This connector provides a unified Source and Sink for
BATCH
and
STREAMING
that reads or writes (partitioned) files to file systems
supported by the
Flink
FileSystem
abstraction
. This filesystem
connector provides the same guarantees for both
BATCH
and
STREAMING
and is designed to provide exactly-once semantics for
STREAMING
execution.
The connector supports reading and writing a set of files from any (distributed) file system (e.g. POSIX, S3, HDFS)
with a
format
(e.g., Avro, CSV, Parquet),
and produces a stream or records.
File Source
The
File Source
is based on the
Source API
,
a unified data source that reads files - both in batch and in streaming mode.
It is divided into the following two parts:
SplitEnumerator
and
SourceReader
.
SplitEnumerator
is responsible for discovering and identifying the files to read and assigns them to the
SourceReader
.
SourceReader
requests the files it needs to process and reads the file from the filesystem.
You will need to combine the File Source with a
format
, which allows you to
parse CSV, decode AVRO, or read Parquet columnar files.
Bounded and Unbounded Streams
A bounded
File Source
lists all files (via SplitEnumerator - a recursive directory list with filtered-out hidden files) and reads them all.
An unbounded
File Source
is created when configuring the enumerator for periodic file discovery.
In this case, the
SplitEnumerator
will enumerate like the bounded case but, after a certain interval, repeats the enumeration.
For any repeated enumeration, the
SplitEnumerator
filters out previously detected files and only sends new ones to the
SourceReader
.
Usage
You can start building a File Source via one of the following API calls:
// reads the contents of a file from a file stream.
FileSource.forRecordStreamFormat(StreamFormat,Path...);// reads batches of records from a file at a time
FileSource.forBulkFileFormat(BulkFormat,Path...);
# reads the contents of a file from a file stream.FileSource.for_record_stream_format(stream_format,*path)# reads batches of records from a file at a timeFileSource.for_bulk_file_format(bulk_format,*path)
This creates a FileSource.FileSourceBuilder on which you can configure all the properties of the File Source.
For the bounded/batch case, the File Source processes all files under the given path(s).
For the continuous/streaming case, the source periodically checks the paths for new files and will start reading those.
When you start creating a File Source (via the FileSource.FileSourceBuilder created through one of the above-mentioned methods),
the source is in bounded/batch mode by default. You can call AbstractFileSource.AbstractFileSourceBuilder.monitorContinuously(Duration)
to put the source into continuous streaming mode.
source=FileSource.for_record_stream_format(...) \
.monitor_continously(Duration.of_millis(5)) \
.build()
Format Types
The reading of each file happens through file readers defined by file formats.
These define the parsing logic for the contents of the file. There are multiple classes that the source supports.
The interfaces are a tradeoff between simplicity of implementation and flexibility/efficiency.
A StreamFormat reads the contents of a file from a file stream. It is the simplest format to implement,
and provides many features out-of-the-box (like checkpointing logic) but is limited in the optimizations it can apply
(such as object reuse, batching, etc.).
A BulkFormat reads batches of records from a file at a time.
It is the most “low level” format to implement, but offers the greatest flexibility to optimize the implementation.
TextLine Format
A StreamFormat reader formats text lines from a file.
The reader uses Java’s built-in InputStreamReader to decode the byte stream using
various supported charset encodings.
This format does not support optimized recovery from checkpoints. On recovery, it will re-read
and discard the number of lines that were processed before the last checkpoint. This is due to
the fact that the offsets of lines in the file cannot be tracked through the charset decoders
with their internal buffering of stream input and charset decoder state.
SimpleStreamFormat Abstract Class
This is a simple version of StreamFormat for formats that are not splittable.
Custom reads of Array or File can be done by implementing SimpleStreamFormat:
The schema for CSV parsing, in this case, is automatically derived based on the fields of the SomePojo class using the Jackson library. (Note: you might need to add @JsonPropertyOrder({field1, field2, ...}) annotation to your class definition with the fields order exactly matching those of the CSV file columns).
If you need more fine-grained control over the CSV schema or the parsing options, use the more low-level forSchema static factory method of CsvReaderFormat:
CsvReaderFormat<T>forSchema(Supplier<CsvMapper>mapperFactory,Function<CsvMapper,CsvSchema>schemaGenerator,TypeInformation<T>typeInformation)
Bulk Format
The BulkFormat reads and decodes batches of records at a time. Examples of bulk formats
are formats like ORC or Parquet.
The outer BulkFormat class acts mainly as a configuration holder and factory for the
reader. The actual reading is done by the BulkFormat.Reader, which is created in the
BulkFormat#createReader(Configuration, FileSourceSplit) method. If a bulk reader is
created based on a checkpoint during checkpointed streaming execution, then the reader is
re-created in the BulkFormat#restoreReader(Configuration, FileSourceSplit) method.
A SimpleStreamFormat can be turned into a BulkFormat by wrapping it in a StreamFormatAdapter:
BulkFormat<SomePojo,FileSourceSplit>bulkFormat=newStreamFormatAdapter<>(CsvReaderFormat.forPojo(SomePojo.class));
Customizing File Enumeration
* A FileEnumerator implementation for hive source, which generates splits based on
* HiveTablePartition.
publicclassHiveSourceFileEnumeratorimplementsFileEnumerator{// reference constructor
publicHiveSourceFileEnumerator(...){ * Generates all file splits for the relevant files under the given paths. The {@code
* minDesiredSplits} is an optional hint indicating how many splits would be necessary to
* exploit parallelism properly.
@OverridepublicCollection<FileSourceSplit>enumerateSplits(Path[]paths,intminDesiredSplits)throwsIOException{// createInputSplits:splitting files into fragmented collections
returnnewArrayList<>(createInputSplits(...)); * A factory to create HiveSourceFileEnumerator.
publicstaticclassProviderimplementsFileEnumerator.Provider{@OverridepublicFileEnumeratorcreate(){returnnewHiveSourceFileEnumerator(...);// use the customizing file enumeration
newHiveSource<>(newHiveSourceFileEnumerator.Provider(partitions!=null?partitions:Collections.emptyList(),newJobConfWrapper(jobConf)),...);
Current Limitations
Watermarking does not work very well for large backlogs of files. This is because watermarks eagerly advance within a file, and the next file might contain data later than the watermark.
For Unbounded File Sources, the enumerator currently remembers paths of all already processed files, which is a state that can, in some cases, grow rather large.
There are plans to add a compressed form of tracking already processed files in the future (for example, by keeping modification timestamps below boundaries).
Behind the Scenes
If you are interested in how File Source works through the new data source API design, you may
want to read this part as a reference. For details about the new data source API, check out the
documentation on data sources and
FLIP-27
for more descriptive discussions.
File Sink
The file sink writes incoming data into buckets. Given that the incoming streams can be unbounded,
data in each bucket is 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 is 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. For Row-encoded Formats (see File Formats) 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. For Bulk-encoded Formats we roll on every checkpoint and the user can
specify additional conditions based on size or time.
IMPORTANT: Checkpointing needs to be enabled when using the FileSink in STREAMING mode. 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.
The FileSink 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:
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
FileSink
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) : Interval for checking time based rolling policies
Basic usage for writing String elements thus looks like this:
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 five built-in BulkWriter factories:
ParquetWriterFactory
AvroWriterFactory
SequenceFileWriterFactory
CompressWriterFactory
OrcBulkWriterFactory
Important Bulk Formats can only have a rolling policy that extends the CheckpointRollingPolicy.
The latter rolls on every checkpoint. A policy can roll additionally based on size or processing time.
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 AvroParquetWriters 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:
schema=AvroSchema.parse_string(JSON_SCHEMA)# The element could be vanilla Python data structure matching the schema,# which is annotated with default Types.PICKLED_BYTE_ARRAY()data_stream=...avro_type_info=GenericRecordAvroTypeInfo(schema)sink=FileSink \
.for_bulk_format(OUTPUT_BASE_PATH,AvroParquetWriters.for_generic_record(schema)) \
.build()# A map to indicate its Avro type info is necessary for serializationdata_stream.map(lambdae:e,output_type=avro_type_info).sink_to(sink)
Similarly, a FileSink that writes Protobuf data to Parquet format can be created like this:
importorg.apache.flink.connector.file.sink.FileSink;importorg.apache.flink.formats.parquet.protobuf.ParquetProtoWriters;// ProtoRecord is a generated protobuf Message class.
DataStream<ProtoRecord>input=...;finalFileSink<ProtoRecord>sink=FileSink.forBulkFormat(outputBasePath,ParquetProtoWriters.forType(ProtoRecord.class)).build();input.sinkTo(sink);
importorg.apache.flink.connector.file.sink.FileSink;importorg.apache.flink.formats.parquet.protobuf.ParquetProtoWriters// ProtoRecord is a generated protobuf Message class.
valinput:DataStream[ProtoRecord]=...valsink:FileSink[ProtoRecord]=FileSink.forBulkFormat(outputBasePath,ParquetProtoWriters.forType(classOf[ProtoRecord])).build()input.sinkTo(sink)
For PyFlink users, ParquetBulkWriters could be used to create a BulkWriterFactory that writes Rows into Parquet files.
row_type=DataTypes.ROW([DataTypes.FIELD('string',DataTypes.STRING()),DataTypes.FIELD('int_array',DataTypes.ARRAY(DataTypes.INT()))sink=FileSink.for_bulk_format(OUTPUT_DIR,ParquetBulkWriters.for_row_type(row_type,hadoop_config=Configuration(),utc_timestamp=True,).build()ds.sink_to(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:
schema=AvroSchema.parse_string(JSON_SCHEMA)# The element could be vanilla Python data structure matching the schema,# which is annotated with default Types.PICKLED_BYTE_ARRAY()data_stream=...avro_type_info=GenericRecordAvroTypeInfo(schema)sink=FileSink \
.for_bulk_format(OUTPUT_BASE_PATH,AvroBulkWriters.for_generic_record(schema)) \
.build()# A map to indicate its Avro type info is necessary for serializationdata_stream.map(lambdae:e,output_type=avro_type_info).sink_to(sink)
For creating customized Avro writers, e.g. enabling compression, users need to create the AvroWriterFactory
with a custom implementation of the AvroBuilder interface:
valfactory=newAvroWriterFactory[Address](newAvroBuilder[Address](){overridedefcreateWriter(out:OutputStream):DataFileWriter[Address]={valschema=ReflectData.get.getSchema(classOf[Address])valdatumWriter=newReflectDatumWriter[Address](schema)valdataFileWriter=newDataFileWriter[Address](datumWriter)dataFileWriter.setCodec(CodecFactory.snappyCodec)dataFileWriter.create(schema,out)dataFileWritervalstream:DataStream[Address]=...stream.sinkTo(FileSink.forBulkFormat(outputBasePath,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:
OrcBulkWriterFactory can also take Hadoop Configuration and Properties so that a custom Hadoop configuration and ORC
writer properties can be provided.
Stringschema=...;Configurationconf=...;PropertieswriterProperties=newProperties();writerProperties.setProperty("orc.compress","LZ4");// Other ORC supported properties can also be set similarly.
finalOrcBulkWriterFactory<Person>writerFactory=newOrcBulkWriterFactory<>(newPersonVectorizer(schema),writerProperties,conf);
valschema:String=...valconf:Configuration=...valwriterProperties:Properties=newProperties()writerProperties.setProperty("orc.compress","LZ4")// Other ORC supported properties can also be set similarly.
valwriterFactory=newOrcBulkWriterFactory(newPersonVectorizer(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.
For PyFlink users, OrcBulkWriters could be used to create BulkWriterFactory to write records to files in Orc format.
row_type=DataTypes.ROW([DataTypes.FIELD('name',DataTypes.STRING()),DataTypes.FIELD('age',DataTypes.INT()),sink=FileSink.for_bulk_format(OUTPUT_DIR,OrcBulkWriters.for_row_type(row_type=row_type,writer_properties=Configuration(),hadoop_config=Configuration(),).build()ds.sink_to(sink)
Hadoop SequenceFile format
To use the SequenceFile bulk encoder in your application you need to add the following dependency:
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:
DateTimeBucketAssigner : Default time based assigner
BasePathBucketAssigner : Assigner that stores all part files in the base path (single global bucket)
Note: PyFlink only supports DateTimeBucketAssigner and BasePathBucketAssigner.
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.
In STREAMING mode, 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. In BATCH mode, part-files become visible at the end of the job but
the rolling policy can control their maximum size.
Flink comes with two built-in RollingPolicies:
DefaultRollingPolicy
OnCheckpointRollingPolicy
Note: PyFlink only supports DefaultRollingPolicy and OnCheckpointRollingPolicy.
Part file lifecycle
In order to use the output of the FileSink 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:
In-progress : The part file that is currently being written to is in-progress
Pending : Closed (due to the specified rolling policy) in-progress files that are waiting to be committed
Finished : On successful checkpoints (STREAMING) or at the end of input (BATCH) pending files transition to “Finished”
Only finished files are safe to read by downstream systems as those are guaranteed to not be modified later.
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:
When the part file part-81fc4980-a6af-41c8-9937-9939408a734b-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-81fc4980-a6af-41c8-9937-9939408a734b-1:
Finished:part-<uid>-<partFileIndex>
where uid is a random id assigned to a subtask of the sink when the subtask is instantiated. This uid is not fault-tolerant
so it is regenerated when the subtask recovers from a failure.
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:
Since version 1.15 FileSink supports compaction of the pending files,
which allows the application to have smaller checkpoint interval without generating a lot of small files,
especially when using the bulk encoded formats
that have to rolling on taking checkpoints.
Once enabled, the compaction happens between the files become pending and get committed. The pending files will
be first committed to temporary files whose path starts with .. Then these files will be compacted according to
the strategy by the compactor specified by the users, and the new compacted pending files will be generated.
Then these pending files will be emitted to the committer to be committed to the formal files. After that, the source files will be removed.
When enabling compaction, you need to specify the
FileCompactStrategy
and the
FileCompactor
FileCompactStrategy
specifies
when and which files get compacted. Currently, there are two parallel conditions: the target file size and the number of checkpoints get passed.
Once the total size of the cached files has reached the size threshold or the number of checkpoints since the last compaction has reached the specified number,
the cached files will be scheduled to compact.
FileCompactor
specifies how to compact
the give list of Path and write the result file. It could be classified into two types according to how to write the file:
OutputStreamBasedFileCompactor
The users can write the compacted results into an output stream. This is useful when the users don’t want to or can’t read records from the input files.
An example is the
ConcatFileCompactor
that concats the list of files directly.
RecordWiseFileCompactor
The compactor can read records one-by-one from the input files and write into the result file similar to the FileWriter.
An example is the
RecordWiseFileCompactor
that reads records from the source files and then writes them with the CompactingFileWriter. Users need to specify how to read records from the source files.
Important Note 1 Once the compaction is enabled, you must explicitly call disableCompact when building the FileSink if you want to disable compaction.
Important Note 2 When the compaction is enabled, the written files need to wait for longer time before they get visible.
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 FileSink 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 FileSink 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 FileSink will refuse to resume and will throw an exception as it cannot locate the
in-progress file.
Important Note 4: Currently, the FileSink only supports five filesystems:
HDFS, S3, OSS, ABFS and Local. Flink will throw an exception when using an unsupported filesystem at runtime.
BATCH-specific
Important Note 1: Although the Writer is executed with the user-specified
parallelism, the Committer is executed with parallelism equal to 1.
Important Note 2: Pending files are committed, i.e. transition to Finished
state, after the whole input has been processed.
Important Note 3: When High-Availability is activated, if a JobManager
failure happens while the Committers are committing, then we may have duplicates. This is going to be fixed in
future Flink versions
(see progress in FLIP-147).
S3-specific
Important Note 1: For S3, the FileSink
supports only the Hadoop-based FileSystem implementation, not
the implementation based on Presto. In case your job uses the
FileSink 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 FileSink 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.
OSS-specific
Important Note: To guarantee exactly-once semantics while
being efficient, the FileSink also uses the Multi-part Upload
feature of OSS(similar with S3).