EXPLAIN - 云原生大数据计算服务 MaxCompute

在实际开发过程中，通常需要分析查询语句或表结构来分析性能瓶颈，MaxCompute SQL为您提供explain语句实现此功能。本文为您介绍explain的功能、命令格式及使用示例。

功能介绍

EXPLAIN语句可以显示MaxCompute SQL对应的DML语句执行计划（执行SQL语义的程序）的结构，帮助您了解SQL语句的处理过程，为优化SQL语句提供帮助。一个查询语句作业会对应多个Job，一个Job对应多个Task。

说明

如果查询语句足够复杂，EXPLAIN的结果较多，超过4 MB则会触发API的限制，无法得到完整的EXPLAIN结果。此时您可以拆分查询语句，对各部分分别执行EXPLAIN语句，以了解Job的结构。

命令格式

EXPLAIN <dml query>;

dml query：必填。SELECT语句，更多信息请参见SELECT语法。

返回说明

EXPLAIN的执行结果包含如下信息：

Job间的依赖关系
例如job0 is root job。如果查询只需要一个Job（job0），只会显示一行信息。
Task间的依赖关系
```
In Job job0:
root Tasks: M1, M2
J3_1_2_Stg1 depends on: M1, M2
```
job0包含三个Task，M1、M2和J3_1_2_Stg1。系统会先执行M1和M2两个Task，执行完成后，再执行J3_1_2_Stg1。
Task的命名规则如下：
- 在MaxCompute中，共有四种Task类型：MapTask、ReduceTask、JoinTask和LocalWork。Task名称的第一个字母表示了当前Task的类型，例如M2Stg1就是一个MapTask。
- 紧跟着第一个字母后的数字，代表了当前Task的ID。这个ID在当前查询对应的所有Task中是唯一的。
- 用下划线（_）分隔的数字代表当前Task的直接依赖，例如J3_1_2_Stg1表示当前Task ID为3，依赖ID为1（M1）和ID为2（M2）的两个Task。

Task中所有Operator的依赖结构。

Operator串描述了一个Task的执行语义。结构示例如下：

In Task M2:
    Data source: mf_mc_bj.sale_detail_jt/sale_date=2013/region=china  # "Data source"描述了当前Task的输入内容。
    TS: mf_mc_bj.sale_detail_jt/sale_date=2013/region=china           # TableScanOperator
        FIL: ISNOTNULL(customer_id)                                   # FilterOperator
            RS: order: +                                              # ReduceSinkOperator
                nullDirection: *
                optimizeOrderBy: False
                valueDestLimit: 0
                dist: HASH
                keys:
                      customer_id
                values:
                      customer_id (string)
                      total_price (double)
                partitions:
                      customer_id


In Task J3_1_2:
    JOIN:                                                           # JoinOperator
         StreamLineRead1 INNERJOIN StreamLineRead2
         keys:
             0:customer_id
             1:customer_id

        AGGREGATE: group by:customer_id                            # GroupByOperator
         UDAF: SUM(total_price) (__agg_0_sum)[Complete],SUM(total_price) (__agg_1_sum)[Complete]
            RS: order: +
                nullDirection: *
                optimizeOrderBy: True
                valueDestLimit: 10
                dist: HASH
                keys:
                      customer_id
                values:
                      customer_id (string)
                      __agg_0 (double)
                      __agg_1 (double)
                partitions:


In Task R4_3:
    SEL: customer_id,__agg_0,__agg_1                               # SelectOperator
        LIM:limit 10                                               # LimitOperator
            FS: output: Screen                                     # FileSinkOperator
                schema:
                  customer_id (string) AS ashop
                  __agg_0 (double) AS ap
                  __agg_1 (double) AS bp

各Operator的含义如下：

TableScanOperator（TS）：描述查询语句中的FROM语句块的逻辑。EXPLAIN结果中会显示输入表的名称（Alias）。
SelectOperator（SEL）：描述查询语句中的SELECT语句块的逻辑。EXPLAIN结果中会显示向下一个Operator传递的列，多个列由逗号分隔。
- 如果是列的引用，则显示为<alias>.<column_name>。
- 如果是表达式的结果，则显示为函数形式，例如func1(arg1_1, arg1_2, func2(arg2_1, arg2_2))。
- 如果是常量，则直接显示常量值。
FilterOperator（FIL）：描述查询语句中的WHERE语句块的逻辑。EXPLAIN结果中会显示一个WHERE条件表达式，形式类似SelectOperator的显示规则。
JoinOperator（JOIN）：描述查询语句中的JOIN语句块的逻辑。EXPLAIN结果中会显示哪些表以哪种方式JOIN在一起。
GroupByOperator（例如AGGREGATE）：描述聚合操作的逻辑。如果查询中使用了聚合函数，就会出现该结构，EXPLAIN结果中会显示聚合函数的内容。
ReduceSinkOperator（RS）：描述Task间数据分发操作的逻辑。如果当前Task的结果会传递给另一个Task，则必然需要在当前Task的最后，使用ReduceSinkOperator执行数据分发操作。EXPLAIN的结果中会显示输出结果的排序方式、分发的Key、Value以及用来求Hash值的列。
FileSinkOperator（FS）：描述最终数据的存储操作。如果查询中有INSERT语句块，EXPLAIN结果中会显示目标表名称。
LimitOperator（LIM）：描述查询语句中的LIMIT语句块的逻辑。EXPLAIN结果中会显示LIMIT数。
MapjoinOperator（HASHJOIN）：类似JoinOperator，描述大表的JOIN操作。

示例数据

为便于理解，本文为您提供源数据，基于源数据提供相关示例。创建表sale_detail和sale_detail_jt，并添加数据，命令示例如下：

--创建分区表sale_detail和sale_detail_jt。
CREATE TABLE if NOT EXISTS sale_detail
(
shop_name     STRING,
customer_id   STRING,
total_price   DOUBLE
)
PARTITIONED BY (sale_date STRING, region STRING);

CREATE TABLE if NOT EXISTS sale_detail_jt
(
shop_name     STRING,
customer_id   STRING,
total_price   DOUBLE
)
PARTITIONED BY (sale_date STRING, region STRING);

--向源表增加分区。
ALTER TABLE sale_detail ADD PARTITION (sale_date='2013', region='china') PARTITION (sale_date='2014', region='shanghai');
ALTER TABLE sale_detail_jt ADD PARTITION (sale_date='2013', region='china');

--向源表追加数据。
INSERT INTO sale_detail PARTITION (sale_date='2013', region='china') VALUES ('s1','c1',100.1),('s2','c2',100.2),('s3','c3',100.3);
INSERT INTO sale_detail PARTITION (sale_date='2014', region='shanghai') VALUES ('null','c5',null),('s6','c6',100.4),('s7','c7',100.5);
INSERT INTO sale_detail_jt PARTITION (sale_date='2013', region='china') VALUES ('s1','c1',100.1),('s2','c2',100.2),('s5','c2',100.2);

--查询表sale_detail和sale_detail_jt中的数据，命令示例如下：
SET odps.sql.allow.fullscan=true;
SELECT * FROM sale_detail;
--返回结果
+------------+-------------+-------------+------------+------------+
| shop_name  | customer_id | total_price | sale_date  | region     |
+------------+-------------+-------------+------------+------------+
| s1         | c1          | 100.1       | 2013       | china      |
| s2         | c2          | 100.2       | 2013       | china      |
| s3         | c3          | 100.3       | 2013       | china      |
| null       | c5          | NULL        | 2014       | shanghai   |
| s6         | c6          | 100.4       | 2014       | shanghai   |
| s7         | c7          | 100.5       | 2014       | shanghai   |
+------------+-------------+-------------+------------+------------+

SET odps.sql.allow.fullscan=true;
SELECT * FROM sale_detail_jt;
-- 返回结果
+------------+-------------+-------------+------------+------------+
| shop_name  | customer_id | total_price | sale_date  | region     |
+------------+-------------+-------------+------------+------------+
| s1         | c1          | 100.1       | 2013       | china      |
| s2         | c2          | 100.2       | 2013       | china      |
| s5         | c2          | 100.2       | 2013       | china      |
+------------+-------------+-------------+------------+------------+

--创建做关联的表。
SET odps.sql.allow.fullscan=true;
CREATE TABLE shop AS SELECT shop_name, customer_id, total_price FROM sale_detail;

使用示例

下述示例均基于示例数据执行。

示例1

查询语句：

SELECT a.customer_id AS ashop, SUM(a.total_price) AS ap,COUNT(b.total_price) AS bp 
FROM (SELECT * FROM sale_detail_jt WHERE sale_date='2013' AND region='china') a 
INNER JOIN (SELECT * FROM sale_detail WHERE sale_date='2013' AND region='china') b 
ON a.customer_id=b.customer_id 
GROUP BY a.customer_id 
ORDER BY a.customer_id 
LIMIT 10;

获取查询语句语义，命令如下：

EXPLAIN 
SELECT a.customer_id AS ashop, SUM(a.total_price) AS ap,COUNT(b.total_price) AS bp 
FROM (SELECT * FROM sale_detail_jt WHERE sale_date='2013' AND region='china') a 
INNER JOIN (SELECT * FROM sale_detail WHERE sale_date='2013' AND region='china') b 
ON a.customer_id=b.customer_id 
GROUP BY a.customer_id 
ORDER BY a.customer_id 
LIMIT 10;

返回结果如下：

job0 is root job

In Job job0:
root Tasks: M1
M2_1 depends on: M1
R3_2 depends on: M2_1
R4_3 depends on: R3_2

In Task M1:
    Data source: doc_****.default.sale_detail/sale_date=2013/region=china
    TS: doc_****.default.sale_detail/sale_date=2013/region=china
        Statistics: Num rows: 3.0, Data size: 324.0
        FIL: ISNOTNULL(customer_id)
            Statistics: Num rows: 2.7, Data size: 291.6
            RS: valueDestLimit: 0
                dist: BROADCAST
                keys:
                values:
                      customer_id (string)
                      total_price (double)
                partitions:

                Statistics: Num rows: 2.7, Data size: 291.6

In Task M2_1:
    Data source: doc_****.default.sale_detail_jt/sale_date=2013/region=china
    TS: doc_****.default.sale_detail_jt/sale_date=2013/region=china
        Statistics: Num rows: 3.0, Data size: 324.0
        FIL: ISNOTNULL(customer_id)
            Statistics: Num rows: 2.7, Data size: 291.6
            HASHJOIN:
                     Filter1 INNERJOIN StreamLineRead1
                     keys:
                         0:customer_id
                         1:customer_id
                     non-equals:
                         0:
                         1:
                     bigTable: Filter1

                Statistics: Num rows: 3.6450000000000005, Data size: 787.32
                RS: order: +
                    nullDirection: *
                    optimizeOrderBy: False
                    valueDestLimit: 0
                    dist: HASH
                    keys:
                          customer_id
                    values:
                          customer_id (string)
                          total_price (double)
                          total_price (double)
                    partitions:
                          customer_id

                    Statistics: Num rows: 3.6450000000000005, Data size: 422.82000000000005

In Task R3_2:
    AGGREGATE: group by:customer_id
     UDAF: SUM(total_price) (__agg_0_sum)[Complete],COUNT(total_price) (__agg_1_count)[Complete]
        Statistics: Num rows: 1.0, Data size: 116.0
        RS: order: +
            nullDirection: *
            optimizeOrderBy: True
            valueDestLimit: 10
            dist: HASH
            keys:
                  customer_id
            values:
                  customer_id (string)
                  __agg_0 (double)
                  __agg_1 (bigint)
            partitions:

            Statistics: Num rows: 1.0, Data size: 116.0

In Task R4_3:
    SEL: customer_id,__agg_0,__agg_1
        Statistics: Num rows: 1.0, Data size: 116.0
        SEL: customer_id ashop, __agg_0 ap, __agg_1 bp, customer_id
            Statistics: Num rows: 1.0, Data size: 216.0
            FS: output: Screen
                schema:
                  ashop (string)
                  ap (double)
                  bp (bigint)

                Statistics: Num rows: 1.0, Data size: 116.0

OK

示例2

查询语句：

SELECT /*+ mapjoin(a) */
       a.customer_id AS ashop, SUM(a.total_price) AS ap,COUNT(b.total_price) AS bp 
FROM (SELECT * FROM sale_detail_jt 
WHERE sale_date='2013' AND region='china') a 
INNER JOIN (SELECT * FROM sale_detail WHERE sale_date='2013' AND region='china') b 
ON a.total_price<b.total_price 
GROUP BY a.customer_id 
ORDER BY a.customer_id 
LIMIT 10;

获取查询语句语义：

EXPLAIN 
SELECT /*+ mapjoin(a) */
       a.customer_id AS ashop, SUM(a.total_price) AS ap,COUNT(b.total_price) AS bp 
FROM (SELECT * FROM sale_detail_jt 
WHERE sale_date='2013' AND region='china') a 
INNER JOIN (SELECT * FROM sale_detail WHERE sale_date='2013' AND region='china') b 
ON a.total_price<b.total_price 
GROUP BY a.customer_id 
ORDER BY a.customer_id 
LIMIT 10;

返回结果如下：

job0 is root job

In Job job0:
root Tasks: M1
M2_1 depends on: M1
R3_2 depends on: M2_1
R4_3 depends on: R3_2

In Task M1:
    Data source: doc_****.sale_detail_jt/sale_date=2013/region=china
    TS: doc_****.sale_detail_jt/sale_date=2013/region=china
        Statistics: Num rows: 3.0, Data size: 324.0
        RS: valueDestLimit: 0
            dist: BROADCAST
            keys:
            values:
                  customer_id (string)
                  total_price (double)
            partitions:

            Statistics: Num rows: 3.0, Data size: 324.0

In Task M2_1:
    Data source: doc_****.sale_detail/sale_date=2013/region=china
    TS: doc_****.sale_detail/sale_date=2013/region=china
        Statistics: Num rows: 3.0, Data size: 24.0
        HASHJOIN:
                 StreamLineRead1 INNERJOIN TableScan2
                 keys:
                     0:
                     1:
                 non-equals:
                     0:
                     1:
                 bigTable: TableScan2

            Statistics: Num rows: 9.0, Data size: 1044.0
            FIL: LT(total_price,total_price)
                Statistics: Num rows: 6.75, Data size: 783.0
                AGGREGATE: group by:customer_id
                 UDAF: SUM(total_price) (__agg_0_sum)[Partial_1],COUNT(total_price) (__agg_1_count)[Partial_1]
                    Statistics: Num rows: 2.3116438356164384, Data size: 268.1506849315069
                    RS: order: +
                        nullDirection: *
                        optimizeOrderBy: False
                        valueDestLimit: 0
                        dist: HASH
                        keys:
                              customer_id
                        values:
                              customer_id (string)
                              __agg_0_sum (double)
                              __agg_1_count (bigint)
                        partitions:
                              customer_id

                        Statistics: Num rows: 2.3116438356164384, Data size: 268.1506849315069

In Task R3_2:
    AGGREGATE: group by:customer_id
     UDAF: SUM(__agg_0_sum)[Final] __agg_0,COUNT(__agg_1_count)[Final] __agg_1
        Statistics: Num rows: 1.6875, Data size: 195.75
        RS: order: +
            nullDirection: *
            optimizeOrderBy: True
            valueDestLimit: 10
            dist: HASH
            keys:
                  customer_id
            values:
                  customer_id (string)
                  __agg_0 (double)
                  __agg_1 (bigint)
            partitions:

            Statistics: Num rows: 1.6875, Data size: 195.75

In Task R4_3:
    SEL: customer_id,__agg_0,__agg_1
        Statistics: Num rows: 1.6875, Data size: 195.75
        SEL: customer_id ashop, __agg_0 ap, __agg_1 bp, customer_id
            Statistics: Num rows: 1.6875, Data size: 364.5
            FS: output: Screen
                schema:
                  ashop (string)
                  ap (double)
                  bp (bigint)

                Statistics: Num rows: 1.6875, Data size: 195.75

OK