Author: 明虚
最近遇到几个客户在HybridDB上做性能测试时，都遇到Count Distinct的性能调优问题。这里我们总结一下HybridDB中，对Count Distinct的几种处理方式。
我们以一个客户的案例来做说明。客户的典型的业务场景是，在用户行为日志中统计对应类别的行为数，类别有几千个，独立的行为的总量很多，有几千万；为分析行为，要查询一段时间内的基于类别的独立行为数，查询如下（test的建表语句见附录）：
select category, count(distinct actionId) as ct from test_user_log
where receivetime between '2017-03-07 11:00:00' and '2017-03-07 12:00:00' group by category
order by ct desc limit 10;
下面我们针对这个查询，来看一下Count Distinct是怎么处理的。
利用explain analyze命令，看一下这个查询执行过程的信息：
test=# explain analyze select category, count(distinct actionId) as ct from test_user_log
where receivetime between '2017-03-07 11:00:00' and '2017-03-07 12:00:00' group by category
order by ct desc limit 10;
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-------------------
Limit (cost=0.00..431.00 rows=10 width=16)
Gather Motion 16:1 (slice2; segments: 16) (cost=5968.98..5968.99 rows=1 width=40)
Merge Key: ct
Rows out: 745 rows at destination with 2469 ms to end, start offset by 77 ms.
-> Sort (cost=5968.98..5968.99 rows=1 width=40)
Sort Key: ct
Rows out: Avg 46.6 rows x 16 workers. Max 55 rows (seg0) with 2461 ms to end, start offset by 85 ms.
Executor memory: 58K bytes avg, 58K bytes max (seg0).
Work_mem used: 58K bytes avg, 58K bytes max (seg0). Workfile: (0 spilling, 0 reused)
-> GroupAggregate (cost=5968.94..5968.97 rows=1 width=40)
Group By: public.test_user_log.category
Rows out: Avg 46.6 rows x 16 workers. Max 55 rows (seg0) with 2460 ms to end, start offset by 85 ms.
-> Sort (cost=5968.94..5968.94 rows=1 width=40)
Sort Key: public.test_user_log.category
Rows out: Avg 461.6 rows x 16 workers. Max 572 rows (seg4) with 2458 ms to end, start offset by 88 ms.
Executor memory: 85K bytes avg, 145K bytes max (seg4).
Work_mem used: 85K bytes avg, 145K bytes max (seg4). Workfile: (0 spilling, 0 reused)
-> Redistribute Motion 16:16 (slice1; segments: 16) (cost=5960.60..5968.93 rows=1 width=40)
Hash Key: public.test_user_log.category
Rows out: Avg 461.6 rows x 16 workers at destination. Max 572 rows (seg4) with 2316 ms to first row, 2458 ms to end, start offset by 88 ms.
-> GroupAggregate (cost=5960.60..5968.91 rows=1 width=40)
Group By: public.test_user_log.category
Rows out: Avg 461.6 rows x 16 workers. Max 472 rows (seg7) with 536 ms to first row, 2455 ms to end, start offset by 89 ms.
Executor memory: 318587K bytes avg, 330108K bytes max (seg7).
Work_mem used: 8544K bytes avg, 8544K bytes max (seg0).
Work_mem wanted: 8414K bytes avg, 8472K bytes max (seg14) to lessen workfile I/O affecting 16 workers.
-> Sort (cost=5960.60..5963.37 rows=70 width=64)
Sort Key: public.test_user_log.category
Rows out: Avg 367982.3 rows x 16 workers. Max 369230 rows (seg8) with 527 ms to first row, 625 ms to end, start offset by 90 ms.
Executor memory: 61433K bytes avg, 61433K bytes max (seg0).
Work_mem used: 61433K bytes avg, 61433K bytes max (seg0). Workfile: (0 spilling, 0 reused)
-> Append (cost=0.00..5904.72 rows=70 width=64)
Rows out: Avg 367982.3 rows x 16 workers. Max 369230 rows (seg8) with 2.710 ms to first row, 265 ms to end, start offset by 91 ms
.
-> Append-only Columnar Scan on test_user_log_1_prt_usual test_user_log (cost=0.00..0.00 rows=1 width=64)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00'::timestamp
without time zone
Rows out: 0 rows (seg0) with 0.542 ms to end, start offset by 93 ms.
-> Append-only Columnar Scan on test_user_log_1_prt_157 test_user_log (cost=0.00..3134.45 rows=37 width=64)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00'::timestamp
without time zone
Rows out: Avg 367882.1 rows x 16 workers. Max 369131 rows (seg8) with 2.178 ms to first row, 132 ms to end, start offset by
91 ms.
-> Append-only Columnar Scan on test_user_log_1_prt_158 test_user_log (cost=0.00..2770.27 rows=33 width=64)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00'::timestamp
without time zone
Rows out: Avg 100.2 rows x 16 workers. Max 124 rows (seg11) with 2.135 ms to first row, 73 ms to end, start offset by 394 m
s.
Slice statistics:
Settings: effective_cache_size=8GB; gp_statistics_use_fkeys=on; optimizer=off
Optimizer status: legacy query optimizer
Total runtime: 2546.862 ms
(51 rows)
可以发现，看似很简单的查询，处理流程却有点复杂。整个处理流程大致如下：
Scan （Columnar Scan + Append) -> Sort(category) -> Group by(category) -> Redistribute -> Sort(category) -> Group by(category) -> Sort -> Gather
从各个节点的实际执行时间的记录可以看出，主要的时间花在了前三步，因为这三步完成后，中间结果只有几百行了。我们重点关注这三个步骤。其实这几个步骤的逻辑比较好理解，扫描出来的数据，直接做排序，排序后，再把结果扫描一遍，同时进行聚合运算。
这里需要注意的一个细节是，查询的表test_user_log是按actionId做分布键的，相同的actionId都会分布在同一个节点上，所以每个节点本地按category做分组后，会在每个分组记录分组中出现的不同actionId值，最终的聚合的结果是category加上一个对应的actionId的计数。
这里有个疑问，其实category的唯一值很少（只有几百个），很适合利用Hash的方式做聚合呢，那么为什么没有选择Hash的方式而是采用了Sort的方式呢？
观察上述查询计划中test_user_log_1_prt_157这个表分区的中间结果估计（(cost=0.00..3134.45 rows=37 width=64)），可以发现预估的结果只有37行，而实际是Rows out: Avg 367882.1 rows，即36万多，这说明表的统计信息不准确。
执行一下Analyze来更新统计信息：
test=# analyze test_user_log_1_prt_157;
analyze test_user_log_1_prt_158;ANALYZE
test=# analyze test_user_log_1_prt_158;
ANALYZE
更新统计信息后，再执行一下查询，查询计划果然发生了变化：排序聚合变成了Hash方式聚合！执行时间也由2546.862ms缩短到2099.144ms。
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-------------------------
Limit (cost=0.00..431.00 rows=10 width=16)
Gather Motion 16:1 (slice2; segments: 16) (cost=320695.70..320695.92 rows=10 width=40)
Merge Key: ct
-> Limit (cost=320695.70..320695.72 rows=1 width=40)
-> Sort (cost=320695.70..320695.94 rows=7 width=40)
Sort Key (Limit): ct
-> HashAggregate (cost=320691.23..320692.46 rows=7 width=40)
Group By: partial_aggregation.category
-> HashAggregate (cost=303756.92..311454.34 rows=38488 width=64)
Group By: public.test_user_log.category, public.test_user_log.actionId
-> Redistribute Motion 16:16 (slice1; segments: 16) (cost=280664.69..292980.55 rows=38488 width=64)
Hash Key: public.test_user_log.category
-> HashAggregate (cost=280664.69..280664.69 rows=38488 width=64)
Group By: public.test_user_log.category, public.test_user_log.actionId
-> Append (cost=0.00..236527.23 rows=367813 width=43)
-> Append-only Columnar Scan on test_user_log_1_prt_usual test_user_log (cost=0.00..0.00 rows=1 width=64)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00'::tim
estamp without time zone
-> Append-only Columnar Scan on test_user_log_1_prt_157 test_user_log (cost=0.00..124267.69 rows=367813 width=42)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00'::tim
estamp without time zone
-> Append-only Columnar Scan on test_user_log_1_prt_158 test_user_log (cost=0.00..112259.54 rows=1 width=42)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00'::tim
estamp without time zone
上面的计划主要流程如下。两次Group by(Hash（category，actionId)）都是为了去除重复值，保证（category，actionId)组合字段值唯一。
Scan （Columnar Scan + Append) -> Group by(Hash(category，actionId)） -> Redistribute(category) -> Group by(Hash(category, acitonId)） -> Group by(Hash（category)） -> Sort -> Gather
那么还有其他可能的处理方式吗？我们知道，HybridDB支持新型的orca优化器，orca考虑更多的查询执行方式。我们下面试试使用orca来生成查询计划。
test=# set optimizer=on;
SET
test=# explain analyze select category, count(distinct actionId) as ct from test_user_log
where receivetime between '2017-03-07 11:00:00' and '2017-03-07 12:00:00' group by category
order by ct desc limit 10;
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-------------------------------
Limit (cost=0.00..431.00 rows=1 width=16)
Rows out: 10 rows with 2690 ms to end, start offset by 0.500 ms.
-> Gather Motion 16:1 (slice2; segments: 16) (cost=0.00..431.00 rows=1 width=16)
Merge Key: ct
Rows out: 10 rows at destination with 2690 ms to end, start offset by 0.501 ms.
-> Sort (cost=0.00..431.00 rows=1 width=16)
Sort Key: ct
Rows out: Avg 46.6 rows x 16 workers. Max 55 rows (seg0) with 2688 ms to end, start offset by 2.356 ms.
Executor memory: 33K bytes avg, 33K bytes max (seg0).
Work_mem used: 33K bytes avg, 33K bytes max (seg0). Workfile: (0 spilling, 0 reused)
-> GroupAggregate (cost=0.00..431.00 rows=1 width=16)
Group By: category
Rows out: Avg 46.6 rows x 16 workers. Max 55 rows (seg0) with 2687 ms to first row, 2688 ms to end, start offset by 2.372 ms.
-> Sort (cost=0.00..431.00 rows=1 width=16)
Sort Key: category
Rows out: Avg 461.6 rows x 16 workers. Max 572 rows (seg4) with 2687 ms to end, start offset by 3.104 ms.
Executor memory: 85K bytes avg, 145K bytes max (seg4).
Work_mem used: 85K bytes avg, 145K bytes max (seg4). Workfile: (0 spilling, 0 reused)
-> Redistribute Motion 16:16 (slice1; segments: 16) (cost=0.00..431.00 rows=1 width=16)
Hash Key: category
Rows out: Avg 461.6 rows x 16 workers at destination. Max 572 rows (seg4) with 2442 ms to first row, 2687 ms to end, start offset by 3.113 ms
.
-> Result (cost=0.00..431.00 rows=1 width=16)
Rows out: Avg 461.6 rows x 16 workers. Max 472 rows (seg7) with 1070 ms to first row, 2583 ms to end, start offset by 3.898 ms.
-> GroupAggregate (cost=0.00..431.00 rows=1 width=16)
Group By: category
Rows out: Avg 461.6 rows x 16 workers. Max 472 rows (seg7) with 1070 ms to first row, 2583 ms to end, start offset by 3.898 ms.
Executor memory: 316808K bytes avg, 328245K bytes max (seg7).
Work_mem used: 8184K bytes avg, 8184K bytes max (seg0).
Work_mem wanted: 8414K bytes avg, 8472K bytes max (seg14) to lessen workfile I/O affecting 16 workers.
-> Sort (cost=0.00..431.00 rows=1 width=16)
Sort Key: category, actionId
Rows out: Avg 367982.3 rows x 16 workers. Max 369230 rows (seg8) with 1064 ms to first row, 1143 ms to end, start offset by
3.812 ms.
Executor memory: 143353K bytes avg, 143353K bytes max (seg0).
Work_mem used: 143353K bytes avg, 143353K bytes max (seg0). Workfile: (0 spilling, 0 reused)
-> Sequence (cost=0.00..431.00 rows=1 width=24)
Rows out: Avg 367982.3 rows x 16 workers. Max 369230 rows (seg8) with 1.905 ms to first row, 241 ms to end, start off
set by 4.032 ms.
-> Partition Selector for test_user_log (dynamic scan id: 1) (cost=10.00..100.00 rows=7 width=4)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00
'::timestamp without time zone
Partitions selected: 3 (out of 745)
Rows out: 0 rows (seg0) with 0.013 ms to end, start offset by 4.033 ms.
-> Dynamic Table Scan on test_user_log (dynamic scan id: 1) (cost=0.00..431.00 rows=1 width=24)
Filter: receivetime >= '2017-03-07 11:00:00'::timestamp without time zone AND receivetime <= '2017-03-07 12:00:00
'::timestamp without time zone
Rows out: Avg 367982.3 rows x 16 workers. Max 369230 rows (seg8) with 1.891 ms to first row, 202 ms to end, sta
rt offset by 4.046 ms.
Partitions scanned: Avg 3.0 (out of 745) x 16 workers. Max 3 parts (seg0).
Slice statistics:
(slice0) Executor memory: 351K bytes.
(slice1) * Executor memory: 152057K bytes avg x 16 workers, 152057K bytes max (seg0). Work_mem: 143353K bytes max, 8472K bytes wanted.
(slice2) Executor memory: 363K bytes avg x 16 workers, 423K bytes max (seg4). Work_mem: 145K bytes max.
Statement statistics:
Memory used: 2047000K bytes
Memory wanted: 34684K bytes
Settings: effective_cache_size=8GB; gp_statistics_use_fkeys=on; optimizer=on
Optimizer status: PQO version 1.609
Total runtime: 2690.675 ms
(54 rows)
使用orca生成的查询计划，又回到了使用Sort+Groupby的方式来做聚合（这是因为，我们使用Analyze只更新了子分区表的统计信息，而orca只会考虑主表上的统计信息，要想是orca的计划转为使用Hash方式，需要在主表上使用Analyze，这里我们不继续讨论）。而上述使用orca生成的计划，与使用缺省优化器有很大不同。orca的查询计划采用了下面的流程：
Scan (Dynamic Scan) -> Sort (category, actionId) -> Group by (category) -> Redistribute -> Sort (category) -> Group by(category) -> Sort -> Gather
注意，第一次Sort用了(category, actionId）两个字段的组合，但后面的Group by时只适应了category一个字段！这是一种特殊的聚合方式。在做这种聚合时，对应一个不同的category，只需保留一个actionId的计数即可，而不是像在缺省优化器计划中那样，对每个不同的category，需要保留所有不同的actionId值，这样省去了建立类似Hash表的数据结构的时间。但由于Sort的时候用了两个字段，时间消耗比使用一个字段高，导致整个查询计划的性能不如缺省优化器产生的计划。
上面的讨论所举的例子中的表，正好是以Count Distinct的字段（即actionId）作为分布键的。如果以其他字段作为分布键，会产生不一样的查询计划，但基本原理都是类似的。
另外，我们没有涉及一个查询中涉及多个字段上有Count Distinct的情况，读者可以自行尝试。
create table test_user_log
(
actionId text,
code text,
receiveTime timestamp,
gmtCreate timestamp,
category text,
version text,
tag text,
siteId int4
)
with (APPENDONLY=true, ORIENTATION=column, BLOCKSIZE=524288)
distributed by (actionId)
PARTITION BY RANGE (receivetime)
(START ('2017-03-07') INCLUSIVE END ('2017-03-07') EXCLUSIVE EVERY (INTERVAL '1 hour' ), DEFAULT PARTITION usual);