数仓中典型的几种不下推语句整改案例

本文分享自华为云社区《GaussDB(DWS)性能调优：典型不下推语句整改案例》，作者：譡里个档。

场景1：With-Recursive仅包含值rte不可发货

根源： SQL语句的某个分支中没有FROM字句（只有VALUES或者类似SELECT 1这样的语句）

案例1：下降驱动分支没有FROM字句

原始语句

SELECT T.RPT_ITEM_ID, --报表项ID

T.RPT_ITEM_CODE,

T.USER_GROUP_CODE AS USER_GROUP_CODE --用户组

FROM BIF.BIF_RPT_ITEM_DEF_T T,

(WITH recursive cte AS (

SELECT DISTINCT TRIM(SUBSTR('' :: text, INSTR('', ',', 1, 1) + 1, INSTR('', ',', 1, 2) - INSTR('', ',', 1, 1) - 1)) AS cte_RPT_ITEM_CODE,

1 AS level

FROM (SELECT '') AS tb0



UNION ALL



SELECT DISTINCT TRIM(SUBSTR('' :: text, INSTR('', ',', 1, cte.level + 1) + 1, INSTR('', ',', 1, cte.level + 2) - INSTR('', ',', 1, cte.level + 1) - 1)),

cte.level + 1

FROM (SELECT '') AS tb0, cte

WHERE cte.level + 1 <= LENGTH('') - LENGTH(REPLACE('', ',', '')) - 1

)

SELECT

DISTINCT cte_RPT_ITEM_CODE AS RPT_ITEM_CODE

FROM cte

) T5

WHERE NVL(INSTR(T.RPT_ITEM_FREQUENCE, 'M'), 0) > 0

AND T.RPT_ITEM_CODE = NVL(T5.RPT_ITEM_CODE, T.RPT_ITEM_CODE)

AND T.RPT_ITEM_TYPE = 1 --是否是叶子报表项，1=是，0=否，基本报表项

AND T.ENABLE_FLAG = 1

AND T.VERSION = '202308' --使用快照,增加条件限制

ORDER BY T.RPT_ITEM_ID

改写语句

SELECT T.RPT_ITEM_ID, --报表项ID

T.RPT_ITEM_CODE,

T.USER_GROUP_CODE AS USER_GROUP_CODE --用户组

FROM BIF.BIF_RPT_ITEM_DEF_T T,

(WITH recursive cte AS (

SELECT DISTINCT TRIM(SUBSTR('' :: text, INSTR('', ',', 1, 1) + 1, INSTR('', ',', 1, 2) - INSTR('', ',', 1, 1) - 1)) AS cte_RPT_ITEM_CODE,

1 AS level

FROM generate_series(1, 1) AS tb0



UNION ALL



SELECT DISTINCT TRIM(SUBSTR('' :: text, INSTR('', ',', 1, cte.level + 1) + 1, INSTR('', ',', 1, cte.level + 2) - INSTR('', ',', 1, cte.level + 1) - 1)),

cte.level + 1

FROM (SELECT '') AS tb0, cte

WHERE cte.level + 1 <= LENGTH('') - LENGTH(REPLACE('', ',', '')) - 1

)

SELECT

DISTINCT cte_RPT_ITEM_CODE AS RPT_ITEM_CODE

FROM cte

) T5

WHERE NVL(INSTR(T.RPT_ITEM_FREQUENCE, 'M'), 0) > 0

AND T.RPT_ITEM_CODE = NVL(T5.RPT_ITEM_CODE, T.RPT_ITEM_CODE)

AND T.RPT_ITEM_TYPE = 1 --是否是叶子报表项，1=是，0=否，基本报表项

AND T.ENABLE_FLAG = 1

AND T.VERSION = '202308' --使用快照,增加条件限制

ORDER BY T.RPT_ITEM_ID

修改点比对

案例2：下降驱动分支没有FROM字句

原始语句

SELECT A.DYNM_COMP_ID,

DECODE(B.LINE_NO, 1, '202308', A.VERSION)

FROM BIF.BIF_DYNM_COMP_SOU_TBL_V A,

(WITH recursive cte AS (

SELECT 1 AS level

UNION ALL

SELECT cte.level + 1

FROM cte

WHERE cte.level + 1 < 3

)

SELECT

level as LINE_NO

FROM cte

) B

WHERE EXISTS (SELECT 1

FROM BIF.BIF_RPT_ITEM_DEF_MV RPT,

BIF.BIF_PUB_SUBJECT_AREA_T SBJ,

BIF.BIF_SNAPSHORT_SUBJECT_V TYP

WHERE A.DYNM_COMP_ID = RPT.DYNM_COMP_ID

AND RPT.VERSION = 'current'

AND RPT.SUBJECT_AREA_ID = SBJ.SUBJECT_AREA_ID

AND SBJ.SUBJECT_AREA_CODE =TYP.SUBJECT_CODE

AND TYP.SUBJECT_TYPE ='TAX')

AND A.VERSION = 'current'

改写语句

SELECT A.DYNM_COMP_ID,

DECODE(B.LINE_NO, 1, '202308', A.VERSION)

FROM BIF.BIF_DYNM_COMP_SOU_TBL_V A,

(SELECT

*

FROM generate_series(1, 2) AS cte(LINE_NO)

) B

WHERE EXISTS (SELECT 1

FROM BIF.BIF_RPT_ITEM_DEF_MV RPT,

BIF.BIF_PUB_SUBJECT_AREA_T SBJ,

BIF.BIF_SNAPSHORT_SUBJECT_V TYP

WHERE A.DYNM_COMP_ID = RPT.DYNM_COMP_ID

AND RPT.VERSION = 'current'

AND RPT.SUBJECT_AREA_ID = SBJ.SUBJECT_AREA_ID

AND SBJ.SUBJECT_AREA_CODE =TYP.SUBJECT_CODE

AND TYP.SUBJECT_TYPE ='TAX')

AND A.VERSION = 'current'

修改点比对

案例3：下降驱动分支是VALUES字句

原始语句

WITH RECURSIVE t(n) AS (

VALUES (1)

UNION ALL

SELECT

n+1

FROM t

WHERE n < (SELECT MAX(LENGTH(COMP_CODE)-LENGTH(REPLACE(COMP_CODE,',','')))+1 MAX_TOKENS

FROM (SELECT DEPT_CODE,

to_char(APPLICABLE_GEO_PC_CODE) COMP_CODE

FROM SDIHR.MDM_CDM_DEPT_ACT_INFO_T_3600)

)

)

SELECT n AS LVL FROM t

改写语句

WITH RECURSIVE t(n) AS (

SELECT * FROM generate_series(1, 1)

UNION ALL

SELECT

n+1

FROM t

WHERE n < (SELECT MAX(LENGTH(COMP_CODE)-LENGTH(REPLACE(COMP_CODE,',','')))+1 MAX_TOKENS

FROM (SELECT DEPT_CODE,

to_char(APPLICABLE_GEO_PC_CODE) COMP_CODE

FROM SDIHR.MDM_CDM_DEPT_ACT_INFO_T_3600)

)

)

SELECT n AS LVL FROM t

修改点比对

案例4：下降驱动分支是VALUES字句

原始语句

WITH RECURSIVE t(n) AS (

VALUES (1)

UNION ALL

SELECT

n+1

FROM t

WHERE n < (SELECT MAX(LENGTH(COMP_CODE)-LENGTH(REPLACE(COMP_CODE,',','')))+1 MAX_TOKENS

FROM (SELECT DEPT_CODE,

to_char(APPLICABLE_GEO_PC_CODE) COMP_CODE

FROM SDIHR.MDM_CDM_DEPT_ACT_INFO_T_3600))

)

SELECT n AS LVL FROM t

改写语句

SELECT

*

FROM generate_series(1, (SELECT MAX(LENGTH(COMP_CODE)-LENGTH(REPLACE(COMP_CODE,',','')))+1 MAX_TOKENS

FROM (SELECT DEPT_CODE,

to_char(APPLICABLE_GEO_PC_CODE) COMP_CODE

FROM SDIHR.MDM_CDM_DEPT_ACT_INFO_T_3600))

) AS t(lvl)

修改点比对

场景2：With-Recursive包含系统表不可发货

根因：连续语句的某个分支中没有FROM字句中只用系统表或者系统视图（DUAL也被视为系统视图）

案例1：下降驱动分支是FROM DUAL查询

原始语句

WITH recursive cte AS (

SELECT

TO_DATE(201701, 'YYYYMM') as level ,TO_DATE(20170131, 'YYYYMMDD') LASTDAY

FROM dual

UNION ALL

SELECT

ADD_MONTHS(cte.LEVEL, 1) AS PERIOD,

LAST_DAY(ADD_MONTHS(cte.LEVEL, 1)) AS LASTDAY

FROM cte WHERE cte.LEVEL <=SYSDATE

)

SELECT

TO_CHAR(cte.level,'YYYYMMDD') AS PERIOD , cte.LASTDAY

FROM cte

WHERE TO_CHAR(cte.level,'YYYYMMDD')<= TO_CHAR(SYSDATE,'YYYYMMDD')

改写语句

WITH recursive cte AS (

SELECT

TO_DATE(201701, 'YYYYMM') as level ,TO_DATE(20170131, 'YYYYMMDD') LASTDAY

FROM generate_series(1, 1)

UNION ALL

SELECT

ADD_MONTHS(cte.LEVEL, 1) AS PERIOD,

LAST_DAY(ADD_MONTHS(cte.LEVEL, 1)) AS LASTDAY

FROM cte WHERE cte.LEVEL <=SYSDATE

)

SELECT

TO_CHAR(cte.level,'YYYYMMDD') AS PERIOD , cte.LASTDAY

FROM cte

WHERE TO_CHAR(cte.level,'YYYYMMDD')<= TO_CHAR(SYSDATE,'YYYYMMDD')

修改点对比

场景3：CN上的SubPlan exec无法发货

根本原因：某个子查询语句只能在CN上执行，通常是这个子查询有不下推关系，比如有系统表、系统视图调用，或者存在不下推函数等

案例1：子查询中系统表/系统视图查询

原始语句

WITH error_log AS NOT MATERIALIZED (

SELECT

upper(log_column_name) AS log_column_name,

log_error_code,

s.char_length AS data_length,

s.data_type,s.nullable

FROM (SELECT

distinct

unnest(string_to_array(bad_log_column_name,',')) AS log_column_name,

unnest(string_to_array(bad_log_error_code,',')) AS log_error_code

FROM stgltc.BAD_cfs_inv_invoice_ad_2500

) T,

(SELECT * FROM user_tab_columns WHERE table_name=lower('dlt_cfs_inv_invoice_ad_2500')) S

WHERE upper(T.log_column_name)=upper(S.column_name)

)

SELECT

CASE WHEN upper('ACTIVITY_NAME') IN (SELECT log_column_name FROM error_log WHERE data_type IN ('varchar','char','character','nchar','character varying','varchar2','nvarchar2','clob','text') AND log_error_code='22001'/*字符超长*/)

THEN SUBSTRB(ACTIVITY_NAME,0,(SELECT distinct DATA_LENGTH FROM error_log WHERE upper(log_column_name)=upper('ACTIVITY_NAME')))

ELSE ACTIVITY_NAME

END AS ACTIVITY_NAME,

CASE WHEN upper('ADJUSTMENT_ID') IN (SELECT log_column_name FROM error_log WHERE data_type IN ('varchar','char','character','nchar','character varying','varchar2','nvarchar2','clob','text') AND log_error_code='22001'/*字符超长*/)

THEN SUBSTRB(ADJUSTMENT_ID,0,(SELECT distinct DATA_LENGTH FROM error_log WHERE upper(log_column_name)=upper('ADJUSTMENT_ID')))

ELSE ADJUSTMENT_ID

END AS ADJUSTMENT_ID

FROM stgltc.BAD_cfs_inv_invoice_ad_2500

改写语句

-- 识别不下推的子查询为WITH error_log字句中的

-- SELECT * FROM user_tab_columns WHERE table_name=lower('dlt_cfs_inv_invoice_ad_2500')

--

-- 因为这部分为系统表查询，无论如何都不能下推，所以此处把这部分结果转储到一个中间表中

-- 中间表创建成行存表

CREATE TEMP TABLE s WITH(orientation=row) DISTRIBUTE BY ROUNDROBIN AS

SELECT * FROM user_tab_columns WHERE table_name=lower('dlt_cfs_inv_invoice_ad_2500')

-- 因为整个查询涉及到的表都是列存表，之后前面创建的临时表s为行存表

-- 所以此处加一个强制走向量化的hint

WITH error_log AS NOT MATERIALIZED (

SELECT

upper(log_column_name) AS log_column_name,

log_error_code,

s.char_length AS data_length,

s.data_type,s.nullable

FROM (SELECT

distinct

unnest(string_to_array(bad_log_column_name,',')) AS log_column_name,

unnest(string_to_array(bad_log_error_code,',')) AS log_error_code

FROM stgltc.bad_cfs_inv_invoice_ad_2500

) T,

pg_temp.S

WHERE upper(T.log_column_name)=upper(S.column_name)

)

SELECT /*+ set global(enable_force_vector_engine on)*/

CASE WHEN upper('ACTIVITY_NAME') IN (SELECT log_column_name FROM error_log WHERE data_type IN ('varchar','char','character','nchar','character varying','varchar2','nvarchar2','clob','text') AND log_error_code='22001'/*字符超长*/)

THEN SUBSTRB(ACTIVITY_NAME,0,(SELECT distinct DATA_LENGTH FROM error_log WHERE upper(log_column_name)=upper('ACTIVITY_NAME')))

ELSE ACTIVITY_NAME

END AS ACTIVITY_NAME,

CASE WHEN upper('ADJUSTMENT_ID') IN (SELECT log_column_name FROM error_log WHERE data_type IN ('varchar','char','character','nchar','character varying','varchar2','nvarchar2','clob','text') AND log_error_code='22001'/*字符超长*/)

THEN SUBSTRB(ADJUSTMENT_ID,0,(SELECT distinct DATA_LENGTH FROM error_log WHERE upper(log_column_name)=upper('ADJUSTMENT_ID')))

ELSE ADJUSTMENT_ID

END AS ADJUSTMENT_ID

FROM stgltc.bad_cfs_inv_invoice_ad_2500

修改点对比

场景4：TargetList中的记录类型无法发货

根本原因：输出列中存在记录类型，这种类型的不下推一般是不会体现在最外层的输出列上，一般此类报错有两个场景

1.SQL书写逻辑错误，导致输出列上出现了（...）形式的输出列

2.SQL业务逻辑正确，这个场景需要了解业务意义，把记录字段强转为文本类型，然后再使用记录字段的位置做特殊装备

案例1：输出列书写错误，出现（...）形式的输出列

原始语句

SELECT

d.id,

coalesce(d.period, 'snull') AS period,

(d.plan_unit_code, 'snull') AS plan_unit_code,

coalesce(d.product_type_model, 'snull') AS product_type_model,

coalesce(d.revision, 'snull') AS revision,

d.start_date

FROM (SELECT *

FROM cdcscm.cdc_mp_d_forecast_t_6120 t

WHERE t.cdc_timestamp > to_date('2023-07-06 00:00:00', 'yyyy-mm-dd hh24:mi:ss') - 7

AND t.cdc_timestamp < to_date('2023-08-08 00:00:00', 'yyyy-mm-dd hh24:mi:ss')

) t1,

sdiscm.mp_d_forecast_t_6120 d

WHERE (t1.audit_op_type = 'delete' AND t1.audit_op_option = 'before')

AND d.id = t1.id

改写语句

SELECT

d.id,

coalesce(d.period, 'snull') AS period,

coalesce(d.plan_unit_code, 'snull') AS plan_unit_code,

coalesce(d.product_type_model, 'snull') AS product_type_model,

coalesce(d.revision, 'snull') AS revision,

d.start_date

FROM (SELECT *

FROM cdcscm.cdc_mp_d_forecast_t_6120 t

WHERE t.cdc_timestamp > to_date('2023-07-06 00:00:00', 'yyyy-mm-dd hh24:mi:ss') - 7

AND t.cdc_timestamp < to_date('2023-08-08 00:00:00', 'yyyy-mm-dd hh24:mi:ss')

) t1,

sdiscm.mp_d_forecast_t_6120 d

WHERE (t1.audit_op_type = 'delete' AND t1.audit_op_option = 'before')

AND d.id = t1.id

修改点对比

备注：修改前后写语句不等价，不等价的原因是因为原始SQL写入有问题，正确的写法是就是coalesce(d.plan_unit_code, 'snull') AS plan_unit_code。

点击关注，第一时间了解华为云新鲜技术~