调参原则

• 禁止通过增加执行资源的方式来达成任务的目的

  • 比如OOM首先考虑的应该是降低每个task数目量，而不是增加内存

• 大多数参数调整无大意义，线上场景也比较复杂，没必要精细化调整每个参数到最优，保持一类任务的参数大致统一才是最佳方案

  • 不出现严重的执行缓慢/task失败/运行不稳定即可

• 尽可能控制任务的资源使用

参数推荐

参数划分标准：根据收集到的资源元数据判定

重要参数详解

1. set spark.sql.adaptive.maxNumPostShufflePartitions =

   2. AE shuffle的最大并行度（也是初始并行度）

   3. 该值过高会引起shuffle过程颗粒度过小而引起资源浪费

   4. 建议值：根据业务场景采用1000 - 5000即可

   5. Note：高于5000并行度的场景可以考虑关闭AE采用spark原生shuffle

2. set spark.sql.adaptive.shuffle.targetPostShuffleInputSize = 268435456/536870912;

   1. shuffle read的期望输入最大值，根据计算逻辑决定，复杂逻辑（有需要大量资源的udf使用；连续按相同key join）采用更小的256M，其他采用512M即可
   2. 不要设置该值过高，该值过高会引起spill
   3. 若shuffle read task的spill过高致使task缓慢，降低该值

3. set spark.sql.files.maxPartitionBytes = 268435456/536870912/

4. 可认为是input read的最大期望值，调整逻辑同上

5. 若input task的spill过高致使task缓慢，降低该值

6. set spark.dynamicAllocation.maxExecutors =

   1. 参与计算的最大executor个数，根据任务量和任务重要程度确定

7. set spark.memory.fraction=;

   1. spark用于计算和缓存rdd的内存比例，该值默认0.6，TQS置为0.7，不建议调整该值

8. set spark.shuffle.hdfs.enabled=true;

   1. 建议在大任务或者重要任务中开启，避免shuffle fetch fail

9. set spark.shuffle.io.maxRetries=1;（该参数禁止设置为0，会导致严重数据错误）

10. set spark.shuffle.io.retryWait=0s;

11. set spark.network.timeout=120s;

12. 启用hdfs shuffle时配合使用，获取local shuffle文件失败立刻路由到hdfs

13. set spark.speculation.multiplier=2.5;

14. set spark.speculation.quantile=0.9;

15. spark推测执行参数，不修改该参数会导致大量speculation task kill，浪费资源

16. set spark.merge.files.enabled=true;

17. set spark.merge.files.number=期望文件个数

广播设置

1. adaptive execution提供的广播，执行过程中动态调整执行计划，必须开启spark.sql.adaptive.enabled才会有效。

2. spark.sql.adaptive.join.enabled=false

3. 是否启用adatptive execution自动广播

4. spark.sql.adaptiveBroadcastJoinThreshold=20971520

5. AE自动广播的阈值，单位byte，对应task的map output size

6. spark非adatptive的广播方式

7. spark.sql.autoBroadcastJoinThreshold=20971520

8. 优先读取hive meta信息；FileScan模式下可以扫描文件size，单位byte（注意是整表size，而不是仅仅所需列的size；text格式表只能走hive serde所以此处不支持分区级别广播，只能通过AE广播）

小型任务：

目前测试：在不手动添加任何参数、平均时长在30min以内、单个shuffle量在500G以下的任务可以使用该模版，但实际任务情况还需跟踪观察。

基础资源
set spark.driver.memory=15g;
set spark.driver.memoryOverhead=4096;
set spark.driver.memoryOverhead=4096;
set spark.executor.memory=5G;
set spark.executor.memoryOverhead=1024;
set spark.executor.cores=2;
set spark.vcore.boost.ratio=2;
--动态executor申请
set spark.dynamicAllocation.minExecutors=10;
set spark.dynamicAllocation.maxExecutors=300;
--ae shuffle partition并⾏度
set spark.sql.adaptive.minNumPostShufflePartitions=10;
set spark.sql.adaptive.maxNumPostShufflePartitions=1000;

--268435456;
set spark.sql.adaptive.shuffle.targetPostShuffleInputSize=536870912;
--开启parquet切分
set spark.sql.parquet.adaptiveFileSplit=true;
--初始task调节，合并⼩⽂件
set spark.sql.files.maxPartitionBytes=536870912;

中型任务：

目前测试：在不手动添加任何参数、平均时长在90min以内、单个shuffle量在2T以下的任务可以使用该模版，但实际任务情况还需跟踪观察。

--
基础资源
 
set spark.driver.memory=25g;
set spark.driver.cores=4;
set spark.driver.memoryOverhead=5120;
set spark.executor.memory=10G;
set spark.executor.memoryOverhead=4096;
set spark.executor.cores=3;
set spark.vcore.boost.ratio=1;
--动态executor申请
 
set spark.dynamicAllocation.minExecutors=10;
set spark.dynamicAllocation.maxExecutors=600;
--ae
set spark.sql.adaptive.minNumPostShufflePartitions=10;
set spark.sql.adaptive.maxNumPostShufflePartitions=1000;
set spark.sql.adaptive.shuffle.targetPostShuffleInputSize= 536870912;
--开启parquet切分，初始task调节，合并⼩⽂件
set spark.sql.parquet.adaptiveFileSplit=true;
set spark.sql.files.maxPartitionBytes=536870912;
--推测
set spark.speculation.multiplier=2.5;
set spark.speculation.quantile=0.8;
--shuffle 落地hdfs 
set spark.shuffle.hdfs.enabled=true;
set spark.shuffle.io.maxRetries=1;
set spark.shuffle.io.retryWait=0s;

大型任务：

目前测试：在不手动添加任何参数、平均时长在120min以内、单个shuffle量在10T以下的任务可以使用该模版，但实际任务情况还需跟踪观察。

--
基础资源
set spark.driver.memory=25g;
set spark.driver.cores=4;
set spark.driver.memoryOverhead=5120;
set spark.executor.memory=15G;
set spark.executor.memoryOverhead=3072;
set spark.executor.cores=3;
set spark.vcore.boost.ratio=1;
--动态executor申请
set spark.dynamicAllocation.minExecutors=10;
set spark.dynamicAllocation.maxExecutors=900;
--ae
set spark.sql.adaptive.minNumPostShufflePartitions=10;
set spark.sql.adaptive.maxNumPostShufflePartitions=3000;
set spark.sql.adaptive.shuffle.targetPostShuffleInputSize= 536870912;
--shuffle 落地hdfs 
set spark.shuffle.hdfs.enabled=true;
set spark.shuffle.io.maxRetries=1;
set spark.shuffle.io.retryWait=0s;
--开启parquet切分，合并⼩⽂件
set spark.sql.parquet.adaptiveFileSplit=true; 
set spark.sql.files.maxPartitionBytes=536870912;
--推测
set spark.speculation.multiplier=2.5;
set spark.speculation.quantile=0.9;

超大型任务：

目前测试：在不手动添加任何参数、平均时长大于120min、单个shuffle量在10T以上的任务可以使用该模版，但实际任务情况还需跟踪观察。

--
基础资源
set spark.driver.memory=30g;
set spark.driver.cores=4;
set spark.driver.memoryOverhead=5120;
set spark.executor.memory=20G;
set spark.executor.memoryOverhead= 5120;
set spark.executor.cores=5;
set spark.vcore.boost.ratio=1;
--动态executor申请
set spark.dynamicAllocation.minExecutors=10;
set spark.dynamicAllocation.maxExecutors=1500;
--ae
set spark.sql.adaptive.minNumPostShufflePartitions=10;
set spark.sql.adaptive.maxNumPostShufflePartitions=7000;
set spark.sql.adaptive.shuffle.targetPostShuffleInputSize= 536870912;
--开启parquet切分,合并小文件
set spark.sql.parquet.adaptiveFileSplit=true;
set spark.sql.files.maxPartitionBytes=536870912;
-- shuffle 落地hdfs 
set spark.shuffle.hdfs.enabled=true;
set spark.shuffle.io.maxRetries=1;
set spark.shuffle.io.retryWait=0s;
--推测
set spark.speculation.multiplier=2.5;
set spark.speculation.quantile=0.9;

备用

--
预留
 
set spark.yarn.priority = 9;
--ae hash join
set spark.sql.adaptive.hashJoin.enabled=true;
set spark.sql.adaptiveHashJoinThreshold=52428800;
--输出文件合并                
set spark.merge.files.byBytes.enabled=true;
set spark.merge.files.byBytes.repartitionNumber=100;
set spark.merge.files.byBytes.fileBytes=134217728;
set spark.merge.files.byBytes.compressionRatio=3;
--skew_join 解析绕过tqs 
set tqs.analysis.skip.hint=true;
--初始task上限
 
set spark.sql.files.openCostInBytes=4194304;
set spark.datasource.splits.max=20000;
--broadcast时间
set spark.sql.broadcastTimeout = 3600;
--（防止get json报错）
set spark.sql.mergeGetMapValue.enabled=true;
--ae 倾斜处理HandlingSkewedJoin  OptimizeSkewedJoin 
set spark.sql.adaptive.allowBroadcastExchange.enabled=true;
set spark.sql.adaptive.hashJoin.enabled=false;
set spark.sql.adaptive.skewedPartitionFactor=3;
set spark.sql.adaptive.skewedPartitionMaxSplits=20;
set spark.sql.adaptive.skewedJoin.enabled=true;
set spark.sql.adaptive.skewedJoinWithAgg.enabled=true;
set spark.sql.adaptive.multipleSkewedJoin.enabled=true;
set spark.shuffle.highlyCompressedMapStatusThreshold=20000;
--并发读文件
set spark.sql.concurrentFileScan.enabled=true;
--filter 按照比例读文件
 
set spark.sql.files.tableSizeFactor={table_name}:{filter 
35 set spark.sql.files.tableSizeFactor=dm_content.tcs_task_dict:10;
--AM failed  时长
set spark.yarn.am.waitTime=200s;
--shuffle service 超时设置
set spark.shuffle.registration.timeout=12000;
set spark.shuffle.registration.maxAttempts=5;
--parquet index  
set spark.sql.parquet.pushdown.inFilterThreshold=30; 
--设置engine 
set tqs.query.engine.type=sparkcli;
--hive metastore 超时
spark.hadoop.hive.metastore.client.socket.timeout=600
--manta备用
spark.sql.adaptive.maxNumPostShufflePartitions 5000
spark.executor.memoryOverhead 8000
spark.sql.adaptive.shuffle.targetPostShuffleInputSize 536870912

默认参数（一些较为重要的参数）

贴这个是为了防止冗余写。

基础资源相关

//基础资源
spark.driver.memory	12G
spark.driver.memoryOverhead	6144
spark.executor.memory	8G
spark.executor.memoryOverhead	6144
spark.executor.cores	4
//动态资源
spark.dynamicAllocation.enabled	true
spark.dynamicAllocation.executorIdleTimeout	120s
spark.dynamicAllocation.initialExecutors	5
spark.dynamicAllocation.maxExecutors	900
spark.dynamicAllocation.minExecutors	5

jvm相关

spark.executor.extraJavaOptions	-XX:+UseG1GC -XX:G1HeapRegionSize=4m -XX: UseGCOverheadLimit -verbose:gc -XX:+PrintAdaptiveSizePolicy XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xss4m XX:MaxDirectMemorySize=4096m
// shuffle相关，spark自带
spark.shuffle.consolidateFiles	true
spark.shuffle.hdfs.rootDir	hdfs://haruna/spark_topi/shuffle/hdfs_dancen
spark.shuffle.highlyCompressedMapStatusThreshold	2000
spark.shuffle.service.enabled	true
spark.shuffle.statistics.verbose	true
spark.speculation	true

Spark ae相关

spark.sql.adaptive.enabled	true
spark.sql.adaptive.join.enabled	true
spark.sql.adaptive.maxNumPostShufflePartitions	500
spark.sql.adaptive.minNumPostShufflePartitions	1
spark.sql.adaptive.shuffle.targetPostShuffleInputSize	67108864 //64MB
spark.sql.adaptive.shuffle.targetPostShuffleRowCount	20000000
spark.sql.adaptive.skewedJoin.enabled	false
spark.sql.adaptive.skewedPartitionFactor	10
spark.sql.adaptive.skewedPartitionRowCountThreshold	10000000
spark.sql.adaptive.skewedPartitionSizeThreshold	67108864
spark.sql.adaptiveBroadcastJoinThreshold
spark.sql.autoBroadcastJoinThreshold	20971520
//初始文件合并
spark.sql.files.maxPartitionBytes	1073741824//1GB
spark.sql.files.openCostInBytes	16777216
// 其他
spark.sql.hive.caseSensitiveInferenceMode N	NEVER_INFER
spark.sqlhive.convertMetastoreOrc	true
.spark.sqlhiveconvertMetastoreParquet	true
.. sparkslparquetadaptiveFileSlit	false
.q..pspark.sqlparquetcompressioncodec	gzip
...sparksqlparquetfilterPushdown	true
... spark.sql.parquet.pushdown.inFilterThreshold	20
spark.sql.skip.adjust.partitioned.file	true
spark.sql.sources.bucketing.enabled	false
//spark 容错以及 yarn 优先级	8
spark.task.maxFailures	1
spark.yarn.maxAppAttempts	1
spark.yarn.priority

贡献者

codingLab

版权所有

版权归属：codingLab

许可证：bugcode