–基础资源
set spark.driver.memory=15g;
set spark.driver.cores=3;
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。
spark.executor.memoryOverhead 每个executor的堆外内存大小，堆外内存主要用于数据IO，对于报堆外OOM的任务要适当调大，单位Mb，与之配合要调大executor JVM参数，例如：set spark.executor.memoryOverhead=3072
set spark.executor.extraJavaOptions=-XX:MaxDirectMemorySize=2560m
–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;
–AQE
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，shuffle文件上传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;
其他常用参数
–ae hash join
set spark.sql.adaptive.hashJoin.enabled=true;
set spark.sql.adaptiveHashJoinThreshold=52428800;
–输出文件合并 byBytes,该功能会生成两个stage，
–第一个stage shuffle的数据量来预估最后生成到hdfs上的文件数据量大小，
–并通过预估的文件数据量大小计算第二个stage的并行度，即最后生成的文件个数。
–该功能只能控制生成的文件大小尽量接近spark.merge.files.byBytes.fileBytes，且有一定的性能损耗，需根据实测情况选择使用。
– 最终文件数量：(totalBytes / fileBytes / compressionRatio).toInt + 1
set spark.merge.files.byBytes.enabled=true;
set spark.merge.files.byBytes.repartitionNumber=100;
–第一个stage的并行读
set spark.merge.files.byBytes.fileBytes=134217728;
– 预期的文件大小
set spark.merge.files.byBytes.compressionRatio=3;
– 压缩比，shuffle文件和最后生成的文件格式和压缩格式都不相同，因此通过该参数调节
–输出文件合并 该功能会在原来job的最后一个stage后面增加1个stage来控制最后生成的文件数量,
–对于动态分区，每个分区生成spark.merge.files.number个文件。
spark.merge.files.enabled=true
spark.merge.files.number=512
–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;
– 某partition数据量大于中位数的3倍，判定为倾斜
set spark.sql.adaptive.skewedPartitionMaxSplits=20;
– 限制某一partition最多拆分多少分，spark3已失效
set spark.sql.adaptive.skewedJoin.enabled=true;
– Normal Join Pattern的优化开关
set spark.sql.adaptive.skewedJoinWithAgg.enabled=true;
– JoinWithAgg Pattern的优化开关，非开源版
set spark.sql.adaptive.multipleSkewedJoin.enabled=true;
– MultipleJoin Pattern的优化开关，非开源版
set spark.shuffle.highlyCompressedMapStatusThreshold=20000;
– 分区数大于20000时 使用HighlyCompressedMapStatus统计每个partition数据量，会降低数据统计进度
–并发读文件
set spark.sql.concurrentFileScan.enabled=true;
–filter按比例读取文件
set spark.sql.files.tableSizeFactor={table_name}:{filter 比例};
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 特性生效，in 条件的个数
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