Migration Guide: Spark Core

• Upgrading from Core 3.5 to 4.0
• Upgrading from Core 3.4 to 3.5
• Upgrading from Core 3.3 to 3.4
• Upgrading from Core 3.2 to 3.3
• Upgrading from Core 3.1 to 3.2
• Upgrading from Core 3.0 to 3.1
• Upgrading from Core 2.4 to 3.0

Upgrading from Core 3.5 to 4.0

• Since Spark 4.0, Spark migrated all its internal reference of servlet API from javax to jakarta

• Since Spark 4.0, Spark will roll event logs to archive them incrementally. To restore the behavior before Spark 4.0, you can set spark.eventLog.rolling.enabled to false.

• Since Spark 4.0, Spark will compress event logs. To restore the behavior before Spark 4.0, you can set spark.eventLog.compress to false.

• Since Spark 4.0, Spark workers will clean up worker and stopped application directories periodically. To restore the behavior before Spark 4.0, you can set spark.worker.cleanup.enabled to false.

• Since Spark 4.0, spark.shuffle.service.db.backend is set to ROCKSDB by default which means Spark will use RocksDB store for shuffle service. To restore the behavior before Spark 4.0, you can set spark.shuffle.service.db.backend to LEVELDB.

• In Spark 4.0, support for Apache Mesos as a resource manager was removed.

• Since Spark 4.0, Spark uses ReadWriteOncePod instead of ReadWriteOnce access mode in persistence volume claims. To restore the legacy behavior, you can set spark.kubernetes.legacy.useReadWriteOnceAccessMode to true.

• Since Spark 4.0, Spark reports its executor pod status by checking all containers of that pod. To restore the legacy behavior, you can set spark.kubernetes.executor.checkAllContainers to false.

• Since Spark 4.0, Spark uses ~/.ivy2.5.2 as Ivy user directory by default to isolate the existing systems from Apache Ivy’s incompatibility. To restore the legacy behavior, you can set spark.jars.ivy to ~/.ivy2.

• Since Spark 4.0, Spark uses the external shuffle service for deleting shuffle blocks for deallocated executors when the shuffle is no longer needed. To restore the legacy behavior, you can set spark.shuffle.service.removeShuffle to false.

• Starting with Spark 4.0, the default logging format for spark-submit has changed from plain text to JSON lines to improve log analysis. If you prefer plain text logs, you have two options:
  • Set the Spark configuration spark.log.structuredLogging.enabled to false. For example, you can use JDK_JAVA_OPTIONS=-Dspark.log.structuredLogging.enabled=false.
  • Use a custom log4j configuration file, such as renaming the template file conf/log4j2.properties.pattern-layout-template to conf/log4j2.properties.

• Since Spark 4.0, the MDC (Mapped Diagnostic Context) key for Spark task names in Spark logs has been changed from mdc.taskName to task_name. To use the key mdc.taskName, you can set spark.log.legacyTaskNameMdc.enabled to true.

• Since Spark 4.0, Spark performs speculative executions less aggressively with spark.speculation.multiplier=3 and spark.speculation.quantile=0.9. To restore the legacy behavior, you can set spark.speculation.multiplier=1.5 and spark.speculation.quantile=0.75.

• Since Spark 4.0, spark.shuffle.unsafe.file.output.buffer is deprecated though still works. Use spark.shuffle.localDisk.file.output.buffer instead.

Upgrading from Core 3.4 to 3.5

• Since Spark 3.5, spark.yarn.executor.failuresValidityInterval is deprecated. Use spark.executor.failuresValidityInterval instead.

• Since Spark 3.5, spark.yarn.max.executor.failures is deprecated. Use spark.executor.maxNumFailures instead.

Upgrading from Core 3.3 to 3.4

• Since Spark 3.4, Spark driver will own PersistentVolumnClaims and try to reuse if they are not assigned to live executors. To restore the behavior before Spark 3.4, you can set spark.kubernetes.driver.ownPersistentVolumeClaim to false and spark.kubernetes.driver.reusePersistentVolumeClaim to false.

• Since Spark 3.4, Spark driver will track shuffle data when dynamic allocation is enabled without shuffle service. To restore the behavior before Spark 3.4, you can set spark.dynamicAllocation.shuffleTracking.enabled to false.

• Since Spark 3.4, Spark will try to decommission cached RDD and shuffle blocks if both spark.decommission.enabled and spark.storage.decommission.enabled are true. To restore the behavior before Spark 3.4, you can set both spark.storage.decommission.rddBlocks.enabled and spark.storage.decommission.shuffleBlocks.enabled to false.

• Since Spark 3.4, Spark will use RocksDB store if spark.history.store.hybridStore.enabled is true. To restore the behavior before Spark 3.4, you can set spark.history.store.hybridStore.diskBackend to LEVELDB.

Upgrading from Core 3.2 to 3.3

• Since Spark 3.3, Spark migrates its log4j dependency from 1.x to 2.x because log4j 1.x has reached end of life and is no longer supported by the community. Vulnerabilities reported after August 2015 against log4j 1.x were not checked and will not be fixed. Users should rewrite original log4j properties files using log4j2 syntax (XML, JSON, YAML, or properties format). Spark rewrites the conf/log4j.properties.template which is included in Spark distribution, to conf/log4j2.properties.template with log4j2 properties format.

Upgrading from Core 3.1 to 3.2

• Since Spark 3.2, spark.scheduler.allocation.file supports read remote file using hadoop filesystem which means if the path has no scheme Spark will respect hadoop configuration to read it. To restore the behavior before Spark 3.2, you can specify the local scheme for spark.scheduler.allocation.file e.g. file:///path/to/file.

• Since Spark 3.2, spark.hadoopRDD.ignoreEmptySplits is set to true by default which means Spark will not create empty partitions for empty input splits. To restore the behavior before Spark 3.2, you can set spark.hadoopRDD.ignoreEmptySplits to false.

• Since Spark 3.2, spark.eventLog.compression.codec is set to zstd by default which means Spark will not fallback to use spark.io.compression.codec anymore.

• Since Spark 3.2, spark.storage.replication.proactive is enabled by default which means Spark tries to replenish in case of the loss of cached RDD block replicas due to executor failures. To restore the behavior before Spark 3.2, you can set spark.storage.replication.proactive to false.

• In Spark 3.2, spark.launcher.childConectionTimeout is deprecated (typo) though still works. Use spark.launcher.childConnectionTimeout instead.

• In Spark 3.2, support for Apache Mesos as a resource manager is deprecated and will be removed in a future version.

• In Spark 3.2, Spark will delete K8s driver service resource when the application terminates by itself. To restore the behavior before Spark 3.2, you can set spark.kubernetes.driver.service.deleteOnTermination to false.

Upgrading from Core 3.0 to 3.1

• In Spark 3.0 and below, SparkContext can be created in executors. Since Spark 3.1, an exception will be thrown when creating SparkContext in executors. You can allow it by setting the configuration spark.executor.allowSparkContext when creating SparkContext in executors.

• In Spark 3.0 and below, Spark propagated the Hadoop classpath from yarn.application.classpath and mapreduce.application.classpath into the Spark application submitted to YARN when Spark distribution is with the built-in Hadoop. Since Spark 3.1, it does not propagate anymore when the Spark distribution is with the built-in Hadoop in order to prevent the failure from the different transitive dependencies picked up from the Hadoop cluster such as Guava and Jackson. To restore the behavior before Spark 3.1, you can set spark.yarn.populateHadoopClasspath to true.

Upgrading from Core 2.4 to 3.0

• The org.apache.spark.ExecutorPlugin interface and related configuration has been replaced with org.apache.spark.api.plugin.SparkPlugin, which adds new functionality. Plugins using the old interface must be modified to extend the new interfaces. Check the Monitoring guide for more details.

• Deprecated method TaskContext.isRunningLocally has been removed. Local execution was removed and it always has returned false.

• Deprecated method shuffleBytesWritten, shuffleWriteTime and shuffleRecordsWritten in ShuffleWriteMetrics have been removed. Instead, use bytesWritten, writeTime and recordsWritten respectively.

• Deprecated method AccumulableInfo.apply have been removed because creating AccumulableInfo is disallowed.

• Deprecated accumulator v1 APIs have been removed and please use v2 APIs instead.

• Event log file will be written as UTF-8 encoding, and Spark History Server will replay event log files as UTF-8 encoding. Previously Spark wrote the event log file as default charset of driver JVM process, so Spark History Server of Spark 2.x is needed to read the old event log files in case of incompatible encoding.

• A new protocol for fetching shuffle blocks is used. It’s recommended that external shuffle services be upgraded when running Spark 3.0 apps. You can still use old external shuffle services by setting the configuration spark.shuffle.useOldFetchProtocol to true. Otherwise, Spark may run into errors with messages like IllegalArgumentException: Unexpected message type: <number>.

• SPARK_WORKER_INSTANCES is deprecated in Standalone mode. It’s recommended to launch multiple executors in one worker and launch one worker per node instead of launching multiple workers per node and launching one executor per worker.