有时候写的 SQL 有性能问题时往往束手无策,而求助于 DBA。今天,我们从使用者、DBA、内核开发三个不同的角度来分析一个有趣的 SQL 性能问题的案例, 从浅入深了解 postgreSQL 的优化器。
问题描述
同事 A 来问我这个假 DBA 一条 SQL 的性能问题:
两条 SQL 生产环境执行情况
limit 10
select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 10;
复制代码
Execution Time: 1.307 ms
limit 1
select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 1;
复制代码
Execution Time: 144.098 ms
分析
执行计划
既然不是缓存问题,那我们先看看执行计划有什么不一样的
limit 1
# explain analyze verbose select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.43..416.25 rows=1 width=73) (actual time=135.213..135.214 rows=1 loops=1)
Output: xxx
-> Index Scan Backward using user_gift_pkey on yay.user_gift (cost=0.43..368000.44 rows=885 width=73) (actual time=135.212..135.212 rows=1 loops=1)
Output: xxx
Filter: ((user_gift.user_id = 11695667) AND (user_gift.user_type = 'default'::user_type))
Rows Removed by Filter: 330192
Planning Time: 0.102 ms
Execution Time: 135.235 ms
(8 rows)
Time: 135.691 ms
复制代码
limit 10
# explain analyze verbose select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 10;
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=868.20..868.22 rows=10 width=73) (actual time=1.543..1.545 rows=10 loops=1)
Output: xxx
-> Sort (cost=868.20..870.41 rows=885 width=73) (actual time=1.543..1.543 rows=10 loops=1)
Output: xxx
Sort Key: user_gift.id DESC
Sort Method: top-N heapsort Memory: 27kB
-> Index Scan using idx_user_type on yay.user_gift (cost=0.56..849.07 rows=885 width=73) (actual time=0.020..1.366 rows=775 loops=1)
Output: xxx
Index Cond: (user_gift.user_id = 11695667)
Filter: (user_gift.user_type = 'default'::user_type)
Planning Time: 0.079 ms
Execution Time: 1.564 ms
(12 rows)
Time: 1.871 ms
复制代码
可以看到,两个 SQL 执行计划不一样:
limit 1
语句 :使用主键进行倒序扫描, Index Scan Backward using user_gift_pkey on yay.user_gift
limit 10
语句 :使用(user_id, user_type)复合索引直接查找用户数据,Index Scan using idx_user_type on yay.user_gift
为什么执行计划不一样?
total cost
其实 postgreSQL 的执行计划并没有“问题”,因为limit 1
的 total costLimit (cost=0.43..416.25 rows=1 width=73)
是 416,run cost 是 416-0.43=415.57。而limit 10
的 total costLimit (cost=868.20..868.22 rows=10 width=73)
是 868.22。
如果使用Index Scan Backward using user_gift_pkey
的方式估算,那么limit 1
成本是 415, limit 2
是 415*2=830, limit 3
是 1245,大于 868,所以当limit 3
的时候会使用Index Scan using idx_user_type
扫索引的计划。
验证
# explain select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 2;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.43..831.95 rows=2 width=73)
-> Index Scan Backward using user_gift_pkey on user_gift (cost=0.43..367528.67 rows=884 width=73)
Filter: ((user_id = 11695667) AND (user_type = 'default'::user_type))
(3 rows)
Time: 0.341 ms
# explain select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 3;
QUERY PLAN
----------------------------------------------------------------------------------------------------------
Limit (cost=866.19..866.20 rows=3 width=73)
-> Sort (cost=866.19..868.40 rows=884 width=73)
Sort Key: id DESC
-> Index Scan using idx_user_type on user_gift (cost=0.56..854.76 rows=884 width=73)
Index Cond: (user_id = 11695667)
Filter: (user_type = 'default'::user_type)
(6 rows)
Time: 0.352 ms
复制代码
结果显示:
实际执行时间
limit 1
时成本估算的是 416.25,比limit 10
的868.22
还是要快的。
但是实际limit 1
执行 cost 是 135.691 ms,而limit 10
执行 cost 是 1.871 ms,比limit 10
慢了 70 倍!!!
我们重新执行下 explain,加上 buffers 选项
# explain (analyze, buffers, verbose) select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.43..416.29 rows=1 width=73) (actual time=451.542..451.544 rows=1 loops=1)
Output: xxx
Buffers: shared hit=214402 read=5280 dirtied=2302
I/O Timings: read=205.027
-> Index Scan Backward using user_gift_pkey on yay.user_gift (cost=0.43..368032.94 rows=885 width=73) (actual time=451.540..451.540 rows=1 loops=1)
Output: xxx
Filter: ((user_gift.user_id = 11695667) AND (user_gift.user_type = 'default'::user_type))
Rows Removed by Filter: 333462
Buffers: shared hit=214402 read=5280 dirtied=2302
I/O Timings: read=205.027
Planning Time: 1.106 ms
Execution Time: 451.594 ms
(12 rows)
复制代码
# explain (analyze, buffers, verbose) select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id desc limit 3;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=860.51..860.52 rows=3 width=73) (actual time=14.633..14.634 rows=3 loops=1)
Output: xxx
Buffers: shared hit=467 read=321
I/O Timings: read=10.112
-> Sort (cost=860.51..862.72 rows=885 width=73) (actual time=14.632..14.632 rows=3 loops=1)
Output: xxx
Sort Key: user_gift.id DESC
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=467 read=321
I/O Timings: read=10.112
-> Index Scan using idx_user_type on yay.user_gift (cost=0.56..849.07 rows=885 width=73) (actual time=0.192..14.424 rows=775 loops=1)
Output: xxx
Index Cond: (user_gift.user_id = 11695667)
Filter: (user_gift.user_type = 'default'::user_type)
Buffers: shared hit=464 read=321
I/O Timings: read=10.112
Planning Time: 0.111 ms
Execution Time: 14.658 ms
(18 rows)
复制代码
可以看出:
从上面输出Buffers: shared hit=214402 read=5280 dirtied=2302
可以看出limit 1
的计划从磁盘读取了 5280 个 blocks(pages)才找到符合 where 条件的记录。
为什么要读取这么多数据呢?我们来看看统计信息:
schemaname | yay
tablename | user_gift
attname | id
inherited | f
null_frac | 0
avg_width | 8
n_distinct | -1
most_common_vals |
most_common_freqs |
histogram_bounds | {93,9817,19893,28177,.......}
correlation | 0.788011
most_common_elems |
most_common_elem_freqs |
elem_count_histogram |
schemaname | yay
tablename | user_gift
attname | user_id
inherited | f
null_frac | 0
avg_width | 4
n_distinct | -0.175761
most_common_vals | {11576819,10299480,14020501,.......,11695667,......}
most_common_freqs | {0.000353333,0.000326667,0.000246667,......,9.33333e-05,......}
histogram_bounds | {3,10002181,10005599,10009672,......,11693300,11698290,......}
correlation | 0.53375
most_common_elems |
most_common_elem_freqs |
elem_count_histogram |
schemaname | yay
tablename | user_gift
attname | user_type
inherited | f
null_frac | 0
avg_width | 4
n_distinct | 3
most_common_vals | {default, invalid, deleted}
most_common_freqs | {0.997923,0.00194,0.000136667}
histogram_bounds |
correlation | 0.99763
most_common_elems |
most_common_elem_freqs |
elem_count_histogram |
复制代码
从统计信息里可以看出:
user_id
字段的most_common_vals
中有 11695667(user_id)的值,则可以直接通过其对应的most_common_freqs
来得到其 selectivity 是 9.33333e-05;
user_type
字段为default
对应的 selectivity 是 0.997923。
所以where user_id=11695667 and user_type='default'
的 selectivity 是 0.0000933333*0.997923 = 0.0000931394467359。
那么可以估算出满足 where 条件的用户数是 0.0000931394467359 * 9499740(总用户数) = 884.8,和执行计划(cost=0.43..367528.67 rows=884 width=73)
的 884 行一样。
而优化器的估算是基于数据分布均匀这个假设的:
那么数据分布真的均匀吗?继续查看数据的实际分布:
# select max(ctid) from user_gift;
max
-------------
(128709,29)
(1 row)
复制代码
# select max(ctid), min(ctid) from user_gift where user_id=11695667;
max | min
-------------+-----------
(124329,22) | (3951,64)
(1 row)
复制代码
# SELECT relpages, reltuples FROM pg_class WHERE relname = 'user_gift';
relpages | reltuples
----------+-------------
128875 | 9.49974e+06
(1 row)
复制代码
每个 page 存储的记录数:9.49974e+06 tuples / 128875 pages = 73.713 tuples/page。
计算:表(main table)的 B+tree 的最大 page 是 128709,而实际用户 11695667 的最大 page 是 124329,128709 - 124329 = 4380,需要扫描 4380 个 page 才能找到符合 where 条件的记录所在的 page,所以过滤的 rows 是 4380 pages * 73.713 tuples/page ≈ 322862。
实际limit 1
时扫描了 5280 个 pages(包含了主键索引的 pages),过滤了 333462 万行记录,和估算的基本一样:
Rows Removed by Filter: 333462
Buffers: shared hit=214402 read=5280 dirtied=2302
I/O Timings: read=205.027
复制代码
所以,此用户数据分布倾斜了:
那么扫描 5280 个 pages 要多久?
需要读取的数据量:5280pages * 8KB/page = 41.2MB 的数据。
[root]$ fio -name iops -rw=randread -bs=8k -runtime=10 -iodepth=1 -filename /dev/sdb -ioengine mmap -buffered=1
...
Run status group 0 (all jobs):
READ: bw=965KiB/s (988kB/s), 965KiB/s-965KiB/s (988kB/s-988kB/s), io=9656KiB (9888kB), run=10005-10005msec
[root]$ fio -name iops -rw=read -bs=8k -runtime=10 -iodepth=1 -filename /dev/sdb -ioengine mmap -direct=1
...
Run status group 0 (all jobs):
READ: bw=513MiB/s (538MB/s), 513MiB/s-513MiB/s (538MB/s-538MB/s), io=5132MiB (5381MB), run=10001-10001msec
复制代码
从fio
结果可以看出,此数据库机器磁盘的顺序读取速度约为 500MB/s,如果数据都是顺序的,那么扫描 40MB 数据需要约 80ms,
如果数据都是随机的,那么需要 40 秒。不是所有的数据都是顺序访问的,而且测试的是非线上机器,没有其他 IO 进程在运行。
到这里问题基本定位清楚了:
postgreSQL 的优化器认为数据分布是均匀的,只需要倒序扫描很快就找到符合条件的记录,而实际上此用户的数据分布在表的前端,就导致了实际执行 start-up time 如此慢了。
从内核视角来分析
我们从 postgreSQL 内核的角度来继续分析几个问题:
优化器如何估算 cost
优化器如何统计 actual time
表的信息
# \d user_gift;
Table "yay.user_gift"
Column | Type | Collation | Nullable | Default
--------------+--------------------------+-----------+----------+------------------------------------------------
id | bigint | | not null | nextval('user_gift_id_seq'::regclass)
user_id | integer | | not null |
ug_name | character varying(100) | | not null |
expired_time | timestamp with time zone | | | now()
created_time | timestamp with time zone | | not null | now()
updated_time | timestamp with time zone | | not null | now()
user_type | user_type | | not null | 'default'::user_type
Indexes:
"user_gift_pkey" PRIMARY KEY, btree (id)
"idx_user_type" btree (user_id, ug_name)
"user_gift_ug_name_idx" btree (ug_name)
复制代码
# SELECT relpages, reltuples FROM pg_class WHERE relname = 'user_gift_pkey';
relpages | reltuples
----------+-------------
40035 | 9.49974e+06
(1 row)
复制代码
# SELECT relpages, reltuples FROM pg_class WHERE relname = 'idx_user_type';
relpages | reltuples
----------+-------------
113572 | 9.49974e+06
(1 row)
复制代码
# SELECT relpages, reltuples FROM pg_class WHERE relname = 'user_gift';
relpages | reltuples
----------+-------------
128875 | 9.49974e+06
(1 row)
复制代码
=# select count(1) from user_gift where user_id=11695667;
count
-------
775
(1 row)
=# select count(1) from user_gift where user_id=11695667 and user_type = 'default' ;
count
-------
775
(1 row)
复制代码
# 主键高度
# select * from bt_metap('user_gift_pkey');
magic | version | root | level | fastroot | fastlevel | oldest_xact | last_cleanup_num_tuples
--------+---------+------+-------+----------+-----------+-------------+-------------------------
340322 | 3 | 412 | 2 | 412 | 2 | 0 | 9.31928e+06
(1 row)
// idx_user_type 高度
# select * from bt_metap('idx_user_type');
magic | version | root | level | fastroot | fastlevel | oldest_xact | last_cleanup_num_tuples
--------+---------+-------+-------+----------+-----------+-------------+-------------------------
340322 | 3 | 15094 | 3 | 15094 | 3 | 0 | 9.49974e+06
(1 row)
复制代码
估算 cost
start-up cost
postgreSQL 对于每种索引的成本估算是不一样的,我们看看 B+tree 的 start-up 成本是如何估算的:
// selfuncs.c
void
btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
Cost *indexStartupCost, Cost *indexTotalCost,
Selectivity *indexSelectivity, double *indexCorrelation,
double *indexPages)
{
......
descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
costs.indexStartupCost += descentCost;
......
// This cost is somewhat arbitrarily set at 50x cpu_operator_cost per page touched
descentCost = (index->tree_height + 1) * 50.0 * cpu_operator_cost;
costs.indexStartupCost += descentCost;
......
}
复制代码
其实 start-up cost 估算很简单,只考虑从 B+tree 的 root page 遍历到 leaf page,且将这个 page 读入第一个 tuple(记录)的 cost。
start-up 估算公式如下:
\left { ceil({\log_2 (N_{index,tuple})}) + (Height_{index} + 1) \times 50 \right }\ \times cpu_operator_cost
N(index,tuple) :索引 tuples(记录)数量
Height(index) : 索引 B+tree 的高度
cpu_operator_cost : 默认值 0.0025
使用 user_gift_pkey 计划的 start-up cost
从上面表信息中可以看出:
所以
\left { ceil({\log_2 (9499740)}) + (2 + 1) \times 50 \right }\ \times cpu_operator_cost = 173 \times 0.0025 = 0.435
和 postgreSQL 估算的 start-up cost=0.43 一样。
使用 idx_user_type 计划的 start-up cost
N(index,tuple) :9.49974e+06,
Height(index) : 3$\left { ceil({\log_2 (9499740)}) + (3 + 1) \times 50 \right }\ \times cpu_operator_cost = 223 \times 0.0025 = 0.5575$和 postgreSQL 估算的 start-up cost=0.56 一样。
run cost
run cost 的估算是比较复杂的,判断的条件非常多,无法用一个固定的公式计算出来,所以这里只是简单描述下,有兴趣的可以看 postgreSQL 源码src/backend/optimizer/path/costsize.c
的cost_index
函数,针对这个案例,一般情况下可以根据此链接的脚本进行来模拟计算 cost。
run cost$runcost=索引成本+主表成本$
索引成本 $索引成本=随机读取索引相关pages的成本+操作相关tuples的成本$
主表成本 $主表成本=maxiocost+indexcorrelation2×(miniocost−maxiocost)$
max io cost(最坏情况下 IO 成本)
所有 pages 都是随机读取
maxiocost=pagesfetched×randompagecost
min_io_cost(最优情况下 IO 成本)
第一个 page 是随机读取,后面 pages 都是顺序读取
miniocost=1×randompagecost+(pagesfetched−1)×seqpagecost
actual start-up time vs estimated start-up cost
刚刚的分析中有一个疑问被忽略了:estimated start-up cost 是开始执行计划到从表中读到的第一个 tuple 的 cost(cost is an arbitrary unit);而 actual start-up time 则是开始执行计划到从表中读取到第一个符合 where 条件的 tuple 的时间。这是为什么呢?
SQL 处理流程:postgreSQL 将 SQL 转化成 AST,然后进行优化,再将 AST 转成执行器(executor)来实现具体的操作。不进行优化的执行器是这样的:
┌──────────────┐
│ projection │
└──────┬───────┘
│
│
┌──────▼──────┐
│ limit │
└──────┬──────┘
│
│
┌──────▼──────┐
│ selection │
└──────┬──────┘
│
│
┌──────▼──────┐
│ index scan │
└─────────────┘
复制代码
简化的执行流程如下:
index scan executor:扫描到一个 tuple,就返回给 selection executor
selection executor:对 tuple 进行过滤,如果符合条件则返回给 limit executor,如果不符合则继续调用 index scan executor
limit executor:当达到 limit 限制则将数据返回给 projection executor
projection executor:过滤掉非select
列的数据
那么如果进行优化,一般会将selection executor
和projection executor
合并到index scan executor
中执行,以减少数据在 executor 之间的传递。
┌─────────────┐
│ limit │
└──────┬──────┘
│
│
┌──────▼──────┐
│ index scan │
│ │
│ + selection │
│ + projection│
└─────────────┘
复制代码
优化后的执行流程:
而通过下面代码可以看出,postgreSQL 对于执行时间的统计是基于 executor 的,
// src/backend/executor/execProcnode.c
static TupleTableSlot *
ExecProcNodeInstr(PlanState *node)
{
TupleTableSlot *result;
InstrStartNode(node->instrument);
result = node->ExecProcNodeReal(node);
// 统计执行指标
InstrStopNode(node->instrument, TupIsNull(result) ? 0.0 : 1.0);
return result;
}
复制代码
所以 actual time 的 start-up 是从启动 executor 直到扫描到符合where
语句的第一条结果为止。
再看看实际的函数调用栈,user_id=xxx
的过滤已经下沉到index scan executor
里面了。
---> int4eq(FunctionCallInfo fcinfo) (/home/ken/cpp/postgres/src/backend/utils/adt/int.c:379)
ExecInterpExpr(ExprState * state, ExprContext * econtext, _Bool * isnull) (/home/ken/cpp/postgres/src/backend/executor/execExprInterp.c:704)
ExecInterpExprStillValid(ExprState * state, ExprContext * econtext, _Bool * isNull) (/home/ken/cpp/postgres/src/backend/executor/execExprInterp.c:1807)
ExecEvalExprSwitchContext(ExprState * state, ExprContext * econtext, _Bool * isNull) (/home/ken/cpp/postgres/src/include/executor/executor.h:322)
---> ExecQual(ExprState * state, ExprContext * econtext) (/home/ken/cpp/postgres/src/include/executor/executor.h:391)
ExecScan(ScanState * node, ExecScanAccessMtd accessMtd, ExecScanRecheckMtd recheckMtd) (/home/ken/cpp/postgres/src/backend/executor/execScan.c:227)
---> ExecIndexScan(PlanState * pstate) (/home/ken/cpp/postgres/src/backend/executor/nodeIndexscan.c:537)
ExecProcNodeInstr(PlanState * node) (/home/ken/cpp/postgres/src/backend/executor/execProcnode.c:466)
ExecProcNodeFirst(PlanState * node) (/home/ken/cpp/postgres/src/backend/executor/execProcnode.c:450)
ExecProcNode(PlanState * node) (/home/ken/cpp/postgres/src/include/executor/executor.h:248)
---> ExecLimit(PlanState * pstate) (/home/ken/cpp/postgres/src/backend/executor/nodeLimit.c:96)
ExecProcNodeInstr(PlanState * node) (/home/ken/cpp/postgres/src/backend/executor/execProcnode.c:466)
ExecProcNodeFirst(PlanState * node) (/home/ken/cpp/postgres/src/backend/executor/execProcnode.c:450)
ExecProcNode(PlanState * node) (/home/ken/cpp/postgres/src/include/executor/executor.h:248)
ExecutePlan(EState * estate, PlanState * planstate, _Bool use_parallel_mode, CmdType operation, _Bool sendTuples, uint64 numberTuples, ScanDirection direction, DestReceiver * dest, _Bool execute_once) (/home/ken/cpp/postgres/src/backend/executor/execMain.c:1632)
standard_ExecutorRun(QueryDesc * queryDesc, ScanDirection direction, uint64 count, _Bool execute_once) (/home/ken/cpp/postgres/src/backend/executor/execMain.c:350)
ExecutorRun(QueryDesc * queryDesc, ScanDirection direction, uint64 count, _Bool execute_once) (/home/ken/cpp/postgres/src/backend/executor/execMain.c:294)
ExplainOnePlan(PlannedStmt * plannedstmt, IntoClause * into, ExplainState * es, const char * queryString, ParamListInfo params, QueryEnvironment * queryEnv, const instr_time * planduration, const BufferUsage * bufusage) (/home/ken/cpp/postgres/src/backend/commands/explain.c:571)
ExplainOneQuery(Query * query, int cursorOptions, IntoClause * into, ExplainState * es, const char * queryString, ParamListInfo params, QueryEnvironment * queryEnv) (/home/ken/cpp/postgres/src/backend/commands/explain.c:404)
ExplainQuery(ParseState * pstate, ExplainStmt * stmt, ParamListInfo params, DestReceiver * dest) (/home/ken/cpp/postgres/src/backend/commands/explain.c:275)
复制代码
下面代码是 scan 的实现,其中的ExecQual(qual, econtext)
是对 tuple 进行过滤,因为 selection 已经合并到 scan 中了。
TupleTableSlot *
ExecScan(ScanState *node, ExecScanAccessMtd accessMtd, ExecScanRecheckMtd recheckMtd)
{
......
for (;;)
{
TupleTableSlot *slot;
slot = ExecScanFetch(node, accessMtd, recheckMtd);
......
econtext->ecxt_scantuple = slot;
// Note : selection判断
if (qual == NULL || ExecQual(qual, econtext))
{
if (projInfo)
{
return ExecProject(projInfo);
}
else
{
return slot;
}
}
else
InstrCountFiltered1(node, 1);
}
}
复制代码
解决方案
禁用走主键扫描
既然计划走的是 user_gift_pkey 倒序扫描,那么我们可以手动避免优化器使用这个索引。
# explain analyze verbose select xxx from user_gift where user_id=11695667 and user_type = 'default' order by id+0 desc limit 1;
复制代码
将order by id
改成order by id+0
,由于id+0
是个表达式所以优化器就就不会使用 user_gift_pkey 这个索引了。
这个方案不适合所有场景,如果数据分布均匀的话则某些情况下使用 user_gift_pkey 扫描更加合理。
增加(user_id, id)索引
create index idx_user_id on user_gift(user_id, id);
复制代码
通过增加 where 条件列和排序键的复合索引,来避免走主键扫描。
写在最后
从排除缓存因素,分析查询计划,定位数据分布倾斜,到调试内核源码来进一步确定原因,最终成功解决性能问题。通过这个有趣的 SQL 优化经历,相信能给大家带来收获。
评论