GaussDB(DWS)性能调优，解决DM区大内存占用问题

gaussdb,dws,dm · 浏览次数 : 45

小编点评

本文主要讨论了在华为云数据库（GaussDB(DWS)）性能调优过程中遇到的一个具体问题及其解决方案。问题出现在维度表与主表关联时，使用会计期作为关联条件，导致大内存占用和数据倾斜。通过优化SQL语句，避免了数据倾斜问题，并提高了SQL执行性能。 **问题描述** 在场景一中，主表和维度表关联过程中将会计期作为关联条件，导致维度表未进行分区剪枝，可能产生大内存占用的情况。 **原始SQL** ```sql SELECT ... FROM DMACC.dm_adp_ar_trx_dtl_tmp F INNER JOIN DMDIM.DM_DIM_REGION_RC_D REG ON F.COA_GEO_PC_KEY = REG.GEO_PC_KEY INNER JOIN DMDIM.DM_DIM_PRODUCT_T_D T9 ON F.PROD_KEY = T9.PROD_KEY AND T9.PROD_POV_ID = 1 INNER JOIN DMDIM.DM_DIM_PROJECT_D J ON F.PROJ_KEY = J.PROJ_KEY INNER JOIN DMDIM.DM_DIM_CONTRACT_D HT ON HT.CONTRACT_KEY = F.CONTRACT_KEY LEFT JOIN DMCOMMON.DWR_CONFIG_DOMESTIC_FINANCE_V FIN ON F.COA_COMPANY_KEY = FIN.COMPANY_KEY AND F.COA_GEO_PC_KEY = FIN.GEO_PC_KEY LEFT JOIN DMAR.DWB_FMD_DIM_INVOICE_PAY_PLAN_D PP ON F.AR_INVOICE_PAY_PLAN_ID = PP.AR_INVOICE_PAY_PLAN_ID AND F.PERIOD_ID = PP.PERIOD_ID LEFT JOIN DMARDI.DWR_DIM_AR_INVOICE_V INV ON F.AR_INVOICE_ID = INV.AR_INVOICE_ID INNER JOIN DMARDI.DWR_DIM_AR_APPLICATION_V APP ON F.AR_APPLICATION_RECORD_ID = APP.AR_APPLICATION_RECORD_ID INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_V RCP ON F.AR_RECEIPT_RECORD_ID = RCP.AR_RECEIPT_RECORD_ID INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_TYPE_V RT ON RCP.RECEIPT_RECORD_TYPE_ID = RT.AR_RECEIPT_TYPE_ID LEFT JOIN ... WHERE ... ``` **性能分析** 从执行计划中可以看出，由于使用会计期作为关联条件，导致维度表未进行分区剪枝，数据量大，产生了数据倾斜，同时还由于数据量大出现了关联下盘，大大降低了SQL执行性能。 **优化方案** 1. 删除left join中的维度表会计期条件，避免主表关联后的结果重分布时将period_id作为分布键之一。 2. 将f.period_id = 维度表.period_id这一关联条件删掉，对SQL进行改写。改写后的SQL语句如下： ```sql SELECT ... FROM DMACC.dm_adp_ar_trx_dtl_tmp F INNER JOIN DMDIM.DM_DIM_REGION_RC_D REG ON F.COA_GEO_PC_KEY = REG.GEO_PC_KEY INNER JOIN DMDIM.DM_DIM_PRODUCT_T_D T9 ON F.PROD_KEY = T9.PROD_KEY AND T9.PROD_POV_ID = 1 INNER JOIN DMDIM.DM_DIM_PROJECT_D J ON F.PROJ_KEY = J.PROJ_KEY INNER JOIN DMDIM.DM_DIM_CONTRACT_D HT ON HT.CONTRACT_KEY = F.CONTRACT_KEY LEFT JOIN DMCOMMON.DWR_CONFIG_DOMESTIC_FINANCE_V FIN ON F.COA_COMPANY_KEY = FIN.COMPANY_KEY AND F.COA_GEO_PC_KEY = FIN.GEO_PC_KEY LEFT JOIN DMAR.DWB_FMD_DIM_INVOICE_PAY_PLAN_D PP ON F.AR_INVOICE_PAY_PLAN_ID = PP.AR_INVOICE_PAY_PLAN_ID AND PP.PERIOD_ID = '202406' LEFT JOIN DMARDI.DWR_DIM_AR_INVOICE_V INV ON F.AR_INVOICE_ID = INV.AR_INVOICE_ID INNER JOIN DMARDI.DWR_DIM_AR_APPLICATION_V APP ON F.AR_APPLICATION_RECORD_ID = APP.AR_APPLICATION_RECORD_ID INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_V RCP ON F.AR_RECEIPT_RECORD_ID = RCP.AR_RECEIPT_RECORD_ID INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_TYPE_V RT ON RCP.RECEIPT_RECORD_TYPE_ID = RT.AR_RECEIPT_TYPE_ID LEFT JOIN ... WHERE ... ``` 经过优化后，执行计划中维度表进行了分区剪枝，数据量减少，缓解了数据倾斜，也避免了关联下盘的问题，提高了SQL执行性能。

正文

本文分享自华为云社区《GaussDB(DWS)性能调优：DM区优化案例——维度表关联条件存在会计期》，作者： O泡果奶~。

当前DM（P1、P3、CBGDM）存在维度表与主表关联时使用会计期作为关联条件，会导致出现大内存占用或未识别数据倾斜的问题

【场景一】f.period_id = 维度表.period_id

1.1、【问题描述】

主表和维度表关联过程中将会计期作为关联条件，导致维度表未进行分区剪枝，可能会产生大内存占用的情况

1.2、【原始SQL】

仅呈现SQL中的问题，详细SQL见附件

FROM
        DMACC.dm_adp_ar_trx_dtl_tmp F
        INNER JOIN DMDIM.DM_DIM_REGION_RC_D REG ON F.COA_GEO_PC_KEY = REG.GEO_PC_KEY
        INNER JOIN DMDIM.DM_DIM_PRODUCT_T_D T9 ON F.PROD_KEY = T9.PROD_KEY 
        AND T9.PROD_POV_ID = 1
        INNER JOIN DMDIM.DM_DIM_PROJECT_D J ON F.PROJ_KEY = J.PROJ_KEY
        INNER JOIN DMDIM.DM_DIM_CONTRACT_D HT ON HT.CONTRACT_KEY = F.CONTRACT_KEY
        LEFT JOIN DMCOMMON.DWR_CONFIG_DOMESTIC_FINANCE_V FIN ON F.COA_COMPANY_KEY = FIN.COMPANY_KEY 
        AND F.COA_GEO_PC_KEY = FIN.GEO_PC_KEY
        LEFT JOIN DMAR.DWB_FMD_DIM_INVOICE_PAY_PLAN_D PP ON F.AR_INVOICE_PAY_PLAN_ID = PP.AR_INVOICE_PAY_PLAN_ID 
        AND F.PERIOD_ID = PP.PERIOD_ID
        LEFT JOIN DMARDI.DWR_DIM_AR_INVOICE_V INV ON F.AR_INVOICE_ID = INV.AR_INVOICE_ID
        INNER JOIN DMARDI.DWR_DIM_AR_APPLICATION_V APP ON F.AR_APPLICATION_RECORD_ID = APP.AR_APPLICATION_RECORD_ID
        INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_V RCP ON F.AR_RECEIPT_RECORD_ID = RCP.AR_RECEIPT_RECORD_ID
        INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_TYPE_V RT ON RCP.RECEIPT_RECORD_TYPE_ID = RT.AR_RECEIPT_TYPE_ID
        LEFT JOIN (
        SELECT C
            .CONTRACT_KEY,
            D.COMPANY_KEY,
            R.FIRST_SHIP_DATE 
        FROM
            DMDIM.dm_dim_contract_d C,
            DMDIM.DM_DIM_COMPANY_D D,
            DMARDI.DWR_CTRCT_FIRST_SHIP_DATE_R R 
        WHERE
            C.CONTRACT_ID = R.CONTRACT_ID 
            AND D.COMPANY_ID = R.COMPANY_ID 
        ) FR ON F.CONTRACT_KEY = FR.CONTRACT_KEY 
        AND F.COA_COMPANY_KEY = FR.COMPANY_KEY
        INNER JOIN DMDIM.DM_DIM_SALES_MODE_D MO ON F.SALES_MODE_KEY = MO.SALES_MODE_KEY
        JOIN DMDIM.DM_DIM_JOURNAL_SOURCE_D T29 ON F.JE_SOURCE_ID = T29.JE_SOURCE_ID
        JOIN DMDIM.DM_DIM_JOURNAL_CATEGORY_D T30 ON F.JE_CATEGORY_ID = T30.JE_CATEGORY_ID

1.3、【性能分析】

从上图的执行计划可以看出，由于用会计期作为关联条件，导致维度表未进行分区剪枝，数据量大，不但产生了数据倾斜，同时还由于数据量大出现了关联下盘，大大降低了sql执行性能。
主表只有一个会计期，可以识别出对应的会计期，然后对SQL进行如下改写：

FROM
        DMACC.dm_adp_ar_trx_dtl_tmp F
        INNER JOIN DMDIM.DM_DIM_REGION_RC_D REG ON F.COA_GEO_PC_KEY = REG.GEO_PC_KEY
        INNER JOIN DMDIM.DM_DIM_PRODUCT_T_D T9 ON F.PROD_KEY = T9.PROD_KEY 
        AND T9.PROD_POV_ID = 1
        INNER JOIN DMDIM.DM_DIM_PROJECT_D J ON F.PROJ_KEY = J.PROJ_KEY
        INNER JOIN DMDIM.DM_DIM_CONTRACT_D HT ON HT.CONTRACT_KEY = F.CONTRACT_KEY
        LEFT JOIN DMCOMMON.DWR_CONFIG_DOMESTIC_FINANCE_V FIN ON F.COA_COMPANY_KEY = FIN.COMPANY_KEY 
        AND F.COA_GEO_PC_KEY = FIN.GEO_PC_KEY
        LEFT JOIN DMAR.DWB_FMD_DIM_INVOICE_PAY_PLAN_D PP ON F.AR_INVOICE_PAY_PLAN_ID = PP.AR_INVOICE_PAY_PLAN_ID 
        AND PP.PERIOD_ID = '202406'
        LEFT JOIN DMARDI.DWR_DIM_AR_INVOICE_V INV ON F.AR_INVOICE_ID = INV.AR_INVOICE_ID
        INNER JOIN DMARDI.DWR_DIM_AR_APPLICATION_V APP ON F.AR_APPLICATION_RECORD_ID = APP.AR_APPLICATION_RECORD_ID
        INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_V RCP ON F.AR_RECEIPT_RECORD_ID = RCP.AR_RECEIPT_RECORD_ID
        INNER JOIN DMARDI.DWR_DIM_AR_RECEIPT_TYPE_V RT ON RCP.RECEIPT_RECORD_TYPE_ID = RT.AR_RECEIPT_TYPE_ID
        LEFT JOIN (
        SELECT C
            .CONTRACT_KEY,
            D.COMPANY_KEY,
            R.FIRST_SHIP_DATE 
        FROM
            DMDIM.dm_dim_contract_d C,
            DMDIM.DM_DIM_COMPANY_D D,
            DMARDI.DWR_CTRCT_FIRST_SHIP_DATE_R R 
        WHERE
            C.CONTRACT_ID = R.CONTRACT_ID 
            AND D.COMPANY_ID = R.COMPANY_ID 
        ) FR ON F.CONTRACT_KEY = FR.CONTRACT_KEY 
        AND F.COA_COMPANY_KEY = FR.COMPANY_KEY
        INNER JOIN DMDIM.DM_DIM_SALES_MODE_D MO ON F.SALES_MODE_KEY = MO.SALES_MODE_KEY
        JOIN DMDIM.DM_DIM_JOURNAL_SOURCE_D T29 ON F.JE_SOURCE_ID = T29.JE_SOURCE_ID
        JOIN DMDIM.DM_DIM_JOURNAL_CATEGORY_D T30 ON F.JE_CATEGORY_ID = T30.JE_CATEGORY_ID

经优化后，执行计划如下图所示，维度表进行了分区剪枝，数据量减少，缓解了数据倾斜，也避免了关联下盘的问题。

【场景二】f left join 维度表 on f.period_id = 维度表.period_id and 维度表.period_id = ‘会计期’

2.1、【问题描述】

主表和维度表关联过程中将会计期作为关联条件，同时还为维度表会计期进行赋值，可能会产生数据倾斜未识别的情况

2.2、【原始SQL】

FROM
        dmdp.dm_dpc_inv_m_dtl_f_TEM_A LT1
        LEFT JOIN dmcommon.dm_dim_prod_key_r LT2 ON LT1.prod_key = LT2.old_key 
        AND LT1.period_id = LT2.period_id 
        AND LT2.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_reg_key_r LT3 ON LT1.period_id = LT3.period_id 
        AND LT1.geo_pc_key = LT3.old_key 
        AND LT3.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT4 ON LT1.period_id = LT4.period_id 
        AND LT1.account_dept_cust_key = LT4.old_key 
        AND LT4.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_proj_key_r LT5 ON LT1.period_id = LT5.period_id 
        AND LT1.proj_key = LT5.old_key 
        AND LT5.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT6 ON LT1.period_id = LT6.period_id 
        AND LT1.enterprise_cust_key = LT6.old_key 
        AND LT6.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_rep_key_r LT7 ON LT1.period_id = LT7.period_id 
        AND LT1.report_item_id = LT7.old_key 
        AND LT7.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_supply_center_key_r LT8 ON LT1.period_id = LT8.period_id 
        AND LT1.supply_center_key = LT8.old_key 
        AND LT8.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_inv_key_r LT9 ON LT1.period_id = LT9.period_id 
        AND LT1.inventory_class_key = LT9.old_key 
        AND LT9.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_bus_key_r LT10 ON LT1.period_id = LT10.period_id 
        AND LT1.business_status_key = LT10.old_key 
        AND LT10.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_hisi_key_r LT11 ON LT1.period_id = LT11.period_id 
        AND LT1.hisi_prod_key = LT11.old_key 
        AND LT11.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_inv_org_key_r LT12 ON LT1.period_id = LT12.period_id 
        AND LT1.inventory_org_key = LT12.old_key 
        AND LT12.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT13 ON LT1.period_id = LT13.period_id 
        AND LT1.end_cust_key = LT13.old_key 
        AND LT13.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT14 ON LT1.period_id = LT14.period_id 
        AND LT1.sign_cust_key = LT14.old_key 
        AND LT14.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT15 ON LT1.period_id = LT15.period_id 
        AND LT1.agent_distribution_cust_key = LT15.old_key 
        AND LT15.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_com_key_r LT16 ON LT1.period_id = LT16.period_id 
        AND LT1.company_key = LT16.old_key 
        AND LT16.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_con_key_r LT17 ON LT1.period_id = LT17.period_id 
        AND LT1.contract_key = LT17.old_key 
        AND LT17.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_con_key_r LT18 ON LT1.period_id = LT18.period_id 
        AND LT1.loan_contract_key = LT18.old_key 
        AND LT18.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_supply_center_key_r LT19 ON LT1.period_id = LT19.period_id 
        AND LT1.target_supply_center_key = LT19.old_key 
        AND LT19.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_subinventory_key_r LT20 ON LT1.period_id = LT20.period_id 
        AND LT1.subinventory_key = LT20.old_key 
        AND LT20.PERIOD_ID = 202406 
    WHERE
        1 = 1 
    AND partition_value IN ( 0, 1 )

2.3、【性能分析】

上图的执行计划可以看出，在主表一开始关联过程中就存在数据倾斜，导致SQL执行性能差。

详细执行计划中，虽然维度表进行了分区剪枝，但由于使用了 left join，导致关联条件中维度表的常量period_id不能直接赋值给主表period_id，主表关联后的结果重分布时将period_id作为了分布键之一，这会影响优化器的倾斜优化。
可以将f.period_id = 维度表.period_id这一关联条件删掉，对sql进行如下改写

FROM
        dmdp.dm_dpc_inv_m_dtl_f_TEM_A LT1
        LEFT JOIN dmcommon.dm_dim_prod_key_r LT2 ON LT1.prod_key = LT2.old_key 
        AND LT2.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_reg_key_r LT3 ON LT1.geo_pc_key = LT3.old_key 
        AND LT3.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT4 ON LT1.account_dept_cust_key = LT4.old_key 
        AND LT4.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_proj_key_r LT5 ON LT1.proj_key = LT5.old_key 
        AND LT5.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT6 ON LT1.enterprise_cust_key = LT6.old_key 
        AND LT6.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_rep_key_r LT7 ON LT1.report_item_id = LT7.old_key 
        AND LT7.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_supply_center_key_r LT8 ON LT1.supply_center_key = LT8.old_key 
        AND LT8.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_inv_key_r LT9 ON LT1.inventory_class_key = LT9.old_key 
        AND LT9.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_bus_key_r LT10 ON LT1.business_status_key = LT10.old_key 
        AND LT10.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_hisi_key_r LT11 ON LT1.hisi_prod_key = LT11.old_key 
        AND LT11.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_inv_org_key_r LT12 ON LT1.inventory_org_key = LT12.old_key 
        AND LT12.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT13 ON LT1.end_cust_key = LT13.old_key 
        AND LT13.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT14 ON LT1.sign_cust_key = LT14.old_key 
        AND LT14.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_cus_key_r LT15 ON LT1.agent_distribution_cust_key = LT15.old_key 
        AND LT15.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_com_key_r LT16 ON LT1.company_key = LT16.old_key 
        AND LT16.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_con_key_r LT17 ON LT1.contract_key = LT17.old_key 
        AND LT17.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_con_key_r LT18 ON LT1.loan_contract_key = LT18.old_key 
        AND LT18.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_supply_center_key_r LT19 ON LT1.target_supply_center_key = LT19.old_key 
        AND LT19.PERIOD_ID = 202406
        LEFT JOIN dmcommon.dm_dim_subinventory_key_r LT20 ON LT1.subinventory_key = LT20.old_key 
        AND LT20.PERIOD_ID = 202406 
    WHERE
        1 = 1 
    AND partition_value IN ( 0, 1 )

改写后，执行计划如下所示

可以看出，执行计划不但进行了分区剪枝，同时优化器还进行了倾斜优化，提高了SQL执行性能

点击关注，第一时间了解华为云新鲜技术~