spark配置详解

对付看把 到这里格式变化太大了,懒得调整了,这是大概spark1.5版本时候的一些参数默认值,现在2.x会有变化

这些皆可在 spark-default.conf配置,或者部分可在 sparkconf().set设置

应用程序属性

|--------------------------------------------------------------------------------------------
| 属性名称                                    | 默认值 | 含义
|--------------------------------------------------------------------------------------------
| spark.app.name                         | (none) | 你的应用程序的名字。这将在UI和日志数据中出现
|--------------------------------------------------------------------------------------------
| spark.driver.cores                      | 1      | driver程序运行需要的cpu内核数
|--------------------------------------------------------------------------------------------
| spark.driver.maxResultSize       | 1g    | 每个Spark action(如collect)所有分区的序列化结果的总大小限制
|                                                    |        | 。设置的值应该不小于1m,0代表没有限制。如果总大小超过这个限制,程序将会终止。
|                                                    |        | 大的限制值可能导致driver出现内存溢出错误(依赖于spark.driver.
|                                                    |        | memory和JVM中对象的内存消耗)。
|--------------------------------------------------------------------------------------------
| spark.driver.memory                  | 512m   | driver进程使用的内存数
|--------------------------------------------------------------------------------------------
| spark.executor.memory             | 512m   | 每个executor进程使用的内存数。和JVM内存串拥有相同的格式(如512m,
|                                                    |             | 2g)
|--------------------------------------------------------------------------------------------
| spark.extraListeners       | (none) | 注册监听器,需要实现SparkListener
|--------------------------------------------------------------------------------------------
| spark.local.dir            | /tmp   | Spark中暂存空间的使用目录。在Spark1.0以及更高的版本中,这个属性被S
|                            |        | PARK_LOCAL_DIRS(Standalone, Mesos)和LOCAL
|                            |        | _DIRS(YARN)环境变量覆盖。
|--------------------------------------------------------------------------------------------
| spark.logConf              | false  | 当SparkContext启动时,将有效的SparkConf记录为INFO。
|--------------------------------------------------------------------------------------------
| spark.master               | (none) | 集群管理器连接的地方
|--------------------------------------------------------------------------------------------
 
运行环境
|----------------------------------------------------------------------------------------------------------------
| 属性名称                                     | 默认值   | 含义
|----------------------------------------------------------------------------------------------------------------
| spark.driver.extraClassPath                  | (none)   | 附加到driver的classpath的额外的classpath实体。
|----------------------------------------------------------------------------------------------------------------
| spark.driver.extraJavaOptions                | (none)   | 传递给driver的JVM选项字符串。例如GC设置或者其它日志设置。注意,在这个
|                                              |          | 选项中设置Spark属性或者堆大小是不合法的。Spark属性需要用--drive
|                                              |          | r-class-path设置。
|----------------------------------------------------------------------------------------------------------------
| spark.driver.extraLibraryPath                | (none)   | 指定启动driver的JVM时用到的库路径
|----------------------------------------------------------------------------------------------------------------
| spark.driver.userClassPathFirst              | false    | (实验性)当在driver中加载类时,是否用户添加的jar比Spark自己的ja
|                                              |          | r优先级高。这个属性可以降低Spark依赖和用户依赖的冲突。它现在还是一个实验性
|                                              |          | 的特征。
|----------------------------------------------------------------------------------------------------------------
| spark.executor.extraClassPath                | (none)   | 附加到executors的classpath的额外的classpath实体。这个
|                                              |          | 设置存在的主要目的是Spark与旧版本的向后兼容问题。用户一般不用设置这个选项
|----------------------------------------------------------------------------------------------------------------
| spark.executor.extraJavaOptions              | (none)   | 传递给executors的JVM选项字符串。例如GC设置或者其它日志设置。注意,
|                                              |          | 在这个选项中设置Spark属性或者堆大小是不合法的。Spark属性需要用Spar
|                                              |          | kConf对象或者spark-submit脚本用到的spark-defaults
|                                              |          | .conf文件设置。堆内存可以通过spark.executor.memory设置
|                                              |          |
|----------------------------------------------------------------------------------------------------------------
| spark.executor.extraLibraryPath              | (none)   | 指定启动executor的JVM时用到的库路径
|----------------------------------------------------------------------------------------------------------------
| spark.executor.logs.rolling.maxRetainedFiles | (none)   | 设置被系统保留的最近滚动日志文件的数量。更老的日志文件将被删除。默认没有开启。
|----------------------------------------------------------------------------------------------------------------
| spark.executor.logs.rolling.size.maxBytes    | (none)   | executor日志的最大滚动大小。默认情况下没有开启。值设置为字节spark1.4 已经废弃 使用spark.executor.logs.rolling.maxSize
|----------------------------------------------------------------------------------------------------------------
| spark.executor.logs.rolling.strategy         | (none)   | 设置executor日志的滚动(rolling)策略。默认情况下没有开启。可以配
|                                              |          | 置为time和size。对于time,用spark.executor.logs.
|                                              |          | rolling.time.interval设置滚动间隔;对于size,用spar
|                                              |          | k.executor.logs.rolling.size.maxBytes设置最
|                                              |          | 大的滚动大小
|----------------------------------------------------------------------------------------------------------------
| spark.executor.logs.rolling.time.interval    | daily    | executor日志滚动的时间间隔。默认情况下没有开启。合法的值是daily,
|                                              |          | hourly, minutely以及任意的秒。
|----------------------------------------------------------------------------------------------------------------
| spark.files.userClassPathFirst               | false    | (实验性)当在Executors中加载类时,是否用户添加的jar比Spark自己
|                                              |          | 的jar优先级高。这个属性可以降低Spark依赖和用户依赖的冲突。它现在还是一个
|                                              |          | 实验性的特征。
|----------------------------------------------------------------------------------------------------------------
| spark.python.worker.memory                   | 512m     | 在聚合期间,每个python worker进程使用的内存数。在聚合期间,如果内存
|                                              |          | 超过了这个限制,它将会将数据塞进磁盘中
|----------------------------------------------------------------------------------------------------------------
| spark.python.profile                         | false    | 在Python worker中开启profiling。通过sc.show_pro
|                                              |          | files()展示分析结果。或者在driver退出前展示分析结果。可以通过sc.
|                                              |          | dump_profiles(path)将结果dump到磁盘中。如果一些分析结果已
|                                              |          | 经手动展示,那么在driver退出前,它们再不会自动展示
|----------------------------------------------------------------------------------------------------------------
| spark.python.profile.dump                    | (none)   | driver退出前保存分析结果的dump文件的目录。每个RDD都会分别dump一
|                                              |          | 个文件。可以通过ptats.Stats()加载这些文件。如果指定了这个属性,分析
|                                              |          | 结果不会自动展示
|----------------------------------------------------------------------------------------------------------------
| spark.python.worker.reuse                    | true     | 是否重用python worker。如果是,它将使用固定数量的Python wo
|                                              |          | rkers,而不需要为每个任务fork()一个Python进程。如果有一个非常大
|                                              |          | 的广播,这个设置将非常有用。因为,广播不需要为每个任务从JVM到Python w
|                                              |          | orker传递一次
|----------------------------------------------------------------------------------------------------------------
| spark.executorEnv.[EnvironmentVariableName]  | (none)   | 通过EnvironmentVariableName添加指定的环境变量到execu
|                                              |          | tor进程。用户可以指定多个EnvironmentVariableName,设置
|                                              |          | 多个环境变量
|----------------------------------------------------------------------------------------------------------------
| spark.mesos.executor.home                    | driver   | side SPARK_HOME     设置安装在Mesos的executor上的Sp
|                                              |          | ark的目录。默认情况下,executors将使用driver的Spark本地(
|                                              |          | home)目录,这个目录对它们不可见。注意,如果没有通过 spark.execu
|                                              |          | tor.uri指定Spark的二进制包,这个设置才起作用
|----------------------------------------------------------------------------------------------------------------
| spark.mesos.executor.memoryOverhead          | executor | memory * 0.07, 最小384m   这个值是spark.executo
|                                              |          | r.memory的补充。它用来计算mesos任务的总内存。另外,有一个7%的硬编
|                                              |          | 码设置。最后的值将选择spark.mesos.executor.memoryOv
|                                              |          | erhead或者spark.executor.memory的7%二者之间的大者
|----------------------------------------------------------------------------------------------------------------
 
 
Shuffle行为
|---------------------------------------------------------------------------------------------------------
| 属性名称                                | 默认值 | 含义
|---------------------------------------------------------------------------------------------------------
| spark.reducer.maxMbInFlight             | 48     | 从递归任务中同时获取的map输出数据的最大大小(mb)。因为每一个输出都需要我们
|                                         |        | 创建一个缓存用来接收,这个设置代表每个任务固定的内存上限,所以除非你有更大的内存
|                                         |        | ,将其设置小一点
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.blockTransferService      | netty  | 实现用来在executor直接传递shuffle和缓存块。有两种可用的实现:ne
|                                         |        | tty和nio。基于netty的块传递在具有相同的效率情况下更简单
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.compress                  | true   | 是否压缩map操作的输出文件。一般情况下,这是一个好的选择。
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.consolidateFiles          | false  | 如果设置为”true”,在shuffle期间,合并的中间文件将会被创建。创建更少
|                                         |        | 的文件可以提供文件系统的shuffle的效率。这些shuffle都伴随着大量递归
|                                         |        | 任务。当用ext4和dfs文件系统时,推荐设置为”true”。在ext3中,因为
|                                         |        | 文件系统的限制,这个选项可能机器(大于8核)降低效率
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.file.buffer.kb            | 32     | 每个shuffle文件输出流内存内缓存的大小,单位是kb。这个缓存减少了创建只中
|                                         |        | 间shuffle文件中磁盘搜索和系统访问的数量
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.io.maxRetries             | 3      | Netty only,自动重试次数
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.io.numConnectionsPerPeer  | 1      | Netty only
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.io.preferDirectBufs       | true   | Netty only
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.io.retryWait              | 5      | Netty only
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.manager                   | sort   | 它的实现用于shuffle数据。有两种可用的实现:sort和hash。基于sor
|                                         |        | t的shuffle有更高的内存使用率
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.memoryFraction            | 0.2    | 如果spark.shuffle.spill为true,shuffle中聚合和合并
|                                         |        | 组操作使用的java堆内存占总内存的比重。在任何时候,shuffles使用的所有
|                                         |        | 内存内maps的集合大小都受这个限制的约束。超过这个限制,spilling数据将
|                                         |        | 会保存到磁盘上。如果spilling太过频繁,考虑增大这个值
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.sort.bypassMergeThreshold | 200    | (Advanced) In the sort-based shuffle man
|                                         |        | ager, avoid merge-sorting data if there
|                                         |        | is no map-side aggregation and there are
|                                         |        |  at most this many reduce partitions
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.spill                     | true   | 如果设置为”true”,通过将多出的数据写入磁盘来限制内存数。通过spark.s
|                                         |        | huffle.memoryFraction来指定spilling的阈值
|---------------------------------------------------------------------------------------------------------
| spark.shuffle.spill.compress            | true   | 在shuffle时,是否将spilling的数据压缩。压缩算法通过spark.i
|                                         |        | o.compression.codec指定。
|---------------------------------------------------------------------------------------------------------
 
 
Spark UI
|-----------------------------------------------------------------------------------------------------------
| 属性名称                | 默认值                   | 含义
|-----------------------------------------------------------------------------------------------------------
| spark.eventLog.compress | false                    | 是否压缩事件日志。需要spark.eventLog.enabled为true
|-----------------------------------------------------------------------------------------------------------
| spark.eventLog.dir      | file:///tmp/spark-events | Spark事件日志记录的基本目录。在这个基本目录下,Spark为每个应用程序创建
|                         |                          | 一个子目录。各个应用程序记录日志到直到的目录。用户可能想设置这为统一的地点,像H
|                         |                          | DFS一样,所以历史文件可以通过历史服务器读取
|-----------------------------------------------------------------------------------------------------------
| spark.eventLog.enabled  | false                    | 是否记录Spark的事件日志。这在应用程序完成后,重新构造web UI是有用的
|-----------------------------------------------------------------------------------------------------------
| spark.ui.killEnabled    | true                     | 运行在web UI中杀死stage和相应的job
|-----------------------------------------------------------------------------------------------------------
| spark.ui.port           | 4040                     | 你的应用程序dashboard的端口。显示内存和工作量数据
|-----------------------------------------------------------------------------------------------------------
| spark.ui.retainedJobs   | 1000                     | 在垃圾回收之前,Spark UI和状态API记住的job数
|-----------------------------------------------------------------------------------------------------------
| spark.ui.retainedStages | 1000                     | 在垃圾回收之前,Spark UI和状态API记住的stage数
|-----------------------------------------------------------------------------------------------------------
 
 
压缩和序列化
|--------------------------------------------------------------------------------------------------------------------------------------------
| 属性名称                               | 默认值                                     | 含义
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.broadcast.compress               | true                                       | 在发送广播变量之前是否压缩它
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.closure.serializer               | org.apache.spark.serializer.JavaSerializer | 闭包用到的序列化类。目前只支持java序列化器
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.io.compression.codec             | snappy                                     | 压缩诸如RDD分区、广播变量、shuffle输出等内部数据的编码解码器。默认情况
|                                        |                                            | 下,Spark提供了三种选择:lz4、lzf和snappy,你也可以用完整的类名
|                                        |                                            | 来制定。
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.io.compression.lz4.block.size    | 32768                                      | LZ4压缩中用到的块大小。降低这个块的大小也会降低shuffle内存使用率
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.io.compression.snappy.block.size | 32768                                      | Snappy压缩中用到的块大小。降低这个块的大小也会降低shuffle内存使用率
|                                        |                                            |
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.kryo.classesToRegister           | (none)                                     | 如果你用Kryo序列化,给定的用逗号分隔的自定义类名列表表示要注册的类
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.kryo.referenceTracking           | true                                       | 当用Kryo序列化时,跟踪是否引用同一对象。如果你的对象图有环,这是必须的设置。
|                                        |                                            | 如果他们包含相同对象的多个副本,这个设置对效率是有用的。如果你知道不在这两个场景
|                                        |                                            | ,那么可以禁用它以提高效率
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.kryo.registrationRequired        | false                                      | 是否需要注册为Kyro可用。如果设置为true,然后如果一个没有注册的类序列化,
|                                        |                                            | Kyro会抛出异常。如果设置为false,Kryo将会同时写每个对象和其非注册类
|                                        |                                            | 名。写类名可能造成显著地性能瓶颈。
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.kryo.registrator                 | (none)                                     | 如果你用Kryo序列化,设置这个类去注册你的自定义类。如果你需要用自定义的方式注
|                                        |                                            | 册你的类,那么这个属性是有用的。否则spark.kryo.classesToRe
|                                        |                                            | gister会更简单。它应该设置一个继承自KryoRegistrator的类
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.kryoserializer.buffer.max.mb     | 64                                         | Kryo序列化缓存允许的最大值。这个值必须大于你尝试序列化的对象
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.kryoserializer.buffer.mb         | 0.064                                      | Kyro序列化缓存的大小。这样worker上的每个核都有一个缓存。如果有需要,缓
|                                        |                                            | 存会涨到spark.kryoserializer.buffer.max.mb设置
|                                        |                                            | 的值那么大。
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.rdd.compress                     | true                                       | 是否压缩序列化的RDD分区。在花费一些额外的CPU时间的同时节省大量的空间
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.serializer                       | org.apache.spark.serializer.JavaSerializer | 序列化对象使用的类。默认的Java序列化类可以序列化任何可序列化的java对象但
|                                        |                                            | 是它很慢。所有我们建议用org.apache.spark.serializer.
|                                        |                                            | KryoSerializer
|--------------------------------------------------------------------------------------------------------------------------------------------
| spark.serializer.objectStreamReset     | 100                                        | 当用org.apache.spark.serializer.JavaSerial
|                                        |                                            | izer序列化时,序列化器通过缓存对象防止写多余的数据,然而这会造成这些对象的垃
|                                        |                                            | 圾回收停止。通过请求’reset’,你从序列化器中flush这些信息并允许收集老
|                                        |                                            | 的数据。为了关闭这个周期性的reset,你可以将值设为-1。默认情况下,每一百个
|                                        |                                            | 对象reset一次
|--------------------------------------------------------------------------------------------------------------------------------------------
 
 
运行时行为
|------------------------------------------------------------------------------------------------------------------------------------------------
| 属性名称                         | 默认值                                               | 含义
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.broadcast.blockSize        | 4096                                                 | TorrentBroadcastFactory传输的块大小,太大值会降低并发,太
|                                  |                                                      | 小的值会出现性能瓶颈
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.broadcast.factory          | org.apache.spark.broadcast.TorrentBroadcastFactory   | broadcast实现类
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.cleaner.ttl                | (infinite)                                           | spark记录任何元数据(stages生成、task生成等)的持续时间。定期清理
|                                  |                                                      | 可以确保将超期的元数据丢弃,这在运行长时间任务是很有用的,如运行7*24的spa
|                                  |                                                      | rkstreaming任务。RDD持久化在内存中的超期数据也会被清理
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.default.parallelism        | 本地模式:机器核数;Mesos:8;其他:max(executor的core,2) | 如果用户不设置,系统使用集群中运行shuffle操作的默认任务数(groupBy
|                                  |                                                      | Key、 reduceByKey等)
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.executor.heartbeatInterval | 10000                                                | executor 向 the driver 汇报心跳的时间间隔,单位毫秒
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.files.fetchTimeout         | 60                                                   | driver 程序获取通过SparkContext.addFile()添加的文件
|                                  |                                                      | 时的超时时间,单位秒
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.files.useFetchCache        | true                                                 | 获取文件时是否使用本地缓存
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.files.overwrite            | false                                                | 调用SparkContext.addFile()时候是否覆盖文件
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.hadoop.cloneConf           | false                                                | 每个task是否克隆一份hadoop的配置文件
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.hadoop.validateOutputSpecs | true                                                 | 是否校验输出
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.storage.memoryFraction     | 0.6                                                  | Spark内存缓存的堆大小占用总内存比例,该值不能大于老年代内存大小,默认值为0
|                                  |                                                      | .6,但是,如果你手动设置老年代大小,你可以增加该值
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.storage.memoryMapThreshold | 2097152                                              | 内存块大小
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.storage.unrollFraction     | 0.2                                                  | Fraction of spark.storage.memoryFraction
|                                  |                                                      |  to use for unrolling blocks in memory.
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.tachyonStore.baseDir       | System.getProperty(“java.io.tmpdir”)                 | Tachyon File System临时目录
|------------------------------------------------------------------------------------------------------------------------------------------------
| spark.tachyonStore.url           | tachyon://localhost:19998                            | Tachyon File System URL
|------------------------------------------------------------------------------------------------------------------------------------------------
 
 
网络
|---------------------------------------------------------------------------------------------------------
| 属性名称                              | 默认值   | 含义
|---------------------------------------------------------------------------------------------------------
| spark.driver.host                     | (local   | hostname)  driver监听的主机名或者IP地址。这用于和execut
|                                       |          | ors以及独立的master通信
|---------------------------------------------------------------------------------------------------------
| spark.driver.port                     | (random) | driver监听的接口。这用于和executors以及独立的master通信
|---------------------------------------------------------------------------------------------------------
| spark.fileserver.port                 | (random) | driver的文件服务器监听的端口
|---------------------------------------------------------------------------------------------------------
| spark.broadcast.port                  | (random) | driver的HTTP广播服务器监听的端口
|---------------------------------------------------------------------------------------------------------
| spark.replClassServer.port            | (random) | driver的HTTP类服务器监听的端口
|---------------------------------------------------------------------------------------------------------
| spark.blockManager.port               | (random) | 块管理器监听的端口。这些同时存在于driver和executors
|---------------------------------------------------------------------------------------------------------
| spark.executor.port                   | (random) | executor监听的端口。用于与driver通信
|---------------------------------------------------------------------------------------------------------
| spark.port.maxRetries                 | 16       | 当绑定到一个端口,在放弃前重试的最大次数
|---------------------------------------------------------------------------------------------------------
| spark.akka.frameSize                  | 10       | 在”control plane”通信中允许的最大消息大小。如果你的任务需要发送大
|                                       |          | 的结果到driver中,调大这个值
|---------------------------------------------------------------------------------------------------------
| spark.akka.threads                    | 4        | 通信的actor线程数。当driver有很多CPU核时,调大它是有用的
|---------------------------------------------------------------------------------------------------------
| spark.akka.timeout                    | 100      | Spark节点之间的通信超时。单位是秒
|---------------------------------------------------------------------------------------------------------
| spark.akka.heartbeat.pauses           | 6000     | This is set to a larger value to disable
|                                       |          |  failure detector that comes inbuilt akk
|                                       |          | a. It can be enabled again, if you plan
|                                       |          | to use this feature (Not recommended). A
|                                       |          | cceptable heart beat pause in seconds fo
|                                       |          | r akka. This can be used to control sens
|                                       |          | itivity to gc pauses. Tune this in combi
|                                       |          | nation of spark.akka.heartbeat.interval
|                                       |          | and spark.akka.failure-detector.threshol
|                                       |          | d if you need to.
|---------------------------------------------------------------------------------------------------------
| spark.akka.failure-detector.threshold | 300.0    | This is set to a larger value to disable
|                                       |          |  failure detector that comes inbuilt akk
|                                       |          | a. It can be enabled again, if you plan
|                                       |          | to use this feature (Not recommended). T
|                                       |          | his maps to akka’s akka.remote.transport
|                                       |          | -failure-detector.threshold. Tune this i
|                                       |          | n combination of spark.akka.heartbeat.pa
|                                       |          | uses and spark.akka.heartbeat.interval i
|                                       |          | f you need to.
|---------------------------------------------------------------------------------------------------------
| spark.akka.heartbeat.interval         | 1000     | This is set to a larger value to disable
|                                       |          |  failure detector that comes inbuilt akk
|                                       |          | a. It can be enabled again, if you plan
|                                       |          | to use this feature (Not recommended). A
|                                       |          |  larger interval value in seconds reduce
|                                       |          | s network overhead and a smaller value (
|                                       |          |  ~ 1 s) might be more informative for ak
|                                       |          | ka’s failure detector. Tune this in comb
|                                       |          | ination of spark.akka.heartbeat.pauses a
|                                       |          | nd spark.akka.failure-detector.threshold
|                                       |          |  if you need to. Only positive use case
|                                       |          | for using failure detector can be, a sen
|                                       |          | sistive failure detector can help evict
|                                       |          | rogue executors really quick. However th
|                                       |          | is is usually not the case as gc pauses
|                                       |          | and network lags are expected in a real
|                                       |          | Spark cluster. Apart from that enabling
|                                       |          | this leads to a lot of exchanges of hear
|                                       |          | t beats between nodes leading to floodin
|                                       |          | g the network with those.
|---------------------------------------------------------------------------------------------------------
 
 
调度相关属性
|--------------------------------------------------------------------------------------------------------------
| 属性名称                        | 默认值              | 含义
|--------------------------------------------------------------------------------------------------------------
| spark.task.cpus                 | 1                   | 为每个任务分配的内核数
|--------------------------------------------------------------------------------------------------------------
| spark.task.maxFailures          | 4                   | Task的最大重试次数
|--------------------------------------------------------------------------------------------------------------
| spark.scheduler.mode            | FIFO                | Spark的任务调度模式,还有一种Fair模式
|--------------------------------------------------------------------------------------------------------------
| spark.cores.max                 | 无                  | 当应用程序运行在Standalone集群或者粗粒度共享模式Mesos集群时,应用
|                                 |                     | 程序向集群请求的最大CPU内核总数(不是指每台机器,而是整个集群)。如果不设置,
|                                 |                     | 对于Standalone集群将使用spark.deploy.defaultCor
|                                 |                     | es中数值,而Mesos将使用集群中可用的内核
|--------------------------------------------------------------------------------------------------------------
| spark.mesos.coarse              | False               | 如果设置为true,在Mesos集群中运行时使用粗粒度共享模式
|--------------------------------------------------------------------------------------------------------------
| spark.speculation               | False               | 以下几个参数是关于Spark推测执行机制的相关参数。此参数设定是否使用推测执行机
|                                 |                     | 制,如果设置为true则spark使用推测执行机制,对于Stage中拖后腿的Ta
|                                 |                     | sk在其他节点中重新启动,并将最先完成的Task的计算结果最为最终结果
|--------------------------------------------------------------------------------------------------------------
| spark.speculation.interval      | 100                 | Spark多长时间进行检查task运行状态用以推测,以毫秒为单位
|--------------------------------------------------------------------------------------------------------------
| spark.speculation.quantile      | 无                  | 推测启动前,Stage必须要完成总Task的百分比
|--------------------------------------------------------------------------------------------------------------
| spark.speculation.multiplier    | 1.5                 | 比已完成Task的运行速度中位数慢多少倍才启用推测
|--------------------------------------------------------------------------------------------------------------
| spark.locality.wait             | 3000                | 以下几个参数是关于Spark数据本地性的。本参数是以毫秒为单位启动本地数据tas
|                                 |                     | k的等待时间,如果超出就启动下一本地优先级别的task。该设置同样可以应用到各优
|                                 |                     | 先级别的本地性之间(本地进程 -> 本地节点 -> 本地机架 -> 任意节点 )
|                                 |                     | ,当然,也可以通过spark.locality.wait.node等参数设置不同
|                                 |                     | 优先级别的本地性
|--------------------------------------------------------------------------------------------------------------
| spark.locality.wait.process     | spark.locality.wait | 本地进程级别的本地等待时间
|--------------------------------------------------------------------------------------------------------------
| spark.locality.wait.node        | spark.locality.wait | 本地节点级别的本地等待时间
|--------------------------------------------------------------------------------------------------------------
| spark.locality.wait.rack        | spark.locality.wait | 本地机架级别的本地等待时间
|--------------------------------------------------------------------------------------------------------------
| spark.scheduler.revive.interval | 1000                | 复活重新获取资源的Task的最长时间间隔(毫秒),发生在Task因为本地资源不足
|                                 |                     | 而将资源分配给其他Task运行后进入等待时间,如果这个等待时间内重新获取足够的资
|                                 |                     | 源就继续计算
|--------------------------------------------------------------------------------------------------------------
 
Dynamic Allocation
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| 属性名称                                                 | 默认值                               | 含义
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| spark.dynamicAllocation.enabled                          | false                                | 是否开启动态资源搜集
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| spark.dynamicAllocation.executorIdleTimeout              | 600                                  |
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| spark.dynamicAllocation.initialExecutors                 | spark.dynamicAllocation.minExecutors |
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| spark.dynamicAllocation.maxExecutors                     | Integer.MAX_VALUE                    |
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| spark.dynamicAllocation.minExecutors                     | 0                                    |
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| spark.dynamicAllocation.schedulerBacklogTimeout          | 5                                    |
|--------------------------------------------------------------------------------------------------------------------------------------------------------
| spark.dynamicAllocation.sustainedSchedulerBacklogTimeout | schedulerBacklogTimeout              |
|--------------------------------------------------------------------------------------------------------------------------------------------------------
 
安全
|---------------------------------------------------------------------------------------------------------
| 属性名称                                | 默认值 | 含义
|---------------------------------------------------------------------------------------------------------
| spark.authenticate                      | false  | 是否Spark验证其内部连接。如果不是运行在YARN上,请看spark.auth
|                                         |        | enticate.secret
|---------------------------------------------------------------------------------------------------------
| spark.authenticate.secret               | None   | 设置Spark两个组件之间的密匙验证。如果不是运行在YARN上,但是需要验证,这
|                                         |        | 个选项必须设置
|---------------------------------------------------------------------------------------------------------
| spark.core.connection.auth.wait.timeout | 30     | 连接时等待验证的实际。单位为秒
|---------------------------------------------------------------------------------------------------------
| spark.core.connection.ack.wait.timeout  | 60     | 连接等待回答的时间。单位为秒。为了避免不希望的超时,你可以设置更大的值
|---------------------------------------------------------------------------------------------------------
| spark.ui.filters                        | None   | 应用到Spark web UI的用于过滤类名的逗号分隔的列表。过滤器必须是标准的
|                                         |        | javax servlet Filter。通过设置java系统属性也可以指定每个
|                                         |        | 过滤器的参数。spark.<class filter="" name="" of="">.para
|                                         |        | ms='param1=value1,param2=value2'。例如-Dspa
|                                         |        | rk.ui.filters=com.test.filter1、-Dspark.c
|                                         |        | om.test.filter1.params='param1=foo,param
|                                         |        | 2=testing'
|---------------------------------------------------------------------------------------------------------
| spark.acls.enable                       | false  | 是否开启Spark acls。如果开启了,它检查用户是否有权限去查看或修改job
|                                         |        | 。UI利用使用过滤器验证和设置用户
|---------------------------------------------------------------------------------------------------------
| spark.ui.view.acls                      | empty  | 逗号分隔的用户列表,列表中的用户有查看Spark web UI的权限。默认情况下
|                                         |        | ,只有启动Spark job的用户有查看权限
|---------------------------------------------------------------------------------------------------------
| spark.modify.acls                       | empty  | 逗号分隔的用户列表,列表中的用户有修改Spark job的权限。默认情况下,只有
|                                         |        | 启动Spark job的用户有修改权限
|---------------------------------------------------------------------------------------------------------
| spark.admin.acls                        | empty  | 逗号分隔的用户或者管理员列表,列表中的用户或管理员有查看和修改所有Spark j
|                                         |        | ob的权限。如果你运行在一个共享集群,有一组管理员或开发者帮助debug,这个选
|                                         |        | 项有用
|---------------------------------------------------------------------------------------------------------
 
 
加密
|----------------------------------------------------------------------------------------------
| 属性名称                     | 默认值 | 含义
|----------------------------------------------------------------------------------------------
| spark.ssl.enabled            | false  | 是否开启ssl
|----------------------------------------------------------------------------------------------
| spark.ssl.enabledAlgorithms  | Empty  | JVM支持的加密算法列表,逗号分隔
|----------------------------------------------------------------------------------------------
| spark.ssl.keyPassword        | None   |
|----------------------------------------------------------------------------------------------
| spark.ssl.keyStore           | None   |
|----------------------------------------------------------------------------------------------
| spark.ssl.keyStorePassword   | None   |
|----------------------------------------------------------------------------------------------
| spark.ssl.protocol           | None   |
|----------------------------------------------------------------------------------------------
| spark.ssl.trustStore         | None   |
|----------------------------------------------------------------------------------------------
| spark.ssl.trustStorePassword | None   |
|----------------------------------------------------------------------------------------------
 
 
Spark Streaming
|------------------------------------------------------------------------------------------------------------------
| 属性名称                                       | 默认值   | 含义
|------------------------------------------------------------------------------------------------------------------
| spark.streaming.blockInterval                  | 200      | 在这个时间间隔(ms)内,通过Spark Streaming receivers
|                                                |          | 接收的数据在保存到Spark之前,chunk为数据块。推荐的最小值为50ms
|------------------------------------------------------------------------------------------------------------------
| spark.streaming.receiver.maxRate               | infinite | 每秒钟每个receiver将接收的数据的最大记录数。有效的情况下,每个流将消耗至
|                                                |          | 少这个数目的记录。设置这个配置为0或者-1将会不作限制
|------------------------------------------------------------------------------------------------------------------
| spark.streaming.receiver.writeAheadLogs.enable | false    | Enable write ahead logs for receivers. A
|                                                |          | ll the input data received through recei
|                                                |          | vers will be saved to write ahead logs t
|                                                |          | hat will allow it to be recovered after
|                                                |          | driver failures
|------------------------------------------------------------------------------------------------------------------
| spark.streaming.unpersist                      | true     | 强制通过Spark Streaming生成并持久化的RDD自动从Spark内存中
|                                                |          | 非持久化。通过Spark Streaming接收的原始输入数据也将清除。设置这个
|                                                |          | 属性为false允许流应用程序访问原始数据和持久化RDD,因为它们没有被自动清除
|                                                |          | 。但是它会造成更高的内存花费
|------------------------------------------------------------------------------------------------------------------
 
 
集群管理
Spark On YARN
|-------------------------------------------------------------------------------------------------------------------------------------------------
| 属性名称                                          | 默认值                               | 含义
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.am.memory                              | 512m                                 | client 模式时,am的内存大小;cluster模式时,使用spark.dr
|                                                   |                                      | iver.memory变量
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.driver.cores                                | 1                                    | claster模式时,driver使用的cpu核数,这时候driver运行在am
|                                                   |                                      | 中,其实也就是am和核数;client模式时,使用spark.yarn.am.c
|                                                   |                                      | ores变量
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.am.cores                               | 1                                    | client 模式时,am的cpu核数
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.am.waitTime                            | 100000                               | 启动时等待时间
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.submit.file.replication                | 3                                    | 应用程序上传到HDFS的文件的副本数
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.preserve.staging.files                 | False                                | 若为true,在job结束后,将stage相关的文件保留而不是删除
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.scheduler.heartbeat.interval-ms        | 5000                                 | Spark AppMaster发送心跳信息给YARN RM的时间间隔
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.max.executor.failures                  | 2倍于executor数,最小值3              | 导致应用程序宣告失败的最大executor失败次数
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.applicationMaster.waitTries            | 10                                   | RM等待Spark AppMaster启动重试次数,也就是SparkContex
|                                                   |                                      | t初始化次数。超过这个数值,启动失败
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.historyServer.address                  | Spark                                | history server的地址(不要加 http://)。这个地址会在Spa
|                                                   |                                      | rk应用程序完成后提交给YARN RM,然后RM将信息从RM UI写到histo
|                                                   |                                      | ry server UI上。
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.dist.archives                          | (none)                               |
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.dist.files                             | (none)                               |
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.executor.instances                          | 2                                    | executor实例个数
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.executor.memoryOverhead                | executorMemory                       | * 0.07, with minimum of 384    executor的堆内
|                                                   |                                      | 存大小设置
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.driver.memoryOverhead                  | driverMemory                         | * 0.07, with minimum of 384    driver的堆内存大
|                                                   |                                      | 小设置
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.am.memoryOverhead                      | AM                                   | memory * 0.07, with minimum of 384     am的堆
|                                                   |                                      | 内存大小设置,在client模式时设置
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.queue                                  | default                              | 使用yarn的队列
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.jar                                    | (none)                               |
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.access.namenodes                       | (none)                               |
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.appMasterEnv.[EnvironmentVariableName] | (none)                               | 设置am的环境变量
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.containerLauncherMaxThreads            | 25                                   | am启动executor的最大线程数
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.am.extraJavaOptions                    | (none)                               |
|-------------------------------------------------------------------------------------------------------------------------------------------------
| spark.yarn.maxAppAttempts                         | yarn.resourcemanager.am.max-attempts | in YARN    am重试次数
|-------------------------------------------------------------------------------------------------------------------------------------------------
 
Spark History Server的属性
|----------------------------------------------------------------------------------------------------------------------------------------------
| 属性名称                           | 默认                                             | 含义
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.provider             | org.apache.spark.deploy.history.FsHistoryProvide | 应用历史后端实现的类名。 目前只有一个实现, 由Spark提供, 它查看存储在文
|                                    |                                                  | 件系统里面的应用日志
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.fs.logDirectory      | file:/tmp/spark-events                           |
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.updateInterval       | 10                                               | 以秒为单位,多长时间Spark history server显示的信息进行更新。
|                                    |                                                  | 每次更新都会检查持久层事件日志的任何变化。
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.retainedApplications | 50                                               | 在Spark history server上显示的最大应用程序数量,如果超过这个
|                                    |                                                  | 值,旧的应用程序信息将被删除。
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.ui.port              | 18080                                            | 官方版本中,Spark history server的默认访问端口
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.kerberos.enabled     | false                                            | 是否使用kerberos方式登录访问history server,对于持久层位于
|                                    |                                                  | 安全集群的HDFS上是有用的。如果设置为true,就要配置下面的两个属性。
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.kerberos.principal   | 空                                               | 用于Spark history server的kerberos主体名称
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.kerberos.keytab      | 空                                               | 用于Spark history server的kerberos keytab文件
|                                    |                                                  | 位置
|----------------------------------------------------------------------------------------------------------------------------------------------
| spark.history.ui.acls.enable       | false                                            | 授权用户查看应用程序信息的时候是否检查acl。如果启用,只有应用程序所有者和sp
|                                    |                                                  | ark.ui.view.acls指定的用户可以查看应用程序信息;如果禁用,不做任
|                                    |                                                  | 何检查。
|----------------------------------------------------------------------------------------------------------------------------------------------

posted @ 2018-03-08 10:35  该名称已被用户使用  阅读(711)  评论(0编辑  收藏  举报