DeltaSource¶

DeltaSource is the Source (Spark Structured Streaming) of the delta data source for streaming queries.

Creating Instance¶

DeltaSource takes the following to be created:

SparkSession (Spark SQL)
DeltaLog
DeltaOptions
Filters (default: empty)

DeltaSource is created when:

DeltaDataSource is requested for a streaming source

Demo¶

val q = spark
  .readStream               // Creating a streaming query
  .format("delta")          // Using delta data source
  .load("/tmp/delta/users") // Over data in a delta table
  .writeStream
  .format("memory")
  .option("queryName", "demo")
  .start
import org.apache.spark.sql.execution.streaming.{MicroBatchExecution, StreamingQueryWrapper}
val plan = q.asInstanceOf[StreamingQueryWrapper]
  .streamingQuery
  .asInstanceOf[MicroBatchExecution]
  .logicalPlan
import org.apache.spark.sql.execution.streaming.StreamingExecutionRelation
val relation = plan.collect { case r: StreamingExecutionRelation => r }.head

import org.apache.spark.sql.delta.sources.DeltaSource
assert(relation.source.asInstanceOf[DeltaSource])

scala> println(relation.source)
DeltaSource[file:/tmp/delta/users]

Latest Available Offset¶

latestOffset(
  startOffset: streaming.Offset,
  limit: ReadLimit): streaming.Offset

latestOffset is part of the SupportsAdmissionControl (Spark Structured Streaming) abstraction.

latestOffset determines the latest offset (currentOffset) based on whether the previousOffset internal registry is initialized or not.

latestOffset prints out the following DEBUG message to the logs (using the previousOffset internal registry).

previousOffset -> currentOffset: [previousOffset] -> [currentOffset]

In the end, latestOffset returns the previousOffset if defined or null.

No previousOffset¶

For no previousOffset, getOffset retrieves the starting offset (with a new AdmissionLimits for the given ReadLimit).

previousOffset Available¶

When the previousOffset is defined (which is when the DeltaSource is requested for another micro-batch), latestOffset gets the changes as an indexed AddFiles (with the previousOffset and a new AdmissionLimits for the given ReadLimit).

latestOffset takes the last AddFile if available.

With no AddFile, latestOffset returns the previousOffset.

With the previousOffset and the last indexed AddFile both available, latestOffset creates a new DeltaSourceOffset for the version, index, and isLast flag from the last indexed AddFile.

Note

isStartingVersion local value is enabled (true) when the following holds:

Version of the last indexed AddFile is equal to the reservoirVersion of the previous ending offset
isStartingVersion flag of the previous ending offset is enabled (true)

getStartingOffset¶

getStartingOffset(
  limits: Option[AdmissionLimits]): Option[Offset]

getStartingOffset...FIXME (review me)

getStartingOffset requests the DeltaLog for the version of the delta table (by requesting for the current state (snapshot) and then for the version).

getStartingOffset takes the last file from the files added (with rate limit) for the version of the delta table, -1L as the fromIndex, and the isStartingVersion flag enabled (true).

getStartingOffset returns a new DeltaSourceOffset for the tableId, the version and the index of the last file added, and whether the last file added is the last file of its version.

getStartingOffset returns None (offset not available) when either happens:

the version of the delta table is negative (below 0)
no files were added in the version

getStartingOffset throws an AssertionError when the version of the last file added is smaller than the delta table's version:

assertion failed: getChangesWithRateLimit returns an invalid version: [v] (expected: >= [version])

getChangesWithRateLimit¶

getChangesWithRateLimit(
  fromVersion: Long,
  fromIndex: Long,
  isStartingVersion: Boolean): Iterator[IndexedFile]

getChangesWithRateLimit gets the changes (as indexed AddFiles) for the given fromVersion, fromIndex, and isStartingVersion flag.

getOffset¶

getOffset: Option[Offset]

getOffset is part of the Source (Spark Structured Streaming) abstraction.

getOffset has been replaced by the newer latestOffset and so throws an UnsupportedOperationException when called:

latestOffset(Offset, ReadLimit) should be called instead of this method

Requesting Micro-Batch DataFrame¶

getBatch(
  start: Option[Offset],
  end: Offset): DataFrame

getBatch is part of the Source (Spark Structured Streaming) abstraction.

getBatch creates an DeltaSourceOffset for the tableId (aka reservoirId) and the given end offset.

getBatch gets the changes...FIXME

getBatch prints out the following DEBUG message to the logs:

start: [start] end: [end] [addFiles]

In the end, getBatch requests the DeltaLog to createDataFrame (for the current snapshot of the DeltaLog, addFiles and isStreaming flag on).

Snapshot Management¶

DeltaSource uses internal registries for the DeltaSourceSnapshot and the version to avoid requesting the DeltaLog for getSnapshotAt.

Snapshot¶

DeltaSource uses initialState internal registry for the DeltaSourceSnapshot of the state of the delta table at the initialStateVersion.

DeltaSourceSnapshot is used for AddFiles of the delta table at a given version.

Initially uninitialized (null).

DeltaSourceSnapshot is created (initialized) when uninitialized or the version requested is different from the current one.

DeltaSourceSnapshot is closed and dereferenced (null) when DeltaSource is requested to cleanUpSnapshotResources (due to version change, another micro-batch or stop).

Version¶

DeltaSource uses initialStateVersion internal registry to keep track of the version of DeltaSourceSnapshot (when requested for AddFiles of the delta table at a given version).

Changes (alongside the initialState) to the version requested when DeltaSource is requested for the snapshot at a given version (only when the versions are different)

Used when:

DeltaSource is requested for AddFiles of the delta table at a given version and to cleanUpSnapshotResources (and unpersist the current snapshot)

Stopping¶

stop(): Unit

stop is part of the Source (Spark Structured Streaming) abstraction.

stop simply cleanUpSnapshotResources.

Previous Offset¶

Ending DeltaSourceOffset of the latest micro-batch

Starts uninitialized (null).

Used when DeltaSource is requested for the latest available offset.

AddFiles of Delta Table at Given Version¶

getSnapshotAt(
  version: Long): Iterator[IndexedFile]

getSnapshotAt requests the DeltaSourceSnapshot for the data files (as indexed AddFiles).

In case the DeltaSourceSnapshot hasn't been initialized yet (null) or the requested version is different from the initialStateVersion, getSnapshotAt does the following:

cleanUpSnapshotResources
Requests the DeltaLog for the state (snapshot) of the delta table at the version
Creates a new DeltaSourceSnapshot for the state (snapshot) as the current DeltaSourceSnapshot
Changes the initialStateVersion internal registry to the requested version

getSnapshotAt is used when:

DeltaSource is requested to getChanges (with isStartingVersion flag enabled)

getChanges¶

getChanges(
  fromVersion: Long,
  fromIndex: Long,
  isStartingVersion: Boolean): Iterator[IndexedFile]

getChanges branches based on isStartingVersion flag (enabled or not):

For isStartingVersion flag enabled (true), getChanges gets the state (snapshot) for the given fromVersion followed by (filtered out) indexed AddFiles for the next version after the given fromVersion
For isStartingVersion flag disabled (false), getChanges simply gives (filtered out) indexed AddFiles for the given fromVersion

Note

isStartingVersion flag simply adds the state (snapshot) before (filtered out) indexed AddFiles when enabled (true).

isStartingVersion flag is enabled when DeltaSource is requested for the following:

Micro-batch with data between start and end offsets and the start offset is not given or is for the starting version
Latest available offset with no end offset of the latest micro-batch or the end offset of the latest micro-batch for the starting version

In the end, getChanges filters out (excludes) indexed AddFiles that are not with the version later than the given fromVersion or the index greater than the given fromIndex.

getChanges is used when:

DeltaSource is requested for the latest available offset (when requested for the files added (with rate limit)) and getBatch

filterAndIndexDeltaLogs¶

filterAndIndexDeltaLogs(
  startVersion: Long): Iterator[IndexedFile]

filterAndIndexDeltaLogs...FIXME

verifyStreamHygieneAndFilterAddFiles¶

verifyStreamHygieneAndFilterAddFiles(
  actions: Seq[Action],
  version: Long): Seq[Action]

verifyStreamHygieneAndFilterAddFiles...FIXME

cleanUpSnapshotResources¶

cleanUpSnapshotResources(): Unit

cleanUpSnapshotResources does the following when the initial DeltaSourceSnapshot internal registry is not empty:

Requests the DeltaSourceSnapshot to close (with the unpersistSnapshot flag based on whether the initialStateVersion is earlier than the snapshot version)
Dereferences (nullifies) the DeltaSourceSnapshot

Otherwise, cleanUpSnapshotResources does nothing.

cleanUpSnapshotResources is used when:

DeltaSource is requested to getSnapshotAt, getBatch and stop

ReadLimit¶

getDefaultReadLimit: ReadLimit

getDefaultReadLimit is part of the SupportsAdmissionControl (Spark Structured Streaming) abstraction.

getDefaultReadLimit creates a AdmissionLimits and requests it for a corresponding ReadLimit.

Retrieving Last Element From Iterator¶

iteratorLast[T](
  iter: Iterator[T]): Option[T]

iteratorLast simply returns the last element of the given Iterator (Scala) or None.

iteratorLast is used when:

DeltaSource is requested to getStartingOffset and getOffset

Logging¶

Enable ALL logging level for org.apache.spark.sql.delta.sources.DeltaSource logger to see what happens inside.

Add the following line to conf/log4j.properties:

log4j.logger.org.apache.spark.sql.delta.sources.DeltaSource=ALL

Refer to Logging.

Last update: 2021-03-26