Package org.apache.iceberg

Class SparkDistributedDataScan

java.lang.Object
org.apache.iceberg.SnapshotScan<BatchScan,ScanTask,ScanTaskGroup<ScanTask>>
org.apache.iceberg.SparkDistributedDataScan
All Implemented Interfaces:
BatchScan, Scan<BatchScan,ScanTask,ScanTaskGroup<ScanTask>>
public class SparkDistributedDataScan extends SnapshotScan<BatchScan,ScanTask,ScanTaskGroup<ScanTask>>
A batch data scan that can utilize Spark cluster resources for planning.

This scan remotely filters manifests, fetching only the relevant data and delete files to the driver. The delete file assignment is done locally after the remote filtering step. Such approach is beneficial if the remote parallelism is much higher than the number of driver cores.

This scan is best suited for queries with selective filters on lower/upper bounds across all partitions, or against poorly clustered metadata. This allows job planning to benefit from highly concurrent remote filtering while not incurring high serialization and data transfer costs. This class is also useful for full table scans over large tables but the cost of bringing data and delete file details to the driver may become noticeable. Make sure to follow the performance tips below in such cases.

Ensure the filtered metadata size doesn't exceed the driver's max result size. For large table scans, consider increasing `spark.driver.maxResultSize` to avoid job failures.

Performance tips:

  • Enable Kryo serialization (`spark.serializer`)
  • Increase the number of driver cores (`spark.driver.cores`)
  • Tune the number of threads used to fetch task results (`spark.resultGetter.threads`)

  • Field Details

    • SCAN_COLUMNS

      protected static final List<String> SCAN_COLUMNS

    • SCAN_WITH_STATS_COLUMNS

      protected static final List<String> SCAN_WITH_STATS_COLUMNS

    • DELETE_SCAN_COLUMNS

      protected static final List<String> DELETE_SCAN_COLUMNS

    • DELETE_SCAN_WITH_STATS_COLUMNS

      protected static final List<String> DELETE_SCAN_WITH_STATS_COLUMNS

    • PLAN_SCANS_WITH_WORKER_POOL

      protected static final boolean PLAN_SCANS_WITH_WORKER_POOL

  • Constructor Details

    • SparkDistributedDataScan

      public SparkDistributedDataScan(org.apache.spark.sql.SparkSession spark, Table table, SparkReadConf readConf)

  • Method Details

    • newRefinedScan

      protected BatchScan newRefinedScan(Table newTable, Schema newSchema, org.apache.iceberg.TableScanContext newContext)

    • remoteParallelism

      protected int remoteParallelism()
      Returns the cluster parallelism.

      This value indicates the maximum number of manifests that can be processed concurrently by the cluster. Implementations should take into account both the currently available processing slots and potential dynamic allocation, if applicable.

      The remote parallelism is compared against the size of the thread pool available locally to determine the feasibility of remote planning. This value is ignored if the planning mode is set explicitly as local or distributed.

    • dataPlanningMode

      protected PlanningMode dataPlanningMode()
      Returns which planning mode to use for data.

    • shouldCopyRemotelyPlannedDataFiles

      protected boolean shouldCopyRemotelyPlannedDataFiles()
      Controls whether defensive copies are created for remotely planned data files.

      By default, this class creates defensive copies for each data file that is planned remotely, assuming the provided iterable can be lazy and may reuse objects. If unnecessary and data file objects can be safely added into a collection, implementations can override this behavior.

    • planDataRemotely

      protected Iterable<CloseableIterable<DataFile>> planDataRemotely(List<ManifestFile> dataManifests, boolean withColumnStats)
      Plans data remotely.

      Implementations are encouraged to return groups of matching data files, enabling this class to process multiple groups concurrently to speed up the remaining work. This is particularly useful when dealing with equality deletes, as delete index lookups with such delete files require comparing bounds and typically benefit from parallelization.

      If the result iterable reuses objects, shouldCopyRemotelyPlannedDataFiles() must return true.

      The input data manifests have been already filtered to include only potential matches based on the scan filter. Implementations are expected to further filter these manifests and only return files that may hold data matching the scan filter.

      Parameters:
      dataManifests - data manifests that may contain files matching the scan filter
      withColumnStats - a flag whether to load column stats
      Returns:
      groups of data files planned remotely

    • deletePlanningMode

      protected PlanningMode deletePlanningMode()
      Returns which planning mode to use for deletes.

    • planDeletesRemotely

      protected org.apache.iceberg.DeleteFileIndex planDeletesRemotely(List<ManifestFile> deleteManifests)
      Plans deletes remotely.

      The input delete manifests have been already filtered to include only potential matches based on the scan filter. Implementations are expected to further filter these manifests and return files that may hold deletes matching the scan filter.

      Parameters:
      deleteManifests - delete manifests that may contain files matching the scan filter
      Returns:
      a delete file index planned remotely

    • doPlanFiles

      protected CloseableIterable<ScanTask> doPlanFiles()
      Specified by:
      doPlanFiles in class SnapshotScan<BatchScan,ScanTask,ScanTaskGroup<ScanTask>>

    • planTasks

      public CloseableIterable<ScanTaskGroup<ScanTask>> planTasks()
      Description copied from interface: Scan
      Plan balanced task groups for this scan by splitting large and combining small tasks.

      Task groups created by this method may read partial input files, multiple input files or both.

      Specified by:
      planTasks in interface Scan<BatchScan,ScanTask,ScanTaskGroup<ScanTask>>
      Returns:
      an Iterable of balanced task groups required by this scan

    • useSnapshotSchema

      protected boolean useSnapshotSchema()
      Overrides:
      useSnapshotSchema in class SnapshotScan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>

    • newManifestGroup

      protected org.apache.iceberg.ManifestGroup newManifestGroup(List<ManifestFile> dataManifests, List<ManifestFile> deleteManifests)

    • newManifestGroup

      protected org.apache.iceberg.ManifestGroup newManifestGroup(List<ManifestFile> dataManifests, boolean withColumnStats)

    • newManifestGroup

      protected org.apache.iceberg.ManifestGroup newManifestGroup(List<ManifestFile> dataManifests, List<ManifestFile> deleteManifests, boolean withColumnStats)

    • table

      public Table table()

    • io

      protected FileIO io()

    • tableSchema

      protected Schema tableSchema()

    • context

      protected org.apache.iceberg.TableScanContext context()

    • options

      protected Map<String,String> options()

    • scanColumns

      protected List<String> scanColumns()

    • shouldReturnColumnStats

      protected boolean shouldReturnColumnStats()

    • columnsToKeepStats

      protected Set<Integer> columnsToKeepStats()

    • shouldIgnoreResiduals

      protected boolean shouldIgnoreResiduals()

    • residualFilter

      protected Expression residualFilter()

    • shouldPlanWithExecutor

      protected boolean shouldPlanWithExecutor()

    • planExecutor

      protected ExecutorService planExecutor()

    • option

      public BatchScan option(String property, String value)
      Description copied from interface: Scan
      Create a new scan from this scan's configuration that will override the Table's behavior based on the incoming pair. Unknown properties will be ignored.
      Specified by:
      option in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Parameters:
      property - name of the table property to be overridden
      value - value to override with
      Returns:
      a new scan based on this with overridden behavior

    • project

      public BatchScan project(Schema projectedSchema)
      Description copied from interface: Scan
      Create a new scan from this with the schema as its projection.
      Specified by:
      project in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Parameters:
      projectedSchema - a projection schema
      Returns:
      a new scan based on this with the given projection

    • caseSensitive

      public BatchScan caseSensitive(boolean caseSensitive)
      Description copied from interface: Scan
      Create a new scan from this that, if data columns where selected via Scan.select(java.util.Collection), controls whether the match to the schema will be done with case sensitivity. Default is true.
      Specified by:
      caseSensitive in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Returns:
      a new scan based on this with case sensitivity as stated

    • isCaseSensitive

      public boolean isCaseSensitive()
      Description copied from interface: Scan
      Returns whether this scan is case-sensitive with respect to column names.
      Specified by:
      isCaseSensitive in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Returns:
      true if case-sensitive, false otherwise.

    • includeColumnStats

      public BatchScan includeColumnStats()
      Description copied from interface: Scan
      Create a new scan from this that loads the column stats with each data file.

      Column stats include: value count, null value count, lower bounds, and upper bounds.

      Specified by:
      includeColumnStats in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Returns:
      a new scan based on this that loads column stats.

    • includeColumnStats

      public BatchScan includeColumnStats(Collection<String> requestedColumns)
      Description copied from interface: Scan
      Create a new scan from this that loads the column stats for the specific columns with each data file.

      Column stats include: value count, null value count, lower bounds, and upper bounds.

      Specified by:
      includeColumnStats in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Parameters:
      requestedColumns - column names for which to keep the stats.
      Returns:
      a new scan based on this that loads column stats for specific columns.

    • select

      public BatchScan select(Collection<String> columns)
      Description copied from interface: Scan
      Create a new scan from this that will read the given data columns. This produces an expected schema that includes all fields that are either selected or used by this scan's filter expression.
      Specified by:
      select in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Parameters:
      columns - column names from the table's schema
      Returns:
      a new scan based on this with the given projection columns

    • filter

      public BatchScan filter(Expression expr)
      Description copied from interface: Scan
      Create a new scan from the results of this filtered by the Expression.
      Specified by:
      filter in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Parameters:
      expr - a filter expression
      Returns:
      a new scan based on this with results filtered by the expression

    • filter

      public Expression filter()
      Description copied from interface: Scan
      Returns this scan's filter Expression.
      Specified by:
      filter in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Returns:
      this scan's filter expression

    • ignoreResiduals

      public BatchScan ignoreResiduals()
      Description copied from interface: Scan
      Create a new scan from this that applies data filtering to files but not to rows in those files.
      Specified by:
      ignoreResiduals in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Returns:
      a new scan based on this that does not filter rows in files.

    • planWith

      public BatchScan planWith(ExecutorService executorService)
      Description copied from interface: Scan
      Create a new scan to use a particular executor to plan. The default worker pool will be used by default.
      Specified by:
      planWith in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Parameters:
      executorService - the provided executor
      Returns:
      a table scan that uses the provided executor to access manifests

    • schema

      public Schema schema()
      Description copied from interface: Scan
      Returns this scan's projection Schema.

      If the projection schema was set directly using Scan.project(Schema), returns that schema.

      If the projection schema was set by calling Scan.select(Collection), returns a projection schema that includes the selected data fields and any fields used in the filter expression.

      Specified by:
      schema in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>
      Returns:
      this scan's projection schema

    • targetSplitSize

      public long targetSplitSize()
      Description copied from interface: Scan
      Returns the target split size for this scan.
      Specified by:
      targetSplitSize in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>

    • splitLookback

      public int splitLookback()
      Description copied from interface: Scan
      Returns the split lookback for this scan.
      Specified by:
      splitLookback in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>

    • splitOpenFileCost

      public long splitOpenFileCost()
      Description copied from interface: Scan
      Returns the split open file cost for this scan.
      Specified by:
      splitOpenFileCost in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>

    • metricsReporter

      public BatchScan metricsReporter(MetricsReporter reporter)
      Description copied from interface: Scan
      Create a new scan that will report scan metrics to the provided reporter in addition to reporters maintained by the scan.
      Specified by:
      metricsReporter in interface Scan<ThisT,T extends ScanTask,G extends ScanTaskGroup<T>>