FlinkSQL【分组聚合-多维分析-性能调优】应用实例分析

发布时间:2024年01月16日

FlinkSQL处理如下实时数据需求:
实时聚合不同 类型/账号/发布时间 的各个指标数据,比如:初始化/初始化后删除/初始化后取消/推送/成功/失败 的指标数据。要求实时产出指标数据,数据源是mysql cdc binlog数据。

代码实例

--SET table.exec.state.ttl=86400s; --24 hour,默认: 0 ms
SET table.exec.state.ttl=2592000s; --30 days,默认: 0 ms
--MiniBatch 聚合
SET table.exec.mini-batch.enabled = true;
SET table.exec.mini-batch.allow-latency = 1s;
SET table.exec.mini-batch.size = 10000;
--Local-Global 聚合
SET table.optimizer.agg-phase-strategy = TWO_PHASE;

CREATE TABLE kafka_table (
     mid bigint,
     db string,
     sch string,
     tab string,
     opt string,
     ts bigint,
     ddl string,
     err string,
     src map<string,string>,
     cur map<string,string>,
     cus map<string,string>,
     account_id AS IF(cur['account_id'] IS NOT NULL , cur['account_id'], src ['account_id']),
     publish_time AS IF(cur['publish_time'] IS NOT NULL , cur['publish_time'], src ['publish_time']),
     msg_status AS IF(cur['msg_status'] IS NOT NULL , cur['msg_status'], src ['msg_status']),
     send_type AS IF(cur['send_type'] IS NOT NULL , cur['send_type'], src ['send_type'])
     --event_time as cast(IF(cur['update_time'] IS NOT NULL , cur['update_time'], src ['update_time']) AS TIMESTAMP(3)), -- TIMESTAMP(3)/TIMESTAMP_LTZ(3)
     --WATERMARK FOR event_time AS event_time - INTERVAL '1' MINUTE     --SECOND
) WITH (
  'connector' = 'kafka',
  'topic' = 't1',
  'properties.bootstrap.servers' = 'xx.xx.xx.xx:9092',
  'properties.group.id' = 'g1',
  'scan.startup.mode' = 'earliest-offset',  --group-offsets/earliest-offset/latest-offset
   --  'properties.enable.auto.commit',= 'true' -- default:false, 如果为false,则在发生checkpoint时触发offset提交
  'format' = 'json'
);



CREATE TABLE es_sink(
     send_type      STRING
    ,account_id     STRING
    ,publish_time   STRING
    ,grouping_id       INTEGER
    ,init           INTEGER
    ,init_cancel    INTEGER
    ,push          INTEGER
    ,succ           INTEGER
    ,fail           INTEGER
    ,init_delete    INTEGER
    ,update_time    STRING
    ,PRIMARY KEY (group_id,send_type,account_id,publish_time) NOT ENFORCED
)
with (
    'connector' = 'elasticsearch-6',
    'index' = 'es_sink',
    'document-type' = 'es_sink',
    'hosts' = 'http://xxx:9200',
    'format' = 'json',
    'filter.null-value'='true',
    'sink.bulk-flush.max-actions' = '1000',
    'sink.bulk-flush.max-size' = '10mb'
);

CREATE view  tmp as
select
    send_type,
    account_id,
    publish_time,
    msg_status,
    case when UPPER(opt) = 'INSERT' and msg_status='0'  then 1 else 0 end AS init,
    case when UPPER(opt) = 'UPDATE' and send_type='1' and msg_status='4' then 1 else 0 end AS init_cancel,
    case when UPPER(opt) = 'UPDATE' and msg_status='3' then 1 else 0 end AS push,
    case when UPPER(opt) = 'UPDATE' and (msg_status='1' or msg_status='5') then 1 else 0 end AS succ,
    case when UPPER(opt) = 'UPDATE' and (msg_status='2' or msg_status='6') then 1 else 0 end AS fail,
    case when UPPER(opt) = 'DELETE' and send_type='1' and msg_status='0' then  1 else 0 end AS init_delete,
    event_time,
    opt,
    ts
FROM kafka_table
where (UPPER(opt) = 'INSERT' and msg_status='0' )
or        (UPPER(opt) = 'UPDATE' and msg_status in ('1','2','3','4','5','6'))
or        (UPPER(opt) = 'DELETE' and send_type='1' and msg_status='0');


--send_type=1          send_type=0
--初始化->0             初始化->0
--取消->4
--推送->3               推送->3
--成功->1               成功->5
--失败->2               失败->6

CREATE view  tmp_groupby as
select
 COALESCE(send_type,'N') AS send_type
,COALESCE(account_id,'N') AS account_id
,COALESCE(publish_time,'N') AS publish_time
,case when send_type is null and account_id is null and publish_time is null then 1
         when send_type is not null and account_id is null and publish_time is null then 2
         when send_type is not null and account_id is not null and publish_time is null then 3
         when send_type is not null and account_id is not null and publish_time is not null then 4
         end grouping_id
,sum(init) as init
,sum(init_cancel) as init_cancel
,sum(push) as push
,sum(succ) as succ
,sum(fail) as fail
,sum(init_delete) as init_delete
from tmp
--GROUP BY GROUPING SETS ((send_type,account_id,publish_time), (send_type,account_id),(send_type), ())
GROUP BY ROLLUP (send_type,account_id,publish_time); --等同于以上

INSERT INTO es_sink
select
     send_type
    ,account_id
    ,publish_time
    ,grouping_id
    ,init
    ,init_cancel
    ,push
    ,succ
    ,fail
    ,init_delete
    ,CAST(LOCALTIMESTAMP AS STRING) as update_time
from tmp_groupby

其他配置

  • flink集群参数
state.backend: rocksdb
state.backend.incremental: true
state.backend.rocksdb.ttl.compaction.filter.enabled: true
state.backend.rocksdb.localdir: /export/io_tmp_dirs/rocksdb
state.checkpoints.dir: hdfs://namenode-host:port/flink-checkpoints
state.savepoints.dir: hdfs://namenode-host:port/flink-savepoints
rest.flamegraph.enabled: true
pipeline.operator-chaining: false
taskmanager.memory.managed.fraction: 0.7
taskmanager.memory.network.min: 128 mb
taskmanager.memory.network.max: 128 mb
taskmanager.memory.framework.off-heap.size: 32mb
taskmanager.memory.task.off-heap.size: 32mb
taskmanager.memory.jvm-metaspace.size: 256mb
taskmanager.memory.jvm-overhead.fraction: 0.03
  • 检查点配置
    在这里插入图片描述

  • job运行资源
    管理节点(JM) 1 个, 节点规格 1 核 4 GB内存, 磁盘 10Gi
    运行节点(TM)10 个, 节点规格 1 核 4 GB内存, 磁盘 80Gi
    单TM槽位数(Slot): 1
    默认并行度:8

  • es mapping

#POST app_cust_syyy_private_domain_syyy_group_msg/app_cust_syyy_private_domain_syyy_group_msg/_mapping
{
    "app_cust_syyy_private_domain_syyy_group_msg": {
        "properties": {
            "send_type": {
                "type": "keyword",
                "ignore_above": 256
            },
            "account_id": {
                "type": "keyword"
            },
           "publish_time": {
           	"type": "keyword",
           	"fields": {
           		"text": {
           			"type": "keyword"
           		},
           		"date": {
           			"type": "date",
           			"format": "yyyy-MM-dd HH:mm:ss.SSS||yyyy-MM-dd HH:mm:ss||yyyy-MM-dd||epoch_millis",
           			"ignore_malformed":"true" # 忽略错误的各式
           		}
           	}
           },
            "grouping_id": {
                "type": "integer"
            },
            "init": {
                "type": "integer"
            },
            "init_cancel": {
                "type": "integer"
            },
            "query": {
                "type": "integer"
            },
            "succ": {
                "type": "integer"
            },
            "fail": {
                "type": "integer"
            },
            "init_delete": {
                "type": "integer"
            },
            "update_time": {
                "type": "date",
                "format": "yyyy-MM-dd HH:mm:ss.SSS||yyyy-MM-dd HH:mm:ss||yyyy-MM-dd||epoch_millis"
            }
        }
    }
}

性能调优

是否开启【MiniBatch 聚合】和【Local-Global 聚合】对分组聚合场景影响巨大,尤其是在数据量大的场景下。

  • 如果未开启,在分组聚合,数据更新状态时,每条数据都会触发聚合运算,进而更新StateBackend (尤其是对于 RocksDB StateBackend,火焰图上反映就是一直在update rocksdb),造成上游算子背压特别大。此外,生产中非常常见的数据倾斜会使这个问题恶化,并且容易导致 job 发生反压。
    在这里插入图片描述

  • 在开启【MiniBatch 聚合】和【Local-Global 聚合】后,配置如下:

--MiniBatch 聚合
SET table.exec.mini-batch.enabled = true;
SET table.exec.mini-batch.allow-latency = 1s;
SET table.exec.mini-batch.size = 10000;
--Local-Global 聚合
SET table.optimizer.agg-phase-strategy = TWO_PHASE;

开启配置好会在DAG上添加两个环节MiniBatchAssignerLocalGroupAggregate
在这里插入图片描述

对结果的影响

开启了【MiniBatch 聚合】和【Local-Global 聚合】后,一天处理不完的数据,在10分钟内处理完毕

输出结果

在这里插入图片描述在这里插入图片描述

参考:
Group Aggregation
Streaming Aggregation Performance Tuning

文章来源:https://blog.csdn.net/weixin_38251332/article/details/135617562
本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。