Tag: Streaming

Spark Structured Streaming Deep Dive (2) – Source

Spark Structured Streaming Deep Dive (2) – Source

As mentioned in the last blog discussing the execution flow of Spark Structured Streaming queries, the Spark Structured Streaming framework consists of three main components, Source, StreamExecution, and Sink. The source interfaces defined by the Spark Structured Streaming framework abstract the input data stream from the external streaming data sources and standarise the interaction patterns between the Spark stream querying engine and the external data sources. All the concrete Spark streaming data source classes, no matter the built-in ones or custom developed, need to implement those interfaces.

This blog post discusses the streaming source interfaces and methods defined in Spark Structured Streaming. In addition, one of the concrete streaming data sources which implements those interfaces, FileStreamSource, is discussed in details. Kafka is the primary and the most used streaming source in the industry. Spark Structured Streaming framework ships the built-in KafkaSource to support Kafka input streams. I have planned a deep dive blog series of Kafka and Kafka-Spark integration. The KafkaSource class will be discussed in details there.

The diagram below depicts the Spark Structured Streaming source class hierarchy.

SparkDataStream is the root interface representing a readable data input stream in a Spark Structured streaming query. It is the top-level abstraction of input streams and all the concrete data sources should implement this interface.

Source trait extends the SparkDataStream root interface. It is defined to abstract and represent all data sources with continually arriving data. Offset is the abstract Source trait uses to represent the monotonically increasing notion of the progress of the input stream. Source trait defines the ‘getOffset‘ method for fetching the maximum available offset for a source. Based on the last committed offset and the latest available offset, Spark can fetch the batch with all new data through the ‘getBatch‘ method defined by the Source trait. The Source trait also defines the ‘commit‘ method which is used to notify the source that the current batch has been processed successfully by Spark and the source can start to conduct the cleaning, garbage collection tasks.

SupportsAdmissionControl interface also extends the SparkDataStream and represents the streaming data sources which are able to control the rate of data ingested into the Spark. The rate of data ingestion can be controlled by the ReadLimit which can be configured by the end-user developers through options when creating the data source. It can also be controlled implicitly from the trigger modes, for example, OneTimeTrigger requires a micro-batch process all data available which fits the one-off historical data loading scenario. SupportsAdmissionControl provides the ‘latestOffset‘ method which takes ReadLimit as a parameter so that the latest offset available to read is decided by considering the given read limit. The concrete data source classes which support data ingestion rate control need to implement this method instead of the ‘getOffset‘ for Source trait.

The concrete sources need to implement the Source trait to support stream inputs. If the sources need to control the ingestion rate, they need to implement the SupportsAdmissionControl interface. Both the FileStreamSource and the KafkaSource implements the Source trait and SupportsAdmissionControl. The end-user developers can also create custom sources to support all sorts of external sources by implementing those interfaces. The development of custom sources is out of the scope of this blog post, but if you are interested, here is a nice article regarding this:

Spark Custom Stream Sources by Nicolas A Perez

In the rest of the blog post, I will look into how FileStreamSource works with details and I will cover the KafkaStreamSource in the subsequent blog posts of Spark-Kafka integration.

FileStreamSource represents a file-system data source which takes the files written in a directory as a stream of data, with the file modification time as the event time. A list file formats are supported, including csv, test, json, Parquet, ORC. The schema of the files can be defined by the end-user developer when creating the DataStreamWriter. Some file formats support schema inference, but it is not recommended.

The main responsibility of the FileStreamSource is to implement the methods defined in the Source trait and SupportsAdmissionControl interface so that a FileStreamSource instance can be consumed by Spark. The core method to implement for FileStreamSource are: latestOffset, getBatch, and commit.


This is a diagram I draw to depict how FileStreamSource gets the lastestOffset:

Here is a list of the key components involved in the process:

  • The ‘fs‘ member (used in ‘fetchAllFiles‘ method), the reference for accessing the directory (source) in a Hadoop compatible file system.
  • The ‘sourceOptions‘, the configuration settings defined by the end-user developers for controlling how the source works, such as max files per trigger (read limit).
  • The ‘metadataLog‘, records the metadata of the batch and file entries.
  • The ‘metadataLogCurrentOffset‘, the latest offset in the metadataLog.
  • The ‘seenFiles‘, the log of the files that have been placed into the current batch to process.
  • The ‘unreadFiles‘, the list of files that was read in the last batch, but was not processed (due to over the max files per trigger limit.

The FileStreamSource defines the fetchMaxOffset method for getting the latestOffset. This method first checks whether or not there is unreadFiles that was fetched, but not processed in the last batch. If there is, set the unreadFiles as newFiles, if no, call the fetchAllFiles method to fetch files from the file system and applies isNewFile method (of seenFiles) to return the unprocessed files only as newFiles.

If the newFiles set contains more files than the maxFilesPerTrigger, it will be split into two sets by the maxFilesPerTrigger value. The current batch, batchFiles, is capped by the maxFilesPerTrigger number, and the remaining files will be cached in the unselectedFiles. The unselectedFiles will be marked as unreadFiles and will be processed in the next batch. However, if the number of files in the unselectedFiles is too small (maxFilesPerTrigger * DISCARD_UNSEEN_FILES_RATIO) that the next batch would have too few files to read, the unselectedFiles will be discarded and will be reread along with other new files in the next batch.

The files to return in the current batch are logged in the seenFiles. A purge operation is then applied on the seenFiles to remove aged entries (based on the maxFileAgeMs option configured by the end-user developers). The metadataLogCurrentOffset increases by 1 to update the latest offset. All the files in the current batch are associated with the new metadataLogCurrentOffset and logged in the metadataLog. The new metadataLogCurentOffset will be returned by the fetchMaxOffset method.


The getBatch method takes the start offset and end offset as arguments and return a DataFrame instance, which represents the data rows in the files recorded in the metadataLog between the start offset and the end offset. The returned Dataframe defines the data rows to process by Saprk Structured Streaming framework in a micro-batch.

Be aware that the ‘Batch’ concept of the getBatch method is different with the ‘Batch’ concept used in the fetchMaxOffset method mentioned above. The batch returned by the getBatch method refers to the batch of rows in the files associating with one of the metadataLog offset which is between the specified start offset and end offset of the getBatch method. The batch used in the fetchMaxOffset method refers to the batch of files associating with one metadataLog offset.


In the Micro-Batch processing mode, the commit method will be called by the MicroBatchExecution to inform the source that the all data up to the given offset is processed so the source can start the post-batch processing.

FileStreamSource implements the commit method for cleaning up completed files. The end-user developers can specify the option to clean, including ‘archive’, ‘delete’ and ‘off’. Internally, FileStreamSource defines the FileStreamSourceCleaner private abstract class and two concrete classes: SourceFileArchiver for ‘archive’ option and SourceFileRemover for ‘delete’ option.

To reduce the performance impacts of the cleaning task on the micro-batch processing thread, a cleanThreadPool is maintained and all the concrete implmentations (such as SourceFileArchiver and SourceFileRemover) need to implement the cleanTask with their specific cleaning logic and submit the cleanTask to execute in a seperate thread. the number of threads used in the cleanThreadPool can be specified with the spark.sql.streaming.fileSource.cleaner.numThreads.

Spark Structured Streaming Deep Dive (1) – Execution Flow

Spark Structured Streaming Deep Dive (1) – Execution Flow

From this blog post, I am starting to write about streaming processing, focusing on Spark Structured Streaming, Kafka, Flink and Kappa architecture. This is the first blog post of the Spark Structured Streaming deep dive series. This blog post digs into the underlying, end-to-end execution flow of Spark streaming queries.

Firstly, let’s have a look at the official structured streaming query example from Spark:

A typical streaming query is defined in three steps: firstly create a data stream reader in the current Spark session that will be used to read streaming data in as a Spark DataFrame, then define the transforms on the DataFrame based on the processing requirements, and finally define the data stream writer and start running the query.

Now, let’s look into how a streaming query is executed by the Spark Structured Streaming framework under the hood.

SparkSession provides a ‘readStream‘ method which can be used by end-user developers to create and access an DataStreamReader instance that is responsible for loading an input stream from external data source. The data source format and input schema can be specified in the data stream reader. The ‘load‘ method of the data stream reader defines the Spark DataFrame which logically representing the input stream.

The transformation of the DataFrame that represents the data processing logics can be defined by the end-user developers by calling the operations available on the DataFrame. Most of the common operations available for batch processing are also supported for stream processing. Here is a list of unsupported operations. Just as how it works in the batch processing workloads, the DataFrame operations for streaming processing are also executed in the lazy mode. Under the hood, a Dataframe represents a logical plan which defines the computations required for the data processing. The computations will only be triggered when an action is invoked.

After data transformations of the DataFrame are defined, a DataStreamWriter instance needs to be created for saving the processed streaming data out into external storage. The DataStreamWriter can be created by calling the ‘writeStream‘ method of the DataFrame. The end-user developers can specify the output mode for writing the data into streaming sink, the partitions of the sink file system where the data written into, and the trigger for the micro-batch streaming jobs. Up to this point, all we have done is just to define the streaming query execution. The physical streaming query execution starts when the ‘start’ method of the DataStreamWriter is called.

The ‘start‘ method of the DataStreamWriter first creates the sink instance which supports continual writing data into external storage, and then passes the sink instance to the StreamingQueryManager of the current Spark session to create and start a StreamExecution instance. The StreamExecution instance is the core component that manages the lifecycle of a streaming query execution. There are two concrete implementations of the StreamExecution (which itself is defined as an abstract class), including the MicroBatchExecution and the ContinuousExecution. The micro-batch processing mode, i.e., processing data streams as a series of small batch jobs, implemented by MicroBatchExecution is the primary stream processing mode supported by Spark Structured Streaming engine. This mode supports exactly-once fault-tolerance guarantees and is capable to achieve end-to-end latencies at 100ms level. The ContinuousExecution supports the low-latency, continuous processing mode that achieves end-to-end latencies at 1ms level. However, this mode only supports map-like operations and can only achieve at-least-once guarantees. This blog post focuses on the micro-batch processing mode. I will dedicate a few blog posts to continuous processing mode in future.

Now, let’s look into a bit deeper on how StreamQueryManager creates and runs a streaming query (i.e. a StreamExecution instance). StreamingQueryManager is responsible to manage all the active streaming queries in a Spark session. It comes with an ‘activeQueries‘ hash map which caches the reference to the active streaming queries. The ‘start’ call from the DataStreamWriter will trigger the ‘startQuery‘ method which conducts three major activities in order, including defining streaming query (i.e. creating the StreamExecution instance), registering the created StreamExecution instance in the activeQueries hash map, and invoking the ‘start’ call of the StreamExecution instance.

To create a streaming query, logical plan of the DataFrame which is defined earlier with the required transformation operations is analysed. Please refer to my earlier blog post for more details of the Spark Catalyst Analyzer. With the analysed logical plan along with other streaming writing configurations, such as sink, output mode, checkpoint location and trigger), a WriteToStreamStatement is created which is the logical node representing a stream writing plan. After the WriteToStreamStatement is analysed, depending on the trigger type (ContinuousTrigger or not), one of the StreamExecution implementation classes, ContinuousExecution or MicroBatchExecution, is created to manage the stream query execution. The created streaming query is then registered in the activeQueries.

After that, the ‘start’ method of the created StreamExecution instance is called to start a separate thread to run the streaming querying which executes repeatedly each time when new data arrives at any source specified in the query plan. Within the thread, after the initialisation of source and metadata objects, the ‘runActivatedStream‘ method is called. The execution flow of the ‘runActivatedStream‘ method is defined in each implementation class of the StreamExecution. The rest of the blog post looks into the implementation of this method of the MicroBatchExecution. However, before digging into the method, let’s have a step back and have a look at the core components in the Sparking Structured Streaming system.

Firstly, we have external streaming sources and sinks which support continuous data read and write. Within the Spark Structured Streaming framework, there are three main components: the Source interface which abstracts the data read from the external streaming sources, the Sink interface which abstracts the data write into the external Streaming sinks, and the StreamExecution which manages the lifecycle of streaming query executions. The Source interface defines three core methods for communicating with external data sources, getOffset(), getBatch(), and commit(), while the Sink interface defines one core method for output the streaming data to sink, addBatch(). I will discuss those methods in details in the subsequent blog posts which digs into the sources and sinks implementation in the Spark Structured Streaming framework.

The execution flow of each micro-batch data processing is defined in the ‘runActivatedStream‘ method of the MicroBatchExecution class. A TriggerExecutor (ProcessingTimeExecutor by default) instance, is created to schedule and trigger micro-batch processing jobs. I draw the following diagram to depict the end-to-end execution flow of a micro-batch job.

(1) The execution of micro-batch processing starts from identifying the start offsets to read from the sources. The start offsets are determined based on the last batch execution recorded in the OffsetLog and CommitLog. The OffsetLog is a Write-Ahead Log (WAL) recording the offsets of the current batch before the batch run while the CommitLog records the batch commit when the batch run is completed successfully. If the last batch id recorded in the OffsetLog equals to the last batch id recorded in the CommitLog, it implies the last batch was successfully processed, so the start offsets of current batch == the end offsets of the last batch. If the last batch id recorded in the OffsetLog equals to the last batch id recorded in the CommitLog + 1, it implies the last batch was failed and needed to rerun, so the start offsets of current batch == the start offsets of the last batch.

(2) Check each source whether new data is available and get the next available offset.

(3) Construct next batch, record the range of offsets that the next batch will process in the WAL OffsetLog.

(4) Call the getBatch method of the sources to return the DataFrame that represents the data between the start offsets and end offsets. Note that no physical data is actually read from the source at this moment, but instead the getBatch method only returns the data plan of the batch.

(5) Replace sources in the logical plan of this stream query with the data plan of the new batch to get the newBatchesPlan. Rewire the newBatchesPlan to use the new attributes that were returned by the sources to get the newAttributePlan.

(6) Create triggerLogicalPlan from newAttributePlan. Convert the newAttributePlan to WriteToMicroBatchDataSource if required.

(7) Create the IncrementalExecution for the current batch run with the triggerLogicalPlan as input. IncrementalExecution is a variant of the QueryExecution (please refer to my earlier blog post) for planning stream queries that allow incremental execution of a given logical plan. Just like the batch query planning, the stream query planning also needs to go through the same planning stages which will be covered in details in the subsequent blog posts in this series.

(8) Based on the generated execution plan, a new Dataset that logically represents the next batch is created.

(9) Add the next batch Dataset to the sink through addBatch method provided by the Sink interface.

(10) Log the batch execution into the CommitLog to record this successful batch run.