffunc ]  [ , MFINALFUNC_EXTRA ]  [ , MINITCOND = minitial_condition ]  [ , SORTOP = sort_operator ]  [ , PARALLEL = { SAFE | RESTRICTED | UNSAFE } ]
)

相比非并行，多了一个过程，那就是combinefunc的过程(也叫partial agg)。

非并行模式的聚合流程大致如下：

循环
sfunc( internal-state, next-data-values ) ---> next-internal-state  最后调用一次(可选)
ffunc( internal-state ) ---> aggregate-value

并行模式的聚合流程大致如下，如果没有写combinefunc，那么实际上聚合过程并没有实现并行而只是扫描并行：

下面这个例子，我们可以观察到一个COUNT操作的并行聚合。

postgres=# set max_parallel_workers=4;
SET
postgres=# set max_parallel_workers_per_gather =4;
SET
postgres=# set parallel_setup_cost =0;
SET
postgres=# set parallel_tuple_cost =0;
SET
postgres=# alter table test set (parallel_workers =4);
ALTER TABLE
postgres=# explain (analyze,verbose,timing,costs,buffers) select count(*) from test;  QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------  -- final并行，可有可无，看具体的聚合算法  Finalize Aggregate  (cost=15837.02..15837.03 rows=1 width=8) (actual time=57.296..57.296 rows=1 loops=1)  Output: count(*)  Buffers: shared hit=3060  ->  Gather  (cost=15837.00..15837.01 rows=4 width=8) (actual time=57.287..57.292 rows=5 loops=1)  Output: (PARTIAL count(*))  Workers Planned: 4  Workers Launched: 4  Buffers: shared hit=3060  -- 一下就是combinefunc完成的聚合并行(显示为PARTIAL agg)  ->  Partial Aggregate  (cost=15837.00..15837.01 rows=1 width=8) (actual time=52.333..52.333 rows=1 loops=5)  Output: PARTIAL count(*)  Buffers: shared hit=12712  Worker 0: actual time=50.917..50.918 rows=1 loops=1  Buffers: shared hit=2397  Worker 1: actual time=51.293..51.294 rows=1 loops=1  Buffers: shared hit=2423  Worker 2: actual time=51.062..51.063 rows=1 loops=1  Buffers: shared hit=2400  Worker 3: actual time=51.436..51.436 rows=1 loops=1  Buffers: shared hit=2432  ->  Parallel Seq Scan on public.test  (cost=0.00..15212.00 rows=250000 width=0) (actual time=0.010..30.499 rows=200000 loops=5)  Buffers: shared hit=12712  Worker 0: actual time=0.013..30.343 rows=190269 loops=1  Buffers: shared hit=2397  Worker 1: actual time=0.010..30.401 rows=192268 loops=1  Buffers: shared hit=2423  Worker 2: actual time=0.013..30.467 rows=190350 loops=1  Buffers: shared hit=2400  Worker 3: actual time=0.009..30.221 rows=192861 loops=1  Buffers: shared hit=2432  Planning time: 0.074 ms  Execution time: 60.169 ms
(31 rows)

了解了并行聚合的原理后，我们就可以写自定义聚合函数的并行计算了。

例子

例如我们要支持一个数组的聚合，并且在聚合过程中我们要实现对元素去重。

1、创建测试表

create table test(id int, col int[]);

2、生成测试数据

CREATE OR REPLACE FUNCTION public.gen_arr(integer, integer)  RETURNS integer[]  LANGUAGE sql  STRICT
AS $function$  select array(select ($1*random())::int from generate_series(1,$2));
$function$;  insert into test select random()*1000, gen_arr(500,10) from generate_series(1,10000);

3、创建聚合函数

例子1，没有combinefunc，只支持扫描并行。

数组去重函数

postgres=# create or replace function uniq(int[]) returns int[] as $$  select array( select unnest($1) group by 1);
$$ language sql strict parallel safe;
CREATE FUNCTION

数组合并与去重函数

postgres=# create or replace function array_uniq_cat(anyarray,anyarray) returns anyarray as $$  select uniq(array_cat($1,$2));
$$ language sql strict parallel safe;
CREATE FUNCTION

聚合函数(不带COMBINEFUNC)

create aggregate arragg (anyarray) (sfunc = array_uniq_cat, stype=anyarray, PARALLEL=safe);

并行查询例子：

postgres=# set max_parallel_workers=4;
SET
postgres=# set max_parallel_workers_per_gather =4;
SET
postgres=# set parallel_setup_cost =0;
SET
postgres=# set parallel_tuple_cost =0;
SET
postgres=# alter table test set (parallel_workers =4);
ALTER TABLE
postgres=# explain (analyze,verbose,timing,costs,buffers) select id, arragg(col) from test group by id ;

很明显没有设置COMBINEFUNC时，未使用并行聚合。

postgres=# explain (analyze,verbose,timing,costs,buffers) select id, arragg(col) from test group by id ;  QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------  HashAggregate  (cost=4139.74..4141.74 rows=200 width=36) (actual time=602.957..603.195 rows=1001 loops=1)  Output: id, arragg(col)  Group Key: test.id  Buffers: shared hit=6  ->  Gather  (cost=0.00..163.37 rows=15748 width=36) (actual time=0.328..43.734 rows=10000 loops=1)  Output: id, col  Workers Planned: 4  Workers Launched: 4  Buffers: shared hit=6  -- 只有并行扫描，没有并行聚合。  ->  Parallel Seq Scan on public.test  (cost=0.00..163.37 rows=3937 width=36) (actual time=0.017..0.891 rows=2000 loops=5)  Output: id, col  Buffers: shared hit=124  Worker 0: actual time=0.019..0.177 rows=648 loops=1  Buffers: shared hit=8  Worker 1: actual time=0.022..0.180 rows=648 loops=1  Buffers: shared hit=8  Worker 2: actual time=0.017..3.772 rows=7570 loops=1  Buffers: shared hit=94  Worker 3: actual time=0.015..0.189 rows=648 loops=1  Buffers: shared hit=8  Planning time: 0.084 ms  Execution time: 603.450 ms
(22 rows)

例子2，有combinefunc，支持并行聚合。

drop aggregate arragg(anyarray);  create aggregate arragg (anyarray) (sfunc = array_uniq_cat, stype=anyarray, COMBINEFUNC = array_uniq_cat, PARALLEL=safe);

使用了并行聚合。

postgres=# explain (analyze,verbose,timing,costs,buffers) select id, arragg(col) from test group by id ;  QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------  Finalize HashAggregate  (cost=1361.46..1363.46 rows=200 width=36) (actual time=285.489..285.732 rows=1001 loops=1)  Output: id, arragg(col)  Group Key: test.id  Buffers: shared hit=36  ->  Gather  (cost=1157.46..1159.46 rows=800 width=36) (actual time=63.654..74.163 rows=4297 loops=1)  Output: id, (PARTIAL arragg(col))  Workers Planned: 4  Workers Launched: 4  Buffers: shared hit=36  -- 并行聚合  ->  Partial HashAggregate  (cost=1157.46..1159.46 rows=200 width=36) (actual time=57.367..57.727 rows=859 loops=5)  Output: id, PARTIAL arragg(col)  Group Key: test.id  Buffers: shared hit=886  Worker 0: actual time=54.788..54.997 rows=857 loops=1  Buffers: shared hit=213  Worker 1: actual time=56.881..57.255 rows=861 loops=1  Buffers: shared hit=213  Worker 2: actual time=55.415..55.813 rows=856 loops=1  Buffers: shared hit=212  Worker 3: actual time=56.453..56.854 rows=838 loops=1  Buffers: shared hit=212  ->  Parallel Seq Scan on public.test  (cost=0.00..163.37 rows=3937 width=36) (actual time=0.011..0.736 rows=2000 loops=5)  Output: id, col  Buffers: shared hit=124  Worker 0: actual time=0.009..0.730 rows=1981 loops=1  Buffers: shared hit=25  Worker 1: actual time=0.012..0.773 rows=2025 loops=1  Buffers: shared hit=25  Worker 2: actual time=0.015..0.741 rows=1944 loops=1  Buffers: shared hit=24  Worker 3: actual time=0.012..0.751 rows=1944 loops=1  Buffers: shared hit=24  Planning time: 0.073 ms  Execution time: 285.949 ms
(34 rows)

实际上并行聚合与分布式数据库聚合阶段原理是一样的，分布式数据库自定义聚合可以参考末尾的文章。

参考

https://www.postgresql.org/docs/10/static/sql-createaggregate.html

https://www.postgresql.org/docs/10/static/xaggr.html#XAGGR-PARTIAL-AGGREGATES

《PostgreSQL aggregate function customize》

《Greenplum 最佳实践 - 估值插件hll的使用(以及hll分式聚合函数优化)》

《Postgres-XC customized aggregate introduction》