Streaming File Sink

This connector provides a Sink that writes partitioned files to filesystems supported by the Flink FileSystem abstraction.

The streaming file sink writes incoming data into buckets. Given that the incoming streams can be unbounded, data in each bucket are organized into part files of finite size. The bucketing behaviour is fully configurable with a default time-based bucketing where we start writing a new bucket every hour. This means that each resulting bucket will contain files with records received during 1 hour intervals from the stream.

Data within the bucket directories are split into part files. Each bucket will contain at least one part file for each subtask of the sink that has received data for that bucket. Additional part files will be created according to the configurable rolling policy. The default policy rolls part files based on size, a timeout that specifies the maximum duration for which a file can be open, and a maximum inactivity timeout after which the file is closed.

IMPORTANT: Checkpointing needs to be enabled when using the StreamingFileSink. Part files can only be finalized on successful checkpoints. If checkpointing is disabled, part files will forever stay in the `in-progress` or the `pending` state, and cannot be safely read by downstream systems.

Streaming File Sink - 图1

File Formats

The StreamingFileSink supports both row-wise and bulk encoding formats, such as Apache Parquet. These two variants come with their respective builders that can be created with the following static methods:

  • Row-encoded sink: StreamingFileSink.forRowFormat(basePath, rowEncoder)
  • Bulk-encoded sink: StreamingFileSink.forBulkFormat(basePath, bulkWriterFactory)

When creating either a row or a bulk encoded sink we have to specify the base path where the buckets will be stored and the encoding logic for our data.

Please check out the JavaDoc for StreamingFileSink for all the configuration options and more documentation about the implementation of the different data formats.

Row-encoded Formats

Row-encoded formats need to specify an Encoder that is used for serializing individual rows to the OutputStream of the in-progress part files.

In addition to the bucket assigner, the RowFormatBuilder allows the user to specify:

  • Custom RollingPolicy : Rolling policy to override the DefaultRollingPolicy
  • bucketCheckInterval (default = 1 min) : Millisecond interval for checking time based rolling policies

Basic usage for writing String elements thus looks like this:

  1. import org.apache.flink.api.common.serialization.SimpleStringEncoder;
  2. import org.apache.flink.core.fs.Path;
  3. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink;
  4. import org.apache.flink.streaming.api.functions.sink.filesystem.rollingpolicies.DefaultRollingPolicy;
  6. DataStream<String> input = ...;
  8. final StreamingFileSink<String> sink = StreamingFileSink
  9.     .forRowFormat(new Path(outputPath), new SimpleStringEncoder<String>("UTF-8"))
  10.     .withRollingPolicy(
  11.         DefaultRollingPolicy.builder()
  12.             .withRolloverInterval(TimeUnit.MINUTES.toMillis(15))
  13.             .withInactivityInterval(TimeUnit.MINUTES.toMillis(5))
  14.             .withMaxPartSize(1024 * 1024 * 1024)
  15.             .build())
  16.     .build();
  18. input.addSink(sink);
  1. import org.apache.flink.api.common.serialization.SimpleStringEncoder
  2. import org.apache.flink.core.fs.Path
  3. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink
  4. import org.apache.flink.streaming.api.functions.sink.filesystem.rollingpolicies.DefaultRollingPolicy
  6. val input: DataStream[String] = ...
  8. val sink: StreamingFileSink[String] = StreamingFileSink
  9.     .forRowFormat(new Path(outputPath), new SimpleStringEncoder[String]("UTF-8"))
  10.     .withRollingPolicy(
  11.         DefaultRollingPolicy.builder()
  12.             .withRolloverInterval(TimeUnit.MINUTES.toMillis(15))
  13.             .withInactivityInterval(TimeUnit.MINUTES.toMillis(5))
  14.             .withMaxPartSize(1024 * 1024 * 1024)
  15.             .build())
  16.     .build()
  18. input.addSink(sink)

This example creates a simple sink that assigns records to the default one hour time buckets. It also specifies a rolling policy that rolls the in-progress part file on any of the following 3 conditions:

  • It contains at least 15 minutes worth of data
  • It hasn’t received new records for the last 5 minutes
  • The file size has reached 1 GB (after writing the last record)

Bulk-encoded Formats

Bulk-encoded sinks are created similarly to the row-encoded ones, but instead of specifying an Encoder, we have to specify a BulkWriter.Factory. The BulkWriter logic defines how new elements are added and flushed, and how a batch of records is finalized for further encoding purposes.

Flink comes with four built-in BulkWriter factories:

IMPORTANT: Bulk Formats can only have `OnCheckpointRollingPolicy`, which rolls (ONLY) on every checkpoint.

Parquet format

Flink contains built in convenience methods for creating Parquet writer factories for Avro data. These methods and their associated documentation can be found in the ParquetAvroWriters class.

For writing to other Parquet compatible data formats, users need to create the ParquetWriterFactory with a custom implementation of the ParquetBuilder interface.

To use the Parquet bulk encoder in your application you need to add the following dependency:

  1. <dependency>
  2.   <groupId>org.apache.flink</groupId>
  3.   <artifactId>flink-parquet_2.11</artifactId>
  4.   <version>1.12.0</version>
  5. </dependency>

A StreamingFileSink that writes Avro data to Parquet format can be created like this:

  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink;
  2. import org.apache.flink.formats.parquet.avro.ParquetAvroWriters;
  3. import org.apache.avro.Schema;
  6. Schema schema = ...;
  7. DataStream<GenericRecord> stream = ...;
  9. final StreamingFileSink<GenericRecord> sink = StreamingFileSink
  10.     .forBulkFormat(outputBasePath, ParquetAvroWriters.forGenericRecord(schema))
  11.     .build();
  13. input.addSink(sink);
  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink
  2. import org.apache.flink.formats.parquet.avro.ParquetAvroWriters
  3. import org.apache.avro.Schema
  5. val schema: Schema = ...
  6. val input: DataStream[GenericRecord] = ...
  8. val sink: StreamingFileSink[GenericRecord] = StreamingFileSink
  9.     .forBulkFormat(outputBasePath, ParquetAvroWriters.forGenericRecord(schema))
  10.     .build()
  12. input.addSink(sink)

Similarly, a StreamingFileSink that writes Protobuf data to Parquet format can be created like this:

  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink;
  2. import org.apache.flink.formats.parquet.protobuf.ParquetProtoWriters;
  4. // ProtoRecord is a generated protobuf Message class.
  5. DataStream<ProtoRecord> stream = ...;
  7. final StreamingFileSink<ProtoRecord> sink = StreamingFileSink
  8.     .forBulkFormat(outputBasePath, ParquetProtoWriters.forType(ProtoRecord.class))
  9.     .build();
  11. input.addSink(sink);
  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink
  2. import org.apache.flink.formats.parquet.protobuf.ParquetProtoWriters
  4. // ProtoRecord is a generated protobuf Message class.
  5. val input: DataStream[ProtoRecord] = ...
  7. val sink: StreamingFileSink[ProtoRecord] = StreamingFileSink
  8.     .forBulkFormat(outputBasePath, ParquetProtoWriters.forType(classOf[ProtoRecord]))
  9.     .build()
  11. input.addSink(sink)

Avro format

Flink also provides built-in support for writing data into Avro files. A list of convenience methods to create Avro writer factories and their associated documentation can be found in the AvroWriters class.

To use the Avro writers in your application you need to add the following dependency:

  1. <dependency>
  2.   <groupId>org.apache.flink</groupId>
  3.   <artifactId>flink-avro</artifactId>
  4.   <version>1.12.0</version>
  5. </dependency>

A StreamingFileSink that writes data to Avro files can be created like this:

  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink;
  2. import org.apache.flink.formats.avro.AvroWriters;
  3. import org.apache.avro.Schema;
  6. Schema schema = ...;
  7. DataStream<GenericRecord> stream = ...;
  9. final StreamingFileSink<GenericRecord> sink = StreamingFileSink
  10.     .forBulkFormat(outputBasePath, AvroWriters.forGenericRecord(schema))
  11.     .build();
  13. input.addSink(sink);
  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink
  2. import org.apache.flink.formats.avro.AvroWriters
  3. import org.apache.avro.Schema
  5. val schema: Schema = ...
  6. val input: DataStream[GenericRecord] = ...
  8. val sink: StreamingFileSink[GenericRecord] = StreamingFileSink
  9.     .forBulkFormat(outputBasePath, AvroWriters.forGenericRecord(schema))
  10.     .build()
  12. input.addSink(sink)

For creating customized Avro writers, e.g. enabling compression, users need to create the AvroWriterFactory with a custom implementation of the AvroBuilder interface:

  1. AvroWriterFactory<?> factory = new AvroWriterFactory<>((AvroBuilder<Address>) out -> {
  2.     Schema schema = ReflectData.get().getSchema(Address.class);
  3.     DatumWriter<Address> datumWriter = new ReflectDatumWriter<>(schema);
  5.     DataFileWriter<Address> dataFileWriter = new DataFileWriter<>(datumWriter);
  6.     dataFileWriter.setCodec(CodecFactory.snappyCodec());
  7.     dataFileWriter.create(schema, out);
  8.     return dataFileWriter;
  9. });
  11. DataStream<Address> stream = ...
  12. stream.addSink(StreamingFileSink.forBulkFormat(
  13.     outputBasePath,
  14.     factory).build());
  1. val factory = new AvroWriterFactory[Address](new AvroBuilder[Address]() {
  2.     override def createWriter(out: OutputStream): DataFileWriter[Address] = {
  3.         val schema = ReflectData.get.getSchema(classOf[Address])
  4.         val datumWriter = new ReflectDatumWriter[Address](schema)
  6.         val dataFileWriter = new DataFileWriter[Address](datumWriter)
  7.         dataFileWriter.setCodec(CodecFactory.snappyCodec)
  8.         dataFileWriter.create(schema, out)
  9.         dataFileWriter
  10.     }
  11. })
  13. val stream: DataStream[Address] = ...
  14. stream.addSink(StreamingFileSink.forBulkFormat(
  15.     outputBasePath,
  16.     factory).build());

ORC Format

To enable the data to be bulk encoded in ORC format, Flink offers OrcBulkWriterFactory which takes a concrete implementation of Vectorizer.

Like any other columnar format that encodes data in bulk fashion, Flink’s OrcBulkWriter writes the input elements in batches. It uses ORC’s VectorizedRowBatch to achieve this.

Since the input element has to be transformed to a VectorizedRowBatch, users have to extend the abstract Vectorizer class and override the vectorize(T element, VectorizedRowBatch batch) method. As you can see, the method provides an instance of VectorizedRowBatch to be used directly by the users so users just have to write the logic to transform the input element to ColumnVectors and set them in the provided VectorizedRowBatch instance.

For example, if the input element is of type Person which looks like:

  1. class Person {
  2.     private final String name;
  3.     private final int age;
  4.     ...
  5. }

Then a child implementation to convert the element of type Person and set them in the VectorizedRowBatch can be like:

  1. import org.apache.hadoop.hive.ql.exec.vector.BytesColumnVector;
  2. import org.apache.hadoop.hive.ql.exec.vector.LongColumnVector;
  4. import java.io.IOException;
  5. import java.io.Serializable;
  6. import java.nio.charset.StandardCharsets;
  8. public class PersonVectorizer extends Vectorizer<Person> implements Serializable {    
  9.     public PersonVectorizer(String schema) {
  10.         super(schema);
  11.     }
  12.     @Override
  13.     public void vectorize(Person element, VectorizedRowBatch batch) throws IOException {
  14.         BytesColumnVector nameColVector = (BytesColumnVector) batch.cols[0];
  15.         LongColumnVector ageColVector = (LongColumnVector) batch.cols[1];
  16.         int row = batch.size++;
  17.         nameColVector.setVal(row, element.getName().getBytes(StandardCharsets.UTF_8));
  18.         ageColVector.vector[row] = element.getAge();
  19.     }
  20. }
  1. import java.nio.charset.StandardCharsets
  2. import org.apache.hadoop.hive.ql.exec.vector.{BytesColumnVector, LongColumnVector}
  4. class PersonVectorizer(schema: String) extends Vectorizer[Person](schema) {
  6.   override def vectorize(element: Person, batch: VectorizedRowBatch): Unit = {
  7.     val nameColVector = batch.cols(0).asInstanceOf[BytesColumnVector]
  8.     val ageColVector = batch.cols(1).asInstanceOf[LongColumnVector]
  9.     nameColVector.setVal(batch.size + 1, element.getName.getBytes(StandardCharsets.UTF_8))
  10.     ageColVector.vector(batch.size + 1) = element.getAge
  11.   }
  13. }

To use the ORC bulk encoder in an application, users need to add the following dependency:

  1. <dependency>
  2.   <groupId>org.apache.flink</groupId>
  3.   <artifactId>flink-orc_2.11</artifactId>
  4.   <version>1.12.0</version>
  5. </dependency>

And then a StreamingFileSink that writes data in ORC format can be created like this:

  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink;
  2. import org.apache.flink.orc.writer.OrcBulkWriterFactory;
  4. String schema = "struct<_col0:string,_col1:int>";
  5. DataStream<Person> stream = ...;
  7. final OrcBulkWriterFactory<Person> writerFactory = new OrcBulkWriterFactory<>(new PersonVectorizer(schema));
  9. final StreamingFileSink<Person> sink = StreamingFileSink
  10.     .forBulkFormat(outputBasePath, writerFactory)
  11.     .build();
  13. input.addSink(sink);
  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink
  2. import org.apache.flink.orc.writer.OrcBulkWriterFactory
  4. val schema: String = "struct<_col0:string,_col1:int>"
  5. val input: DataStream[Person] = ...
  6. val writerFactory = new OrcBulkWriterFactory(new PersonVectorizer(schema));
  8. val sink: StreamingFileSink[Person] = StreamingFileSink
  9.     .forBulkFormat(outputBasePath, writerFactory)
  10.     .build()
  12. input.addSink(sink)

OrcBulkWriterFactory can also take Hadoop Configuration and Properties so that a custom Hadoop configuration and ORC writer properties can be provided.

  1. String schema = ...;
  2. Configuration conf = ...;
  3. Properties writerProperties = new Properties();
  5. writerProps.setProperty("orc.compress", "LZ4");
  6. // Other ORC supported properties can also be set similarly.
  8. final OrcBulkWriterFactory<Person> writerFactory = new OrcBulkWriterFactory<>(
  9.     new PersonVectorizer(schema), writerProperties, conf);
  1. val schema: String = ...
  2. val conf: Configuration = ...
  3. val writerProperties: Properties = new Properties()
  5. writerProps.setProperty("orc.compress", "LZ4")
  6. // Other ORC supported properties can also be set similarly.
  8. val writerFactory = new OrcBulkWriterFactory(
  9.     new PersonVectorizer(schema), writerProperties, conf)

The complete list of ORC writer properties can be found here.

Users who want to add user metadata to the ORC files can do so by calling addUserMetadata(...) inside the overriding vectorize(...) method.

  1. public class PersonVectorizer extends Vectorizer<Person> implements Serializable {    
  2.     @Override
  3.     public void vectorize(Person element, VectorizedRowBatch batch) throws IOException {
  4.         ...
  5.         String metadataKey = ...;
  6.         ByteBuffer metadataValue = ...;
  7.         this.addUserMetadata(metadataKey, metadataValue);
  8.     }
  9. }
  1. class PersonVectorizer(schema: String) extends Vectorizer[Person](schema) {
  3.   override def vectorize(element: Person, batch: VectorizedRowBatch): Unit = {
  4.     ...
  5.     val metadataKey: String = ...
  6.     val metadataValue: ByteBuffer = ...
  7.     addUserMetadata(metadataKey, metadataValue)
  8.   }
  10. }

Hadoop SequenceFile format

To use the SequenceFile bulk encoder in your application you need to add the following dependency:

  1. <dependency>
  2.   <groupId>org.apache.flink</groupId>
  3.   <artifactId>flink-sequence-file</artifactId>
  4.   <version>1.12.0</version>
  5. </dependency>

A simple SequenceFile writer can be created like this:

  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink;
  2. import org.apache.flink.configuration.GlobalConfiguration;
  3. import org.apache.hadoop.conf.Configuration;
  4. import org.apache.hadoop.io.LongWritable;
  5. import org.apache.hadoop.io.SequenceFile;
  6. import org.apache.hadoop.io.Text;
  9. DataStream<Tuple2<LongWritable, Text>> input = ...;
  10. Configuration hadoopConf = HadoopUtils.getHadoopConfiguration(GlobalConfiguration.loadConfiguration());
  11. final StreamingFileSink<Tuple2<LongWritable, Text>> sink = StreamingFileSink
  12.   .forBulkFormat(
  13.     outputBasePath,
  14.     new SequenceFileWriterFactory<>(hadoopConf, LongWritable.class, Text.class))
  15.     .build();
  17. input.addSink(sink);
  1. import org.apache.flink.streaming.api.functions.sink.filesystem.StreamingFileSink
  2. import org.apache.flink.configuration.GlobalConfiguration
  3. import org.apache.hadoop.conf.Configuration
  4. import org.apache.hadoop.io.LongWritable
  5. import org.apache.hadoop.io.SequenceFile
  6. import org.apache.hadoop.io.Text;
  8. val input: DataStream[(LongWritable, Text)] = ...
  9. val hadoopConf: Configuration = HadoopUtils.getHadoopConfiguration(GlobalConfiguration.loadConfiguration())
  10. val sink: StreamingFileSink[(LongWritable, Text)] = StreamingFileSink
  11.   .forBulkFormat(
  12.     outputBasePath,
  13.     new SequenceFileWriterFactory(hadoopConf, LongWritable.class, Text.class))
  14.     .build()
  16. input.addSink(sink)

The SequenceFileWriterFactory supports additional constructor parameters to specify compression settings.

Bucket Assignment

The bucketing logic defines how the data will be structured into subdirectories inside the base output directory.

Both row and bulk formats (see File Formats) use the DateTimeBucketAssigner as the default assigner. By default the DateTimeBucketAssigner creates hourly buckets based on the system default timezone with the following format: yyyy-MM-dd--HH. Both the date format (i.e. bucket size) and timezone can be configured manually.

We can specify a custom BucketAssigner by calling .withBucketAssigner(assigner) on the format builders.

Flink comes with two built in BucketAssigners:

Rolling Policy

The RollingPolicy defines when a given in-progress part file will be closed and moved to the pending and later to finished state. Part files in the “finished” state are the ones that are ready for viewing and are guaranteed to contain valid data that will not be reverted in case of failure. The Rolling Policy in combination with the checkpointing interval (pending files become finished on the next checkpoint) control how quickly part files become available for downstream readers and also the size and number of these parts.

Flink comes with two built-in RollingPolicies:

Part file lifecycle

In order to use the output of the StreamingFileSink in downstream systems, we need to understand the naming and lifecycle of the output files produced.

Part files can be in one of three states:

  1. In-progress : The part file that is currently being written to is in-progress
  2. Pending : Closed (due to the specified rolling policy) in-progress files that are waiting to be committed
  3. Finished : On successful checkpoints pending files transition to “Finished”

Only finished files are safe to read by downstream systems as those are guaranteed to not be modified later.

IMPORTANT: Part file indexes are strictly increasing for any given subtask (in the order they were created). However these indexes are not always sequential. When the job restarts, the next part index for all subtask will be the `max part index + 1` where `max` is computed across all subtasks.

Each writer subtask will have a single in-progress part file at any given time for every active bucket, but there can be several pending and finished files.

Part file example

To better understand the lifecycle of these files let’s look at a simple example with 2 sink subtasks:

  1. └── 2019-08-25--12
  2.     ├── part-0-0.inprogress.bd053eb0-5ecf-4c85-8433-9eff486ac334
  3.     └── part-1-0.inprogress.ea65a428-a1d0-4a0b-bbc5-7a436a75e575

When the part file part-1-0 is rolled (let’s say it becomes too large), it becomes pending but it is not renamed. The sink then opens a new part file: part-1-1:

  1. └── 2019-08-25--12
  2.     ├── part-0-0.inprogress.bd053eb0-5ecf-4c85-8433-9eff486ac334
  3.     ├── part-1-0.inprogress.ea65a428-a1d0-4a0b-bbc5-7a436a75e575
  4.     └── part-1-1.inprogress.bc279efe-b16f-47d8-b828-00ef6e2fbd11

As part-1-0 is now pending completion, after the next successful checkpoint, it is finalized:

  1. └── 2019-08-25--12
  2.     ├── part-0-0.inprogress.bd053eb0-5ecf-4c85-8433-9eff486ac334
  3.     ├── part-1-0
  4.     └── part-1-1.inprogress.bc279efe-b16f-47d8-b828-00ef6e2fbd11

New buckets are created as dictated by the bucketing policy, and this doesn’t affect currently in-progress files:

  1. └── 2019-08-25--12
  2.     ├── part-0-0.inprogress.bd053eb0-5ecf-4c85-8433-9eff486ac334
  3.     ├── part-1-0
  4.     └── part-1-1.inprogress.bc279efe-b16f-47d8-b828-00ef6e2fbd11
  5. └── 2019-08-25--13
  6.     └── part-0-2.inprogress.2b475fec-1482-4dea-9946-eb4353b475f1

Old buckets can still receive new records as the bucketing policy is evaluated on a per-record basis.

Part file configuration

Finished files can be distinguished from the in-progress ones by their naming scheme only.

By default, the file naming strategy is as follows:

  • In-progress / Pending: part-<subtaskIndex>-<partFileIndex>.inprogress.uid
  • Finished: part-<subtaskIndex>-<partFileIndex>

Flink allows the user to specify a prefix and/or a suffix for his/her part files. This can be done using an OutputFileConfig. For example for a prefix “prefix” and a suffix “.ext” the sink will create the following files:

  1. └── 2019-08-25--12
  2.     ├── prefix-0-0.ext
  3.     ├── prefix-0-1.ext.inprogress.bd053eb0-5ecf-4c85-8433-9eff486ac334
  4.     ├── prefix-1-0.ext
  5.     └── prefix-1-1.ext.inprogress.bc279efe-b16f-47d8-b828-00ef6e2fbd11

The user can specify an OutputFileConfig in the following way:

  1. OutputFileConfig config = OutputFileConfig
  2.  .builder()
  3.  .withPartPrefix("prefix")
  4.  .withPartSuffix(".ext")
  5.  .build();
  7. StreamingFileSink<Tuple2<Integer, Integer>> sink = StreamingFileSink
  8.  .forRowFormat((new Path(outputPath), new SimpleStringEncoder<>("UTF-8"))
  9.  .withBucketAssigner(new KeyBucketAssigner())
  10.  .withRollingPolicy(OnCheckpointRollingPolicy.build())
  11.  .withOutputFileConfig(config)
  12.  .build();
  1. val config = OutputFileConfig
  2.  .builder()
  3.  .withPartPrefix("prefix")
  4.  .withPartSuffix(".ext")
  5.  .build()
  7. val sink = StreamingFileSink
  8.  .forRowFormat(new Path(outputPath), new SimpleStringEncoder[String]("UTF-8"))
  9.  .withBucketAssigner(new KeyBucketAssigner())
  10.  .withRollingPolicy(OnCheckpointRollingPolicy.build())
  11.  .withOutputFileConfig(config)
  12.  .build()

Important Considerations

General

Important Note 1: When using Hadoop < 2.7, please use the OnCheckpointRollingPolicy which rolls part files on every checkpoint. The reason is that if part files “traverse” the checkpoint interval, then, upon recovery from a failure the StreamingFileSink may use the truncate() method of the filesystem to discard uncommitted data from the in-progress file. This method is not supported by pre-2.7 Hadoop versions and Flink will throw an exception.

Important Note 2: Given that Flink sinks and UDFs in general do not differentiate between normal job termination (e.g. finite input stream) and termination due to failure, upon normal termination of a job, the last in-progress files will not be transitioned to the “finished” state.

Important Note 3: Flink and the StreamingFileSink never overwrites committed data. Given this, when trying to restore from an old checkpoint/savepoint which assumes an in-progress file which was committed by subsequent successful checkpoints, the StreamingFileSink will refuse to resume and it will throw an exception as it cannot locate the in-progress file.

Important Note 4: Currently, the StreamingFileSink only supports three filesystems: HDFS, S3, and Local. Flink will throw an exception when using an unsupported filesystem at runtime.

S3-specific

Important Note 1: For S3, the StreamingFileSink supports only the Hadoop-based FileSystem implementation, not the implementation based on Presto. In case your job uses the StreamingFileSink to write to S3 but you want to use the Presto-based one for checkpointing, it is advised to use explicitly “s3a://” (for Hadoop) as the scheme for the target path of the sink and “s3p://” for checkpointing (for Presto). Using “s3://” for both the sink and checkpointing may lead to unpredictable behavior, as both implementations “listen” to that scheme.

Important Note 2: To guarantee exactly-once semantics while being efficient, the StreamingFileSink uses the Multi-part Upload feature of S3 (MPU from now on). This feature allows to upload files in independent chunks (thus the “multi-part”) which can be combined into the original file when all the parts of the MPU are successfully uploaded. For inactive MPUs, S3 supports a bucket lifecycle rule that the user can use to abort multipart uploads that don’t complete within a specified number of days after being initiated. This implies that if you set this rule aggressively and take a savepoint with some part-files being not fully uploaded, their associated MPUs may time-out before the job is restarted. This will result in your job not being able to restore from that savepoint as the pending part-files are no longer there and Flink will fail with an exception as it tries to fetch them and fails.