一、背景

目前实时数仓提供的投放实时指标优先级别越来越重要，特别下游为规则引擎提供的数仓数据，直接对投放运营的广告投放产生直接影响，数据延迟或者异常均可能产生直接或者间接的资产损失；从投放管理平台的链路全景图投放全景图来看，目前投放是一个闭环的运行流程，实时数仓处于数据链路中的关键节点，实时数据直接支持规则引擎的自动化操作，以及投放管理平台的手动控盘；实时节点事故，将可能导致整个投放链路无法正常运行；为使投放链路达到99.9%的稳定性，需要对链路任务做相关的稳定性提升，优先级提升。

研发测试综合评估方案对投放实时链路增加一条备链路，投放需求迭代，通过备链路进行迭代修改，完成修改后进行主备链路Diff，确保Diff通过率99.9%，即可上线。

二、实现方案

• 数据准备：主备链路产出的数据分别实时写入到ODPS中。

• 数据采集：测试工具服务同时采集主备链路数据切片，保留2份同一个时间周期的数据。

• 数据降噪&Diff：工具采集数据后将进行第一步的降噪处理；主备数据开始对比&第二步降噪处理。

• 数据Diff结果：加工数据对比的结果，判断出每个字段的差异量，再最终判断出整体数据的差异量，给出结果。

三、搭建主备链路

实时链路解释：源头数据写入Kafka，Flink消费Kafka数据作为数据源（Source），结合属性字段做算子加工处理（Transformatin），处理结果写入Kafka（Sink），做下一步处理。经过一个个Flink任务节点加工分流到应用数据库中。

四、数据准备-数据切片

时间窗口切片

根据测试时间点，进行切片，取当天0点～执行时间段数据进行固定，确保数据不再更新。

业务场景切片

不同业务场景迭代进行切片，下发数据流提供多种下游场景数据，针对发生迭代的业务场景数据进行切片固定。如：fields_a='b'

五、主备链路数据Diff-去噪

数据漂移问题

问题现象：数据流在不断更新，同一条业务数据的数据流更新的最新的一条，主链路可能进入当天分区中，备链路可能进入到第二天分区中。

去噪方案：数据流取末尾1条数据。

数据更新频率问题

问题现象：同一条业务数据在更新过程中，主链路可能发生了10次更新，后面五次数据不发生改动，备链路只发生了5次更新。

去噪方案：同一个业务数据取数据流的N条数据。

数据更新时效问题

问题现象：同一条业务数据更新过程中，主链路更新三个数据为11.68、12.9、13.05；备链路更新三个数据为11.68、12.9、13.1；可以看出后面1次更新的数据并不一样。

去噪方案：同一个业务数据的数据流融合成一个list，主备相互判断末尾数据是否存在于对方截取的数据流list中。

属性字段值不统一问题

问题现象：存在空字符和null、0和0.0的情况，Diff结果为不通过，实际业务含义是OK的。

去噪方案：统一转换后进行Diff。

主备链路message字段解析属性字段不一致问题

问题现象：message字段存储数据JSON格式。同一条业务数据，主备链路解析的JSON对应的属性字段并不是完全一致的，两者之间存在差异。

去噪方案：通过代码解析出全量的属性字段，确保可以完全Diff。

message范本：

 {"fields_a":"20230628","fields_b":"2023-06-22 19:48:24","fields_c":"2","fields_d":"plan","fields_e":"3******","fields_f":"0.0","fields_g":"2","fields_h":"4*****","fields_i":"ext","fields_j":"binlog+odps","fields_k":"2","fields_l":"STATUS_*****","fields_m":"1********","fields_n":"孙**","fields_o":"2023-06-28T22:19:43.872"} 

转换JSON：

 { "fields_a": "20230717", "fields_d": "plan", "fields_e": "3******", "fields_aj": "33761.125", "fields_p": "37934.0", "fields_r": "1250.412", "fields_s": "1250.412", "fields_t": "33761.125", "fields_w": "33761.125", "fields_m": "1*********", "fields_v": "33761.125", "fields_y": "33761.125", "fields_n": "孙**", "fields_z": "1250.412", "fields_ai": "27", "fields_ak": "", "fields_aa": "33761.125", "fields_ab": "33761.125", "fields_ac": "33761.0", "fields_al": "0.1002", "fields_i": "***", "fields_j": "***", "fields_k": "2", "fields_ad": "1.0", "fields_ak": "37934.0", "fields_x": "1250.412", "fields_y": "0.0", "fields_ag": "27", "fields_af": "27", "fields_ah": "0.0", "fields_al": "0.0", "fields_am": "0.0", "fields_ao": "37934.0", "fields_ap": "37934.0", "fields_an": "33761.125", "fields_aq": "1*********", "fields_ae": "27", "fields_o": "2023-07-17T23:59:00.103", "fields_ar": "0.1002" } 

以上五点问题可以通过SQL进行去噪，整体去噪SQL范本如下：

 SET odps.sql.mapper.split.size = 64; SET odps.stage.joiner.num = 4000; SET odps.stage.reducer.num = 1999; CREATE TABLE table_diff AS SELECT a.fields_as AS fields_as_main ,b.fields_as AS fields_as_branch ,a.fields_at AS fields_at_main ,b.fields_at AS fields_at_branch ,a.fields_d AS fields_d_main ,b.fields_d AS fields_d_branch ,a.fields_i AS fields_i_main ,b.fields_i AS fields_i_branch ,a.fields_j AS fields_j_main ,b.fields_j AS fields_j_branch ,a.fields_aw AS fields_aw_main ,b.fields_aw AS fields_aw_branch ,a.fields_k_json_key AS fields_k_json_key_main ,b.fields_k_json_key AS fields_k_json_key_branch ,a.fields_k_json_key_list AS fields_k_json_key_list_main ,b.fields_k_json_key_list AS fields_k_json_key_list_branch ,CASE WHEN a.fields_k_json_key = b.fields_k_json_key THEN 0 WHEN b.fields_k_json_key_list RLIKE a.fields_k_json_key THEN 0 WHEN a.fields_k_json_key_list RLIKE b.fields_k_json_key THEN 0 ELSE 1 END AS fields_k_json_key_diff_flag FROM ( SELECT fields_as ,fields_at ,fields_d ,fields_i ,fields_j ,fields_aw ,MAX(CASE WHEN rn = 1 THEN fields_k_json_key END) AS fields_k_json_key ,CONCAT_WS(',',COLLECT_SET(fields_k_json_key)) AS fields_k_json_key_list FROM ( SELECT * ,CASE WHEN NVL(GET_JSON_OBJECT(message,'$.fields_k'),'') = '' THEN '---' WHEN GET_JSON_OBJECT(message,'$.fields_k') IN ('0','0.0') THEN '0-0-0' ELSE GET_JSON_OBJECT(message,'$.fields_k') END AS fields_k_json_key ,ROW_NUMBER() OVER (PARTITION BY fields_as,fields_at,fields_d,fields_i,fields_j,fields_aw ORDER BY offset DESC ) AS rn FROM table_main WHERE pt = 20230628 -- AND fields_i = 'realMetric' ) WHERE rn < 6 GROUP BY fields_as ,fields_at ,fields_d ,fields_i ,fields_j ,fields_aw ) a LEFT JOIN ( SELECT fields_as ,fields_at ,fields_d ,fields_i ,fields_j ,fields_aw ,MAX(CASE WHEN rn = 1 THEN fields_k_json_key END) AS fields_k_json_key ,CONCAT_WS(',',COLLECT_SET(fields_k_json_key)) AS fields_k_json_key_list FROM ( SELECT * ,CASE WHEN NVL(GET_JSON_OBJECT(message,'$.fields_k'),'') = '' THEN '---' WHEN GET_JSON_OBJECT(message,'$.fields_k') IN ('0','0.0') THEN '0-0-0' ELSE GET_JSON_OBJECT(message,'$.fields_k') END AS fields_k_json_key ,ROW_NUMBER() OVER (PARTITION BY fields_as,fields_at,fields_d,fields_i,fields_j,fields_aw ORDER BY offset DESC ) AS rn FROM table_branch WHERE pt = 20230628 -- AND fields_i = 'realMetric' and fields_d !='group' ) WHERE rn < 6 GROUP BY fields_as ,fields_at ,fields_d ,fields_i ,fields_j ,fields_aw ) b ON NVL(a.fields_as,'-00') = NVL(b.fields_as,'-00') AND NVL(a.fields_at,'-00') = NVL(b.fields_at,'-00') AND NVL(a.fields_d,'-00') = NVL(b.fields_d,'-00') AND NVL(a.fields_i,'-00') = NVL(b.fields_i,'-00') AND NVL(a.fields_j,'-00') = NVL(b.fields_j,'-00') AND NVL(a.fields_aw,'-00') = NVL(b.fields_aw,'-00') ; 

字段去噪问题

问题现象：涉及字段逻辑修改的情况下，Diff结果是不通过的，影响Diff结果。

去噪方案：需要对逻辑修改的字段抛弃，不再判断发生逻辑修改的字段，通过Java灵活控制。

 String[] jsonColumnListStrings = jsonColumnList.split(","); List<String> jsonColumnLists = new ArrayList<String>(); String[] iterationColumnStrings = iterationColumn.split(","); List<String> iterationColumnLists = Arrays.asList(iterationColumnStrings); for (String s:jsonColumnListStrings){ if(!iterationColumnLists.contains(s)){//判断字段是否为去噪字段 jsonColumnLists.add(s); } } 

六、Diff结果分析

根据主备Diff合成的SQL可以产出对比的结果表，对执行结果分析既可以判断本次执行是否通过。

分析逻辑1：判断每一个对比字段通过占比

提供研发分析哪一个解析的字段通过率低.

分析逻辑2：判断所有字段通过占比总记录数

此指标即可判断本次Diff是否通过，如果占比99.9%，表示通过。

分析SQL样本：

 SELECT round(SUM(CASE WHEN fields_k_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_k_ratio ,round(SUM(CASE WHEN fields_m_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_m_ratio ,round(SUM(CASE WHEN fields_e_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_e_ratio ,round(SUM(CASE WHEN fields_a_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_aratio ,round(SUM(CASE WHEN fields_n_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_n_ratio ,round(SUM(CASE WHEN fields_p_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_p_ratio ,round(SUM(CASE WHEN fields_ac_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_ac_ratio ,round(SUM(CASE WHEN fields_ar_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS fields_ar_ratio ,round(SUM(CASE WHEN fields_k_json_key_diff_flag = 0 AND fields_m_json_key_diff_flag = 0 AND fields_e_json_key_diff_flag = 0 AND fields_a_json_key_diff_flag = 0 AND fields_n_json_key_diff_flag = 0 AND fields_p_json_key_diff_flag = 0 AND fields_ac_json_key_diff_flag = 0 AND fields_ar_json_key_diff_flag = 0 THEN 1 ELSE 0 END) / COUNT(1) * 100,4) AS total_ratio ,COUNT(1) AS total_cnt FROM table_diff ; 

七、工具服务化

后端服务化处理逻辑

主备对比SQL合成

将Diff的SQL植入到代码中，通过代码控制数据切片、去噪等场景，完成测试SQL合成。

 for(String s:jsonColumnLists){ selectSql1 = selectSql1 + " case when NVL(GET_JSON_OBJECT(message,'$." + s + "'),'')='' then '---' when get_json_object(message,'$." + s + "') in ('0','0.0') then '0-0-0' else get_json_object(message,'$." + s + "') end AS " + s + "_json_key,"; selectSql2 = selectSql2 + " max(case when rn =1 then " + s + "_json_key end) as " + s + "_json_key,concat_ws(',',collect_set(" + s + "_json_key)) as " + s + "_json_key_list,"; mergeSql = mergeSql + " a." + s + "_json_key as " + s + "_json_key_main,b." + s + "_json_key as " + s + "_json_key_branch,a." + s + "_json_key_list as " + s + "_json_key_list_main,b." + s + "_json_key_list as " + s + "_json_key_list_branch,case when a." + s + "_json_key = b." + s + "_json_key then 0 when b." + s + "_json_key_list rlike a." + s + "_json_key then 0 when a." + s + "_json_key_list rlike b." + s + "_json_key then 0 else 1 end as " + s + "_json_key_diff_flag,"; } rowNumberSql ="ROW_NUMBER() OVER (PARTITION BY fields_as,fields_at,fields_d,fields_i,fields_j,fields_aw ORDER BY offset DESC ) AS rn "; selectSql1 = selectSql1 + rowNumberSql; whereSql1 = whereSql1 + bizdate + " AND fields_i = 'realMetric' "; String pretreatmentSqlMain = ""; String pretreatmentSqlBranch = ""; pretreatmentSqlBranch = selectSql2.substring(0,selectSql2.length()-1) + " from(" + selectSql1 + " from " + branchLinkTableName + whereSql1 + ")" + whereSql2 + groupSql.substring(0,groupSql.length()-1); pretreatmentSqlMain = selectSql2.substring(0,selectSql2.length()-1) + " from(" + selectSql1 + " from " + masterLinkTableName + whereSql1 + ")" + whereSql2 + groupSql.substring(0,groupSql.length()-1); mergeSql = mergeSql.substring(0,mergeSql.length()-1) + " from (" + pretreatmentSqlMain + ")a left join (" + pretreatmentSqlBranch + ")b " + joinSql.substring(0,joinSql.length()-3) + ";"; SimpleDateFormat simpleDateFormat = new SimpleDateFormat("yyyyMMddHHmmss"); String dateStr = simpleDateFormat.format(new Date()); this.resultDataCreateSql = "set odps.sql.mapper.split.size=64;set odps.stage.joiner.num=4000;set odps.stage.reducer.num=1999; create table du_temp.diff_main_branch_" + dateStr + "_test as " + mergeSql; log.info(resultDataCreateSql); this.resultDataTable = "du_temp.diff_main_branch_" + dateStr + "_test"; log.info(resultDataTable); //合成过滤结果数据的sql String resultSql = " select "; String totalResultSql = "round(sum(case when "; for(String s:jsonColumnLists){ resultSql = resultSql + " round(sum(case when " + s + "_json_key_diff_flag = 0 then 1 else 0 end)/count(1)*100,4) as " + s + "_ratio,"; totalResultSql = totalResultSql + " " + s + "_json_key_diff_flag = 0 and"; } this.resultDataFiltrate = resultSql + totalResultSql.substring(0,totalResultSql.length()-3) + " then 1 else 0 end)/count(1)*100,4) as total_ratio , count(1) as total_cnt from " + this.resultDataTable + ";"; log.info(resultDataFiltrate); 

Diff结果报告解析

 ...} else if(testType.equals("主备diff")) { for (Map.Entry entry:testResultRecord.entrySet()) { List<String> listValue = (List<String>) entry.getValue(); this.resultData.put(entry.getKey().toString(),listValue.get(0)) ; if(Double.parseDouble(listValue.get(0))< 99.9 & !entry.getKey().toString().equals("total_cnt")){ this.failDetail.put(entry.getKey().toString(),listValue.get(0)) ; } } if(failDetail.size()>0){ this.testStatus = "失败"; }else { this.testStatus = "成功"; } } 

平台可视化

• 创建任务

• 执行列表

• 结果报告-平台展示

如下图：一次执行失败的结果，通过率为99.8471，未达到99.99%。

• 结果报告-飞书通知

如下样例：

执行需求名称：主备Diff-521 执行者：*** 执行类型：主备Diff 执行编号：20230628204636 执行备链路表名：table_main 执行主链路表名：table_branch 执行备链路表分区：20230628 执行结果明细表：table_diff 执行结果明细汇总： fields_am_ratio:99.9958 fields_z_ratio:99.9826 fields_af_ratio:99.9856 fields_ba_ratio:99.9964 fields_al_ratio:99.9915 fields_ad_ratio:99.9873 fields_r_ratio:99.9826 fields_aa_ratio:99.9906 fields_ai_ratio:99.9856 fields_v_ratio:99.9917 fields_ak_ratio:99.9909 fields_m_ratio:99.9969 fields_ak_ratio:99.9945 fields_bb_ratio:99.9964 fields_bc_ratio:99.9957 fields_bd_ratio:99.9954 fields_ae_ratio:99.9856 fields_be_ratio:99.9952 fields_bf_ratio:99.9955 fields_t_ratio:99.9917
fields_ag_ratio:99.9856 fields_p_ratio:99.9909 fields_bg_ratio:99.9948 fields_a_ratio:99.9969 fields_d_ratio:99.9969 fields_x_ratio:99.9826 fields_an_ratio:99.9917
fields_ap_ratio:99.9909 fields_ar_ratio:99.9915 fields_y_ratio:99.9917 fields_bh_ratio:99.9955 fields_aj_ratio:99.9916 fields_bi_ratio:99.987 fields_ac_ratio:99.9908 fields_s_ratio:99.9826 fields_ab_ratio:99.9906 fields_i_ratio:99.9969 fields_bj_ratio:99.9951 fields_ah_ratio:99.9959
fields_k_ratio:99.9969
fields_e_ratio:99.9969 fields_bk_ratio:99.9962 fields_bl_ratio:99.8748 fields_al_ratio:99.9958 fields_j_ratio:99.9969
fields_bm_ratio:99.9951 fields_n_ratio:99.9969 fields_ao_ratio:99.9909 fields_w_ratio:99.9906 fields_bn_ratio:99.9965 fields_bo_ratio:99.9912 fields_bcrate_ratio:99.987 fields_y_ratio:99.9958 主备diff执行结果汇总数据： total_ratio:99.8471
total_cnt:714259 执行结果失败明细：
fields_bl_ratio:99.8748
total_ratio:99.8471% 执行结果状态：失败

八、主备diff工具接入发布流程

投放备链路最终经过主备Diff工具测试通过的情况下，完成上线，目前相当于一条备用生产线。

后续版本迭代，需求上线前通过Diff工具验证通过，即可符合上线要求。

九、总结

实时计算不同于离线数仓，数据的稳定性和准确性很难把控，复杂的链路通过简单的测试无法保障整体数据的质量，双链路Diff的形式可以在迭代中更好保障实时数据的质量。

对于主备Diff的实现中：最大的痛点往往是数据的噪点非常的大，需要通过技术手段进行降噪，确保数据对比结果的准确性和可靠性。

*文/诗雨

本文属得物技术原创，更多精彩文章请看：得物技术官网

未经得物技术许可严禁转载，否则依法追究法律责任！