另外,join 节点的并发可以再增加一些,提升 join 的处理性能。
On Wed, 18 Nov 2020 at 11:34, Jark Wu <[email protected]> wrote: > 估计是你的全量数据太大,但是 join 节点处理性能太慢,导致半小时内全量数据还没有处理完,而 checkpoint 超时了。 > 注意全量数据阶段,是做不了 checkpoint 的。 具体可以看下这篇文章的第四点。 > https://mp.weixin.qq.com/s/Mfn-fFegb5wzI8BIHhNGvQ > > 解决办法文中也有提及: > > 解决办法:在 flink-conf.yaml 配置 failed checkpoint 容忍次数,以及失败重启策略,如下: > > execution.checkpointing.interval: 10min # checkpoint间隔时间 > execution.checkpointing.tolerable-failed-checkpoints: 100 # checkpoint > 失败容忍次数 > restart-strategy: fixed-delay # 重试策略 > restart-strategy.fixed-delay.attempts: 2147483647 # 重试次数 > > Best, > Jark > > On Wed, 18 Nov 2020 at 11:28, 丁浩浩 <[email protected]> wrote: > >> 即使我将not >> exists改成了join,join节点的checkpoint也无法完成,是我设置的checkpint时间太短了嘛,我设置的是每隔半小时发起checkpoint一次,超时时间也是半小时。 >> 下面是截图,(我上传图片每次都看不了啥情况) >> https://imgchr.com/i/DeqixU >> https://imgchr.com/i/DeqP2T >> >> > 在 2020年11月16日,上午10:29,Jark Wu <[email protected]> 写道: >> > >> > 瓶颈应该在两个 not exists 上面,not exists 目前只能单并发来做,所以无法水平扩展性能。 >> > 可以考虑把 not exists 替换成其他方案,比如 udf,维表 join。 >> > >> > Best, >> > Jark >> > >> > On Mon, 16 Nov 2020 at 10:05, 丁浩浩 <[email protected]> wrote: >> > >> >> select >> >> ri.sub_clazz_number, >> >> prcrs.rounds, >> >> count(*) as num >> >> from >> >> subclazz gs >> >> JOIN >> >> (SELECT gce.number, min( gce.extension_value ) AS grade FROM >> >> course_extension gce WHERE gce.isdel = 0 AND gce.extension_type = 4 >> GROUP >> >> BY gce.number) AS temp >> >> ON >> >> temp.number = gs.course_number AND temp.grade>30 >> >> JOIN >> >> right_info ri >> >> ON >> >> gs.number = ri.sub_clazz_number >> >> join >> >> wide_subclazz ws >> >> on >> >> ws.number = ri.sub_clazz_number >> >> join >> >> course gc >> >> on >> >> gc.number = ws.course_number and gc.course_category_id in (30,40) >> >> left join >> >> performance_regular_can_renewal_sign prcrs >> >> on prcrs.order_number = ri.order_number and prcrs.rounds in (1,2) >> >> where ri.is_del = 0 and ri.end_idx = -1 and prcrs.rounds is not null >> >> and not exists (select 1 from internal_staff gis where gis.user_id = >> >> ri.user_id) >> >> and not exists (select 1 from clazz_extension ce where ws.clazz_number >> = >> >> ce.number >> >> and ce.extension_type = 3 and ce.isdel = 0 >> >> and ce.extension_value in (1,3,4,7,8,11)) >> >> group by ri.sub_clazz_number, prcrs.rounds >> >> Sql代码是这样的。 >> >> 瓶颈在所有的join节点上,每次的checkpoint无法完成的节点都是join节点。 >> >> >> >>> 在 2020年11月14日,下午5:53,Jark Wu <[email protected]> 写道: >> >>> >> >>> 能展示下你的代码吗?是用的维表关联的语法 (FOR SYSTEM TIME AS OF)? >> >>> 需要明确下,到底是什么节点慢了。 >> >>> >> >>> On Fri, 13 Nov 2020 at 19:02, 丁浩浩 <[email protected]> wrote: >> >>> >> >>>> 我用flink cdc 对mysql的表进行关联查询,发现flink只能两张表关联之后再跟下一张表关联,导致最后落库的延迟非常大。 >> >>>> 有没有比较好的优化方案能缓解这样的问题? >> >> >> >> >> >> >> >> >>
