Release Notes - Flink 1.11

These release notes discuss important aspects, such as configuration, behavior, or dependencies, that changed between Flink 1.10 and Flink 1.11. Please read these notes carefully if you are planning to upgrade your Flink version to 1.11.

Clusters & Deployment

Support for Application Mode (FLIP-85)

The user can now submit applications and choose to execute their main() method on the cluster rather than the client. This allows for more light-weight application submission. For more details, see the Application Mode documentation.

Web Submission behaves the same as detached mode.

With FLINK-16657 the web submission logic changes and it exposes the same behavior as submitting a job through the CLI in detached mode. This implies that, for instance, jobs based on the DataSet API that were using sinks like print(), count() or collect() will now throw an exception while before the output was simply never printed. See also comments on related PR.

Flink project does not provide any updated “flink-shaded-hadoop-*” jars. Users need to provide Hadoop dependencies through the HADOOP_CLASSPATH environment variable (recommended) or via lib/ folder. Also, the include-hadoop Maven profile has been removed.

There is no need to download manually jar files for flink-csv and flink-json formats as they are now bundled in the lib folder.

Flink no longer supports the legacy scheduler. Hence, setting jobmanager.scheduler: legacy will no longer work and fail with an IllegalArgumentException. The only valid option for jobmanager.scheduler is the default value ng.

The user code class loader is being reused by the TaskExecutor as long as there is at least a single slot allocated for the respective job. This changes Flink’s recovery behaviour slightly so that it will not reload static fields. The benefit is that this change drastically reduces pressure on the JVM’s metaspace.

For Standalone Setups, the file with the worker nodes is no longer called slaves but workers. Previous setups that use the start-cluster.sh and stop-cluster.sh scripts need to rename that file.

The examples of Dockerfiles and docker image build.sh scripts have been removed from the Flink Github repository. The examples will no longer be maintained by community in the Flink Github repository, including the examples of integration with Bluemix. Therefore, the following modules have been deleted from the Flink Github repository:

  • flink-contrib/docker-flink
  • flink-container/docker
  • flink-container/kubernetes

Check the updated user documentation for Flink Docker integration instead. It now describes in detail how to use and customize the Flink official docker image: configuration options, logging, plugins, adding more dependencies and installing software. The documentation also includes examples for Session and Job cluster deployments with:

Memory Management

New JobManager Memory Model

Overview

With FLIP-116, a new memory model has been introduced for the JobManager. New configuration options have been introduced to control the memory consumption of the JobManager process. This affects all types of deployments: standalone, YARN, Mesos, and the new active Kubernetes integration.

Please, check the user documentation for more details.

If you try to reuse your previous Flink configuration without any adjustments, the new memory model can result in differently computed memory parameters for the JVM and, thus, performance changes or even failures. In order to start the JobManager process, you have to specify at least one of the following options jobmanager.memory.flink.size, jobmanager.memory.process.size or jobmanager.memory.heap.size. See also the migration guide for more information.

Deprecation and breaking changes

The following options are deprecated:

  • jobmanager.heap.size
  • jobmanager.heap.mb

If these deprecated options are still used, they will be interpreted as one of the following new options in order to maintain backwards compatibility:

The following options have been removed and have no effect anymore:

  • containerized.heap-cutoff-ratio
  • containerized.heap-cutoff-min

There is no container cut-off anymore.

JVM arguments

The direct and metaspace memory of the JobManager’s JVM process are now limited by configurable values:

See also JVM Parameters.

Attention These new limits can produce the respective OutOfMemoryError exceptions if they are not configured properly or there is a respective memory leak. See also the troubleshooting guide.

The mesos.resourcemanager.tasks.mem option, deprecated in 1.10 in favour of taskmanager.memory.process.size, has been completely removed and will have no effect anymore in 1.11+.

Table API & SQL

The default table planner has been changed to blink.

Due to various issues with packages org.apache.flink.table.api.scala/java all classes from those packages were relocated. Moreover the scala expressions were moved to org.apache.flink.table.api as announced in Flink 1.9.

If you used one of:

  • org.apache.flink.table.api.java.StreamTableEnvironment
  • org.apache.flink.table.api.scala.StreamTableEnvironment
  • org.apache.flink.table.api.java.BatchTableEnvironment
  • org.apache.flink.table.api.scala.BatchTableEnvironment

And you do not convert to/from DataStream, switch to:

  • org.apache.flink.table.api.TableEnvironment

If you do convert to/from DataStream/DataSet, change your imports to one of:

  • org.apache.flink.table.api.bridge.java.StreamTableEnvironment
  • org.apache.flink.table.api.bridge.scala.StreamTableEnvironment
  • org.apache.flink.table.api.bridge.java.BatchTableEnvironment
  • org.apache.flink.table.api.bridge.scala.BatchTableEnvironment

For the Scala expressions use the import:

  • org.apache.flink.table.api._ instead of org.apache.flink.table.api.bridge.scala._

Additionally, if you use Scala’s implicit conversions to/from DataStream/DataSet, import org.apache.flink.table.api.bridge.scala._ instead of org.apache.flink.table.api.scala._

The existing StreamTableSink implementations should remove emitDataStream method.

The existing BatchTableSink implementations should rename emitDataSet to consumeDataSet and return DataSink.

In previous versions, TableEnvironment.execute() and StreamExecutionEnvironment.execute() can both trigger table and DataStream programs. Since Flink 1.11.0, table programs can only be triggered by TableEnvironment.execute(). Once table program is converted into DataStream program (through toAppendStream() or toRetractStream() method), it can only be triggered by StreamExecutionEnvironment.execute().

In previous versions, BatchTableEnvironment.execute() and ExecutionEnvironment.execute() can both trigger table and DataSet programs for legacy batch planner. Since Flink 1.11.0, batch table programs can only be triggered by BatchEnvironment.execute(). Once table program is converted into DataSet program (through toDataSet() method), it can only be triggered by ExecutionEnvironment.execute().

An additional change flag called RowKind was added to the Row type. This changed the serialization format and will trigger a state migration.

Configuration

The logging properties files log4j-yarn-session.properties and logback-yarn.xml have been renamed to log4j-session.properties and logback-session.xml. Moreover, yarn-session.sh and kubernetes-session.sh use these logging properties files.

State

The StateTtlConfig#cleanupInBackground has been removed, because the method was deprecated and the background TTL was enabled by default in 1.10.

The TTL compaction filter in RocksDB has been enabled in 1.10 by default and it is now always enabled in 1.11+. Because of that the following option and methods have been removed in 1.11:

  • state.backend.rocksdb.ttl.compaction.filter.enabled
  • StateTtlConfig#cleanupInRocksdbCompactFilter()
  • RocksDBStateBackend#isTtlCompactionFilterEnabled
  • RocksDBStateBackend#enableTtlCompactionFilter
  • RocksDBStateBackend#disableTtlCompactionFilter
  • (state_backend.py) is_ttl_compaction_filter_enabled
  • (state_backend.py) enable_ttl_compaction_filter
  • (state_backend.py) disable_ttl_compaction_filter

Starting from Flink 1.11 the StateBackendFactory#createFromConfig interface now takes ReadableConfig instead of Configuration. A Configuration class is still a valid argument to that method, as it implements the ReadableConfig interface. Implementors of custom StateBackend should adjust their implementations.

The deprecated OptionsFactory and ConfigurableOptionsFactory classes have been removed. Please use RocksDBOptionsFactory and ConfigurableRocksDBOptionsFactory instead. Please also recompile your application codes if any class extends DefaultConfigurableOptionsFactory.

Since Flink-1.11 the option setTotalOrderSeek will be enabled by default for RocksDB’s ReadOptions. This is in order to prevent user from miss using optimizeForPointLookup. For backward compatibility we support customizing ReadOptions through RocksDBOptionsFactory. Please set setTotalOrderSeek back to false if any performance regression observed (it shouldn’t happen according to our testing).

The default value of state.backend.fs.memory-threshold has been increased from 1K to 20K to prevent too many small files created on remote FS for small states. Jobs with large parallelism on source or stateful operators may have “JM OOM” or “RPC message exceeding maximum frame size” problem with this change. If you encounter such issues please manually set the configuration back to 1K.

DataTypes can be configured with some parameters, e.g., precision. However in previous releases, the precision provided by users was not taking any effect and default value for the precision was being used. To avoid confusion since Flink 1.11 exceptions will be thrown if the value is not supported to make it more visible to users. Changes include:

  • the precision for TimeType can only be 0
  • the length for VarBinaryType/VarCharType can only be 0x7fffffff
  • the precision/scale for DecimalType can only be 38/18
  • the precision for TimestampType/LocalZonedTimestampType can only be 3
  • the resolution for DayTimeIntervalType can only be SECOND and the fractionalPrecision can only be 3
  • the resolution for YearMonthIntervalType can only be MONTH and the yearPrecision can only be 2
  • the CharType/BinaryType/ZonedTimestampType is not supported

Monitoring

All MetricReporters that come with Flink have been converted to plugins. They should no longer be placed into /lib directory (doing so may result in dependency conflicts!), but /plugins/<some_directory> instead.

The DataDog metrics reporter now reports counts as the number of events over the reporting interval, instead of the total count. This aligns the count semantics with the DataDog documentation.

Flink now uses Log4j2 by default. Users who wish to revert back to Log4j1 can find instructions to do so in the logging documentation.

Requesting an unavailable log or stdout file from the JobManager’s HTTP server returns status code 404 now. In previous releases, the HTTP server would return a file with (file unavailable) as its content.

Note that the metric lastCheckpointAlignmentBuffered has been removed, because the upstream task will not send any data after emitting a checkpoint barrier until the alignment has been completed on the downstream side. The web UI still displays this value but it is always 0 now.

Connectors

The Kafka 0.8 and 0.9 connectors are no longer under active development and were removed.

The Elasticsearch 2 connector is no longer under active development and was removed. Prior version of these connectors will continue to work with Flink.

Deprecated KafkaPartitioner was removed. Please see the release notes of Flink 1.3.0 how to migrate from that interface.

By default, if there is an official filesystem plugin for a given schema, it will not be allowed to use fallback filesystem factories (like HADOOP libraries on the classpath) to load it. Added fs.allowed-fallback-filesystems configuration option to override this behaviour.

org.apache.flink.core.fs.FileSystem#getKind method has been formally deprecated, as it was not used by Flink.

Runtime

Failures in synchronous part of checkpointing (like an exceptions thrown by an operator) will fail its Task (and job) immediately, regardless of the configuration parameters. Since Flink 1.5 such failures could be ignored by setting setTolerableCheckpointFailureNumber(...) or its deprecated setFailTaskOnCheckpointError(...) predecessor. Now both options will only affect asynchronous failures.

Checkpoint timeouts will now be treated as normal checkpoint failures and checked against value configured by CheckpointConfig#setTolerableCheckpointFailureNumber(...).

Miscellaneous Interface Changes

DataStream API no longer provides StreamTask#getCheckpointLock method, which was deprecated in Flink 1.10. Users should use MailboxExecutor to run actions that require synchronization with the task’s thread (e.g. collecting output produced by an external thread). MailboxExecutor#yield or MailboxExecutor#tryYield methods can be used for actions that should give control to other actions temporarily (equivalent of StreamTask#getCheckpointLock().wait()), if the current operator is blocked. MailboxExecutor can be accessed by using YieldingOperatorFactory. Example usage can be found in the AsyncWaitOperator.

Note, SourceFunction.SourceContext.getCheckpointLock is still available for custom implementations of SourceFunction interface.

Starting from Flink 1.11.0, the flink-streaming-java module does not have a dependency on flink-clients anymore. If your project was depending on this transitive dependency you now have to add flink-clients as an explicit dependency.

AsyncWaitOperator will be allowed to be chained by default with all operators, except of tasks with SourceFunction. This mostly revert limitation introduced as a bug fix for FLINK-13063.

The argument type of methods ShuffleEnvironment#createInputGates and #createResultPartitionWriters are adjusted from Collection to List for satisfying the order guarantee requirement in unaligned checkpoint. It will break the compatibility if users already implemented a custom ShuffleService based on ShuffleServiceFactory interface.

The boolean isOuterSnapshotCompatible(TypeSerializer) on the CompositeTypeSerializerSnapshot class has been deprecated, in favor of a new OuterSchemaCompatibility resolveOuterSchemaCompatibility(TypeSerializer) method. Please implement that instead. Compared to the old method, the new method allows composite serializers to signal state schema migration based on outer schema and configuration.

The long-deprecated TimestampExtractor was removed along with API methods in the DataStream API. Please use the new TimestampAssigner and WatermarkStrategies for working with timestamps and watermarks in the DataStream API.

The ListCheckpointed interface has been deprecated because it uses Java Serialization for checkpointing state which is problematic for savepoint compatibility. Use the CheckpointedFunction interface instead, which gives more control over state serialization.

We removed deprecated state access methods RuntimeContext#getFoldingState(), OperatorStateStore#getSerializableListState() and OperatorStateStore#getOperatorState(). This means that some code that was compiled against Flink 1.10 will not work with a Flink 1.11 cluster. An example of this is our Kafka connector which internally used OperatorStateStore.getSerializableListState.