是的，可以测一下，理论上 mr writer不应该有较大性能差距。
> 为何要强制滚动文件 因为要保证Exactly-Once， 像Orc和parquet类似的 format，它并不能把一个文件拆成多次来写。 On Thu, Sep 17, 2020 at 2:05 PM kandy.wang <kandy1...@163.com> wrote: > > > > ok. 就是用hadoop mr writer vs flink 自实现的native > writer之间的性能对比了。至少目前看了一下table.exec.hive.fallback-mapred-writer > 改成false是可以满足我们的写hive需求了 > 还有一个问题，之前问过你，你还没回复： > HiveRollingPolicy为什么 shouldRollOnCheckpoint true 为何要强制滚动文件，这个可以抽取成一个配置参数么？ > 如果强制滚动的话，基本上sink.rolling-policy.rollover-interval、 > sink.rolling-policy.rollover-interval参数就不work了，如果5min一个分区，2min做一次checkpoint，那文件还不到几十M就滚动了。配置的参数就没意义了 > 在 2020-09-17 13:43:04，"Jingsong Li" <jingsongl...@gmail.com> 写道： > >可以再尝试下最新的1.11.2吗？ > > > >https://flink.apache.org/downloads.html > > > >On Thu, Sep 17, 2020 at 1:33 PM kandy.wang <kandy1...@163.com> wrote: > > > >> 是master分支代码 > >> 那你说的这个情况，刚好是table.exec.hive.fallback-mapred-writer默认是true 的情况 > >> 出现的，现在改成false 就走到else 部分 就暂时没这个问题了 > >> if (userMrWriter) { > >> builder = bucketsBuilderForMRWriter(recordWriterFactory, sd, > assigner, > >> rollingPolicy, outputFileConfig); > >> LOG.info("Hive streaming sink: Use MapReduce RecordWriter writer."); > >> } else { > >> Optional<BulkWriter.Factory<RowData>> bulkFactory = > >> createBulkWriterFactory(partitionColumns, sd); > >> if (bulkFactory.isPresent()) { > >> builder = StreamingFileSink.forBulkFormat( > >> new org.apache.flink.core.fs.Path(sd.getLocation()), > >> new > >> FileSystemTableSink.ProjectionBulkFactory(bulkFactory.get(), > partComputer)) > >> .withBucketAssigner(assigner) > >> .withRollingPolicy(rollingPolicy) > >> .withOutputFileConfig(outputFileConfig); > >> LOG.info("Hive streaming sink: Use native parquet&orc writer."); > >> } else { > >> builder = bucketsBuilderForMRWriter(recordWriterFactory, sd, > >> assigner, rollingPolicy, outputFileConfig); > >> LOG.info("Hive streaming sink: Use MapReduce RecordWriter writer because > >> BulkWriter Factory not available."); > >> } > >> } > >> 在 2020-09-17 13:21:40，"Jingsong Li" <jingsongl...@gmail.com> 写道： > >> >是最新的代码吗？ > >> >1.11.2解了一个bug：https://issues.apache.org/jira/browse/FLINK-19121 > >> >它是影响性能的，1.11.2已经投票通过，即将发布 > >> > > >> >On Thu, Sep 17, 2020 at 12:46 PM kandy.wang <kandy1...@163.com> wrote: > >> > > >> >> @Jingsong Li > >> >> > >> >> public TableSink createTableSink(TableSinkFactory.Context context) { > >> >> CatalogTable table = checkNotNull(context.getTable()); > >> >> Preconditions.checkArgument(table instanceof CatalogTableImpl); > >> >> > >> >> boolean isGeneric = > >> >> > >> > Boolean.parseBoolean(table.getProperties().get(CatalogConfig.IS_GENERIC)); > >> >> > >> >> if (!isGeneric) { > >> >> return new HiveTableSink( > >> >> context.getConfiguration().get( > >> >> > HiveOptions.TABLE_EXEC_HIVE_FALLBACK_MAPRED_WRITER), > >> >> context.isBounded(), > >> >> new JobConf(hiveConf), > >> >> context.getObjectIdentifier(), > >> >> table); > >> >> } else { > >> >> return TableFactoryUtil.findAndCreateTableSink(context); > >> >> } > >> >> } > >> >> > >> >> > >> > HiveTableFactory中，有个配置table.exec.hive.fallback-mapred-writer默认是true，控制是否使用Hadoop > >> >> 自带的mr writer还是用flink native 实现的 writer去写orc parquet格式。 > >> >> > >> >> If it is false, using flink native writer to write parquet and orc > >> files; > >> >> > >> >> If it is true, using hadoop mapred record writer to write parquet and > >> orc > >> >> files > >> >> > >> >> 将此参数调整成false后，同样的资源配置下，tps达到30W > >> >> > >> >> 这个不同的ORC实现，可能性能本身就存在差异吧？ 另外我们的存储格式是orc，orc有没有一些可以优化的参数，async flush > >> >> 一些相关的参数 ？ > >> >> > >> >> > >> >> > >> >> > >> >> > >> >> 在 2020-09-17 11:21:43，"Jingsong Li" <jingsongl...@gmail.com> 写道： > >> >> >Sink并行度 > >> >> >我理解是配置Sink并行度，这个一直在讨论，还没结论 > >> >> > > >> >> >HDFS性能 > >> >> >具体可以看HDFS到底什么瓶颈，是网络还是请求数还是连接数还是磁盘IO > >> >> > > >> >> >On Wed, Sep 16, 2020 at 8:16 PM kandy.wang <kandy1...@163.com> > wrote: > >> >> > > >> >> >> 场景很简单，就是kafka2hive > >> >> >> --5min入仓Hive > >> >> >> > >> >> >> INSERT INTO hive.temp_.hive_5min > >> >> >> > >> >> >> SELECT > >> >> >> > >> >> >> arg_service, > >> >> >> > >> >> >> time_local > >> >> >> > >> >> >> ..... > >> >> >> > >> >> >> FROM_UNIXTIME((UNIX_TIMESTAMP()/300 * 300) ,'yyyyMMdd'), > >> >> >> FROM_UNIXTIME((UNIX_TIMESTAMP()/300 * 300) ,'HHmm') 5min产生一个分区 > >> >> >> > >> >> >> FROM hive.temp_.kafka_source_pageview/*+ OPTIONS(' > >> properties.group.id > >> >> '='kafka_hive_test', > >> >> >> 'scan.startup.mode'='earliest-offset') */; > >> >> >> > >> >> >> > >> >> >> > >> >> >> --kafka source表定义 > >> >> >> > >> >> >> CREATE TABLE hive.temp_vipflink.kafka_source_pageview ( > >> >> >> > >> >> >> arg_service string COMMENT 'arg_service', > >> >> >> > >> >> >> .... > >> >> >> > >> >> >> )WITH ( > >> >> >> > >> >> >> 'connector' = 'kafka', > >> >> >> > >> >> >> 'topic' = '...', > >> >> >> > >> >> >> 'properties.bootstrap.servers' = '...', > >> >> >> > >> >> >> 'properties.group.id' = 'flink_etl_kafka_hive', > >> >> >> > >> >> >> 'scan.startup.mode' = 'group-offsets', > >> >> >> > >> >> >> 'format' = 'json', > >> >> >> > >> >> >> 'json.fail-on-missing-field' = 'false', > >> >> >> > >> >> >> 'json.ignore-parse-errors' = 'true' > >> >> >> > >> >> >> ); > >> >> >> --sink hive表定义 > >> >> >> CREATE TABLE temp_vipflink.vipflink_dm_log_app_pageview_5min ( > >> >> >> .... > >> >> >> ) > >> >> >> PARTITIONED BY (dt string , hm string) stored as orc location > >> >> >> 'hdfs://ssdcluster/....._5min' TBLPROPERTIES( > >> >> >> 'sink.partition-commit.trigger'='process-time', > >> >> >> 'sink.partition-commit.delay'='0 min', > >> >> >> 'sink.partition-commit.policy.class'='...CustomCommitPolicy', > >> >> >> > >> 'sink.partition-commit.policy.kind'='metastore,success-file,custom', > >> >> >> 'sink.rolling-policy.check-interval' ='30s', > >> >> >> 'sink.rolling-policy.rollover-interval'='10min', > >> >> >> 'sink.rolling-policy.file-size'='128MB' > >> >> >> ); > >> >> >> 初步看下来，感觉瓶颈在写hdfs，hdfs 这边已经是ssd hdfs了，kafka的分区数=40 > >> >> >> ，算子并行度=40，tps也就达到6-7万这样子，并行度放大，性能并无提升。 > >> >> >> 就是flink sql可以 > >> >> >> > >> >> > >> > 改局部某个算子的并行度，想单独改一下StreamingFileWriter算子的并行度，有什么好的办法么？然后StreamingFileWriter > >> >> >> 这块，有没有什么可以提升性能相关的优化参数？ > >> >> >> > >> >> >> > >> >> >> > >> >> >> > >> >> >> 在 2020-09-16 19:29:50，"Jingsong Li" <jingsongl...@gmail.com> 写道： > >> >> >> >Hi, > >> >> >> > > >> >> >> >可以分享下具体的测试场景吗？有对比吗？比如使用手写的DataStream作业来对比下，性能的差距？ > >> >> >> > > >> >> >> >另外，压测时是否可以看下jstack？ > >> >> >> > > >> >> >> >Best, > >> >> >> >Jingsong > >> >> >> > > >> >> >> >On Wed, Sep 16, 2020 at 2:03 PM kandy.wang <kandy1...@163.com> > >> wrote: > >> >> >> > > >> >> >> >> 压测下来，发现streaming方式写入hive StreamingFileWriter ，在kafka > partition=40 > >> >> >> ，source > >> >> >> >> writer算子并行度 =40的情况下，kafka从头消费，tps只能达到 7w > >> >> >> >> 想了解一下，streaming方式写Hive 这块有压测过么？性能能达到多少 > >> >> >> > > >> >> >> > > >> >> >> > > >> >> >> >-- > >> >> >> >Best, Jingsong Lee > >> >> >> > >> >> > > >> >> > > >> >> >-- > >> >> >Best, Jingsong Lee > >> >> > >> > > >> > > >> >-- > >> >Best, Jingsong Lee > >> > > > > > >-- > >Best, Jingsong Lee > -- Best, Jingsong Lee
