Spark 配置指南

目录 [−]

  1. Spark属性
  2. 动态加载Spark属性
  3. 查看Spark属性
  4. 可用的属性
    1. 应用属性
    2. 运行时环境Runtime Environment
    3. Shuffle Behavior
    4. Spark UI
    5. Compression and Serialization
    6. Execution Behavior
    7. Networking
    8. Scheduling
    9. Security
    10. Spark Streaming
    11. 集群管理器Cluster Managers
  5. 环境变量
  6. 配置日志
  7. 改变配置文件夹路径


  • Spark属性控制大部分的应用参数。 这些属性可以通过SparkConf对象, 或者Java系统属性.
  • 环境变量可以为每台机器配置,比如IP地址, 通过每个节点上的conf/spark-env.sh脚本.
  • 可同通过log4j.properties配置日志.


Spark属性控制应用的大部分设置, 可以为不同的应用分别设置. 这些属性在SparkConf对象上设置, SparkConf被传给SparkContext. SparkConf允许你配置一些通用的属性(比如master URL 和应用名), 也可以通过set() 方法设置键值对. 例如,我们可以这样初始化一个应用:

val conf = new SparkConf()
.set("spark.executor.memory", "1g")
val sc = new SparkContext(conf)


在一些情况下你可能想避免在SparkConf上硬编码. 举例来说, 如果你想在不同的master上或者不同的内存上运行同样的应用, Spark允许你简单创建一个空的conf:

val sc = new SparkContext(new SparkConf())


./bin/spark-submit --name "My app" --master local[4] --conf spark.shuffle.spill=false
--conf "spark.executor.extraJavaOptions=-XX:+PrintGCDetails -XX:+PrintGCTimeStamps" myApp.jar

Spark shell 和 spark-submit脚本支持两个动态加载配置的方法. 第一种是命令行参数, 如上面用到的 --master. spark-submit通过--conf可以接收任意的spark属性, 但会使用一些其它参数来启动Spark应用. 运行./bin/spark-submit --hp 会显示完整的参数列表.

bin/spark-submit会从conf/spark-defaults.conf读取缺省的配置参数, 每一行包括一个键和一个值, 由空格分隔. 比如下面的例子:

spark.master spark://
spark.executor.memory 512m
spark.eventLog.enabled true
spark.serializer org.apache.spark.serializer.KryoSerializer

命令行参数和文件中配置的属性都会传给应用,由SparkConf合并这些配置. SparkConf上设置的属性有最高优先级,然后是命令行传入给spark-submit或spark-shell的参数, 最后才是缺省文件中配置的属性.


应用的web UI (http://:4040)在"Environment"标签页列出了所有的Spark属性. 这是一个很有用的页面,可以帮助你检查你的属性是否设置正确。 注意只有显式在spark-defaults.conf 或 SparkConf 设置的属性才显示。 其它的配置属性将使用缺省值。




属性名缺省值意义 name of your application. This will appear in the UI and in log data.
spark.master(none)The cluster manager to connect to. See the list ofallowed master URL's.
spark.executor.memory512mAmount of memory to use per executor process, in the same format as JVM memory strings (e.g.512m,2g).
Class to use for serializing objects that will be sent over the network or need to be cached in serialized form. The default of Java serialization works with any Serializable Java object but is quite slow, so we recommendusingorg.apache.spark.serializer.KryoSerializer and configuring Kryo serialization when speed is necessary. Can be any subclass oforg.apache.spark.Serializer.
spark.kryo.registrator(none)If you use Kryo serialization, set this class to register your custom classes with Kryo. It should be set to a class that extendsKryoRegistrator. See thetuning guide for more details.
spark.local.dir/tmpDirectory to use for "scratch" space in Spark, including map output files and RDDs that get stored on disk. This should be on a fast, local disk in your system. It can also be a comma-separated list of multiple directories on different disks. NOTE: In Spark 1.0 and later this will be overriden by SPARK_LOCAL_DIRS (Standalone, Mesos) or LOCAL_DIRS (YARN) environment variables set by the cluster manager.
spark.logConffalseLogs the effective SparkConf as INFO when a SparkContext is started.


运行时环境Runtime Environment

spark.executor.extraJavaOptions(none)A string of extra JVM options to pass to executors. For instance, GC settings or other logging. Note that it is illegal to set Spark properties or heap size settings with this option. Spark properties should be set using a SparkConf object or the spark-defaults.conf file used with the spark-submit script. Heap size settings can be set with spark.executor.memory.
spark.executor.extraClassPath(none)Extra classpath entries to append to the classpath of executors. This exists primarily for backwards-compatibility with older versions of Spark. Users typically should not need to set this option.
spark.executor.extraLibraryPath(none)Set a special library path to use when launching executor JVM's.
spark.files.userClassPathFirstfalse(Experimental) Whether to give user-added jars precedence over Spark's own jars when loading classes in Executors. This feature can be used to mitigate conflicts between Spark's dependencies and user dependencies. It is currently an experimental feature.
spark.python.worker.memory512mAmount of memory to use per python worker process during aggregation, in the same format as JVM memory strings (e.g.512m,2g). If the memory used during aggregation goes above this amount, it will spill the data into disks.
spark.executorEnv.[EnvironmentVariableName](none)Add the environment variable specified byEnvironmentVariableName to the Executor process. The user can specify multiple of these and to set multiple environment variables.
spark.mesos.executor.homedriver sideSPARK_HOMESet the directory in which Spark is installed on the executors in Mesos. By default, the executors will simply use the driver's Spark home directory, which may not be visible to them. Note that this is only relevant if a Spark binary package is not specified throughspark.executor.uri.
spark.mesos.executor.memoryOverheadexecutor memory * 0.07, with minimum of 384This value is an additive forspark.executor.memory, specified in MiB, which is used to calculate the total Mesos task memory. A value of384 implies a 384MiB overhead. Additionally, there is a hard-coded 7% minimum overhead. The final overhead will be the larger of either spark.mesos.executor.memoryOverhead or 7% of spark.executor.memory.

Shuffle Behavior

spark.shuffle.consolidateFilesfalseIf set to "true", consolidates intermediate files created during a shuffle. Creating fewer files can improve filesystem performance for shuffles with large numbers of reduce tasks. It is recommended to set this to "true" when using ext4 or xfs filesystems. On ext3, this option might degrade performance on machines with many (>8) cores due to filesystem limitations.
spark.shuffle.spilltrueIf set to "true", limits the amount of memory used during reduces by spilling data out to disk. This spilling threshold is specified byspark.shuffle.memoryFraction.
spark.shuffle.spill.compresstrueWhether to compress data spilled during shuffles. Compression will
spark.shuffle.memoryFraction0.2Fraction of Java heap to use for aggregation and cogroups during shuffles, ifspark.shuffle.spill is true. At any given time, the collective size of all in-memory maps used for shuffles is bounded by this limit, beyond which the contents will begin to spill to disk. If spills are often, consider increasing this value at the expense
spark.shuffle.compresstrueWhether to compress map output files. Generally a good idea. Compression will
spark.shuffle.file.buffer.kb32Size of the in-memory buffer for each shuffle file output stream, in kilobytes. These buffers reduce the number of disk seeks and system calls made in creating intermediate shuffle files.
spark.reducer.maxMbInFlight48Maximum size (in megabytes) of map outputs to fetch simultaneously from each reduce task. Since each output requires us to create a buffer to receive it, this represents a fixed memory overhead per reduce task, so keep it small unless you have a large amount of memory.
spark.shuffle.managerHASHImplementation to use for shuffling data. A hash-based shuffle manager is the default, but starting in Spark 1.1 there is an experimental sort-based shuffle manager that is more memory-efficient in environments with small executors, such as YARN. To use that, change this value toSORT.
spark.shuffle.sort.bypassMergeThreshold200(Advanced) In the sort-based shuffle manager, avoid merge-sorting data if there is no map-side aggregation and there are at most this many reduce partitions.

Spark UI

spark.ui.port4040Port for your application's dashboard, which shows memory and workload data.
spark.ui.retainedStages1000How many stages the Spark UI remembers before garbage collecting.
spark.ui.killEnabledtrueAllows stages and corresponding jobs to be killed from the web ui.
spark.eventLog.enabledfalseWhether to log Spark events, useful for reconstructing the Web UI after the application has finished.
spark.eventLog.compressfalseWhether to compress logged events, ifspark.eventLog.enabled is true.
spark.eventLog.dirfile:///tmp/spark-eventsBase directory in which Spark events are logged, ifspark.eventLog.enabled is true. Within this base directory, Spark creates a sub-directory for each application, and logs the events specific to the application in this directory. Users may want to set this to a unified location like an HDFS directory so history files can be read by the history server.

Compression and Serialization

spark.broadcast.compresstrueWhether to compress broadcast variables before sending them. Generally a good idea.
spark.rdd.compressfalseWhether to compress serialized RDD partitions (e.g. forStorageLevel.MEMORY_ONLY_SER). Can save substantial space at the cost of some extra CPU time. codec used to compress internal data such as RDD partitions and shuffle outputs. By default, Spark provides three codecs:lz4,lzf, andsnappy. You can also use fully qualified class names to specify the codec,,, size (in bytes) used in Snappy compression, in the case when Snappy compression codec is used. Lowering this block size will also lower shuffle memory usage when Snappy is used. size (in bytes) used in LZ4 compression, in the case when LZ4 compression codec is used. Lowering this block size will also lower shuffle memory usage when LZ4 is used.
Serializer class to use for closures. Currently only the Java serializer is supported.
spark.serializer.objectStreamReset100When serializing using org.apache.spark.serializer.JavaSerializer, the serializer caches objects to prevent writing redundant data, however that stops garbage collection of those objects. By calling 'reset' you flush that info from the serializer, and allow old objects to be collected. To turn off this periodic reset set it to -1. By default it will reset the serializer every 100 objects.
spark.kryo.referenceTrackingtrueWhether to track references to the same object when serializing data with Kryo, which is necessary if your object graphs have loops and useful for efficiency if they contain multiple copies of the same object. Can be disabled to improve performance if you know this is not the case.
spark.kryo.registrationRequiredfalseWhether to require registration with Kryo. If set to 'true', Kryo will throw an exception if an unregistered class is serialized. If set to false (the default), Kryo will write unregistered class names along with each object. Writing class names can cause significant performance overhead, so enabling this option can enforce strictly that a user has not omitted classes from registration.
spark.kryoserializer.buffer.mb0.064Initial size of Kryo's serialization buffer, in megabytes. Note that there will be one bufferper core on each worker. This buffer will grow up tospark.kryoserializer.buffer.max.mb if needed.
spark.kryoserializer.buffer.max.mb64Maximum allowable size of Kryo serialization buffer, in megabytes. This must be larger than any object you attempt to serialize. Increase this if you get a "buffer limit exceeded" exception inside Kryo.

Execution Behavior

  • Local mode: number of cores on the local machine
  • Mesos fine grained mode: 8
  • Others: total number of cores on all executor nodes or 2, whichever is larger
Default number of tasks to use across the cluster for distributed shuffle operations (groupByKey,reduceByKey, etc) when not set by user.
Which broadcast implementation to use.
spark.broadcast.blockSize4096Size of each piece of a block in kilobytes forTorrentBroadcastFactory. Too large a value decreases parallelism during broadcast (makes it slower); however, if it is too small,BlockManager might take a performance hit.
spark.files.overwritefalseWhether to overwrite files added through SparkContext.addFile() when the target file exists and its contents do not match those of the source.
spark.files.fetchTimeoutfalseCommunication timeout to use when fetching files added through SparkContext.addFile() from the driver. of Java heap to use for Spark's memory cache. This should not be larger than the "old" generation of objects in the JVM, which by default is given 0.6 of the heap, but you can increase it if you configure your own old generation size. to use for unrolling blocks in memory. This is dynamically allocated by dropping existing blocks when there is not enough free storage space to unroll the new block in its entirety.
spark.tachyonStore.baseDirSystem.getProperty("")Directories of the Tachyon File System that store RDDs. The Tachyon file system's URL is set byspark.tachyonStore.url. It can also be a comma-separated list of multiple directories on Tachyon file system. of a block, in bytes, above which Spark memory maps when reading a block from disk. This prevents Spark from memory mapping very small blocks. In general, memory mapping has high overhead for blocks close to or below the page size of the operating system.
spark.tachyonStore.urltachyon://localhost:19998The URL of the underlying Tachyon file system in the TachyonStore.
spark.cleaner.ttl(infinite)Duration (seconds) of how long Spark will remember any metadata (stages generated, tasks generated, etc.). Periodic cleanups will ensure that metadata older than this duration will be forgotten. This is useful for running Spark for many hours / days (for example, running 24/7 in case of Spark Streaming applications). Note that any RDD that persists in memory for more than this duration will be cleared as well.
spark.hadoop.validateOutputSpecstrueIf set to true, validates the output specification (e.g. checking if the output directory already exists) used in saveAsHadoopFile and other variants. This can be disabled to silence exceptions due to pre-existing output directories. We recommend that users do not disable this except if trying to achieve compatibility with previous versions of Spark. Simply use Hadoop's FileSystem API to delete output directories by hand.
spark.hadoop.cloneConffalseIf set to true, clones a new HadoopConfiguration object for each task. This option should be enabled to work aroundConfiguration thread-safety issues (seeSPARK-2546 for more details). This is disabled by default in order to avoid unexpected performance regressions for jobs that are not affected by these issues.
spark.executor.heartbeatInterval10000Interval (milliseconds) between each executor's heartbeats to the driver. Heartbeats let the driver know that the executor is still alive and update it with metrics for in-progress tasks.


属性名缺省值意义 hostname)Hostname or IP address for the driver to listen on. This is used for communicating with the executors and the standalone Master.
spark.driver.port(random)Port for the driver to listen on. This is used for communicating with the executors and the standalone Master.
spark.fileserver.port(random)Port for the driver's HTTP file server to listen on.
spark.broadcast.port(random)Port for the driver's HTTP broadcast server to listen on. This is not relevant for torrent broadcast.
spark.replClassServer.port(random)Port for the driver's HTTP class server to listen on. This is only relevant for the Spark shell.
spark.blockManager.port(random)Port for all block managers to listen on. These exist on both the driver and the executors.
spark.executor.port(random)Port for the executor to listen on. This is used for communicating with the driver.
spark.port.maxRetries16Default maximum number of retries when binding to a port before giving up.
spark.akka.frameSize10Maximum message size to allow in "control plane" communication (for serialized tasks and task results), in MB. Increase this if your tasks need to send back large results to the driver (e.g. usingcollect() on a large dataset).
spark.akka.threads4Number of actor threads to use for communication. Can be useful to increase on large clusters when the driver has a lot of CPU cores.
spark.akka.timeout100Communication timeout between Spark nodes, in seconds.
spark.akka.heartbeat.pauses600This is set to a larger value to disable failure detector that comes inbuilt akka. It can be enabled again, if you plan to use this feature (Not recommended). Acceptable heart beat pause in seconds for akka. This can be used to control sensitivity to gc pauses. Tune this in combination of spark.akka.heartbeat.interval and spark.akka.failure-detector.threshold if you need to.
spark.akka.failure-detector.threshold300.0This is set to a larger value to disable failure detector that comes inbuilt akka. It can be enabled again, if you plan to use this feature (Not recommended). This maps to akka's akka.remote.transport-failure-detector.threshold. Tune this in combination of spark.akka.heartbeat.pauses and spark.akka.heartbeat.interval if you need to.
spark.akka.heartbeat.interval1000This is set to a larger value to disable failure detector that comes inbuilt akka. It can be enabled again, if you plan to use this feature (Not recommended). A larger interval value in seconds reduces network overhead and a smaller value ( ~ 1 s) might be more informative for akka's failure detector. Tune this in combination of spark.akka.heartbeat.pauses and spark.akka.failure-detector.threshold if you need to. Only positive use case for using failure detector can be, a sensistive failure detector can help evict rogue executors really quick. However this 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 heart beats between nodes leading to flooding the network with those.


spark.task.cpus1Number of cores to allocate for each task.
spark.task.maxFailures4Number of individual task failures before giving up on the job. Should be greater than or equal to 1. Number of allowed retries = this value - 1.
spark.scheduler.modeFIFOThescheduling mode between jobs submitted to the same SparkContext. Can be set toFAIR to use fair sharing instead of queueing jobs one after another. Useful for multi-user services.
spark.cores.max(not set)When running on astandalone deploy cluster or aMesos cluster in "coarse-grained" sharing mode, the maximum amount of CPU cores to request for the application from across the cluster (not from each machine). If not set, the default will bespark.deploy.defaultCores on Spark's standalone cluster manager, or infinite (all available cores) on Mesos.
spark.mesos.coarsefalseIf set to "true", runs over Mesos clusters in"coarse-grained" sharing mode, where Spark acquires one long-lived Mesos task on each machine instead of one Mesos task per Spark task. This gives lower-latency scheduling for short queries, but leaves resources in use for the whole duration of the Spark job.
spark.speculationfalseIf set to "true", performs speculative execution of tasks. This means if one or more tasks are running slowly in a stage, they will be re-launched.
spark.speculation.interval100How often Spark will check for tasks to speculate, in milliseconds.
spark.speculation.quantile0.75Percentage of tasks which must be complete before speculation is enabled for a particular stage.
spark.speculation.multiplier1.5How many times slower a task is than the median to be considered for speculation.
spark.locality.wait3000Number of milliseconds to wait to launch a data-local task before giving up and launching it on a less-local node. The same wait will be used to step through multiple locality levels (process-local, node-local, rack-local and then any). It is also possible to customize the waiting time for each level by settingspark.locality.wait.node, etc. You should increase this setting if your tasks are long and see poor locality, but the default usually works well.
spark.locality.wait.processspark.locality.waitCustomize the locality wait for process locality. This affects tasks that attempt to access cached data in a particular executor process.
spark.locality.wait.nodespark.locality.waitCustomize the locality wait for node locality. For example, you can set this to 0 to skip node locality and search immediately for rack locality (if your cluster has rack information).
spark.locality.wait.rackspark.locality.waitCustomize the locality wait for rack locality.
spark.scheduler.revive.interval1000The interval length for the scheduler to revive the worker resource offers to run tasks (in milliseconds).
spark.scheduler.minRegisteredResourcesRatio0The minimum ratio of registered resources (registered resources / total expected resources) (resources are executors in yarn mode, CPU cores in standalone mode) to wait for before scheduling begins. Specified as a double between 0 and 1. Regardless of whether the minimum ratio of resources has been reached, the maximum amount of time it will wait before scheduling begins is controlled by configspark.scheduler.maxRegisteredResourcesWaitingTime.
spark.scheduler.maxRegisteredResourcesWaitingTime30000Maximum amount of time to wait for resources to register before scheduling begins (in milliseconds).
spark.localExecution.enabledfalseEnables Spark to run certain jobs, such as first() or take() on the driver, without sending tasks to the cluster. This can make certain jobs execute very quickly, but may require shipping a whole partition of data to the driver.


spark.authenticatefalseWhether Spark authenticates its internal connections. Seespark.authenticate.secret if not running on YARN.
spark.authenticate.secretNoneSet the secret key used for Spark to authenticate between components. This needs to be set if not running on YARN and authentication is enabled.
spark.core.connection.auth.wait.timeout30Number of seconds for the connection to wait for authentication to occur before timing out and giving up.
spark.core.connection.ack.wait.timeout60Number of seconds for the connection to wait for ack to occur before timing out and giving up. To avoid unwilling timeout caused by long pause like GC, you can set larger value.
spark.ui.filtersNoneComma separated list of filter class names to apply to the Spark web UI. The filter should be a standardjavax servlet Filter. Parameters to each filter can also be specified by setting a java system property of:
spark.<class name of filter>.params='param1=value1,param2=value2'
For example:
spark.acls.enablefalseWhether Spark acls should are enabled. If enabled, this checks to see if the user has access permissions to view or modify the job. Note this requires the user to be known, so if the user comes across as null no checks are done. Filters can be used with the UI to authenticate and set the user.
spark.ui.view.aclsEmptyComma separated list of users that have view access to the Spark web ui. By default only the user that started the Spark job has view access.
spark.modify.aclsEmptyComma separated list of users that have modify access to the Spark job. By default only the user that started the Spark job has access to modify it (kill it for example).
spark.admin.aclsEmptyComma separated list of users/administrators that have view and modify access to all Spark jobs. This can be used if you run on a shared cluster and have a set of administrators or devs who help debug when things work.

Spark Streaming

spark.streaming.blockInterval200Interval (milliseconds) at which data received by Spark Streaming receivers is coalesced into blocks of data before storing them in Spark.
spark.streaming.receiver.maxRateinfiniteMaximum rate (per second) at which each receiver will push data into blocks. Effectively, each stream will consume at most this number of records per second. Setting this configuration to 0 or a negative number will put no limit on the rate.
spark.streaming.unpersisttrueForce RDDs generated and persisted by Spark Streaming to be automatically unpersisted from Spark's memory. The raw input data received by Spark Streaming is also automatically cleared. Setting this to false will allow the raw data and persisted RDDs to be accessible outside the streaming application as they will not be cleared automatically. But it comes at the cost of higher memory usage in Spark.
spark.executor.logs.rolling.strategy(none)Set the strategy of rolling of executor logs. By default it is disabled. It can be set to "time" (time-based rolling) or "size" (size-based rolling). For "time", usespark.executor.logs.rolling.time.interval to set the rolling interval. For "size", usespark.executor.logs.rolling.size.maxBytes to set the maximum file size for rolling.
spark.executor.logs.rolling.time.intervaldailySet the time interval by which the executor logs will be rolled over. Rolling is disabled by default. Valid values are daily, hourly, minutely or any interval in seconds. Seespark.executor.logs.rolling.maxRetainedFiles for automatic cleaning of old logs.
spark.executor.logs.rolling.size.maxBytes(none)Set the max size of the file by which the executor logs will be rolled over. Rolling is disabled by default. Value is set in terms of bytes. Seespark.executor.logs.rolling.maxRetainedFiles for automatic cleaning of old logs.
spark.executor.logs.rolling.maxRetainedFiles(none)Sets the number of latest rolling log files that are going to be retained by the system. Older log files will be deleted. Disabled by default.

集群管理器Cluster Managers

每种集群管理都有自己额外的配置参数. 可以在下面的页面中找到每种管理器相应的配置:


有些Spark设置可以通过环境变量来设置, 从Spark文件夹中的conf/spark-env.sh脚本中读取(Windows操作系统中用conf/spark-env.cmd). 在Standalone 和 Mesos 模式下, 这个文件可以给机器特定的信息如hostnames. It is also sourced when running local Spark applications or submission scripts.

注意conf/spark-env.sh缺省情况下并不存在。 然而,你可以从conf/复制一份. 请确保这个复制脚本可执行.


JAVA_HOMELocation where Java is installed (if it's not on your default PATH).
PYSPARK_PYTHONPython binary executable to use for PySpark.
SPARK_LOCAL_IPIP address of the machine to bind to.
SPARK_PUBLIC_DNSHostname your Spark program will advertise to other machines.

除此之外,还有一些参数用来设置Spark standalone 集群脚本, 比如每台机器使用的核数和最大内存。

既然spark-env.sh是一个shell脚本, 可以通过编程的方式设置。 举例来说,你可能查找一个特定网络的IP来设置SPARK_LOCAL_IP.


Spark使用log4j记录日志. 你可以在conf增加一个log4j.properties文件. 其文件夹下有一个log4j的模版.


为了使用一个其它的配置文件夹而不是“SPARK_HOME/conf”, 你可以设置SPARK_CONF_DIR. Spark会使用你指定的文件夹的文件(spark-defaults.conf,,, 等等).

翻译自 Spark Configuration