Interface SupportsRuntimeFiltering

All Superinterfaces:
Scan, SupportsRuntimeV2Filtering

@Experimental public interface SupportsRuntimeFiltering extends SupportsRuntimeV2Filtering
A mix-in interface for Scan. Data sources can implement this interface if they can filter initially planned InputPartitions using predicates Spark infers at runtime.

Note that Spark will push runtime filters only if they are beneficial.

Since:
3.2.0
  • Method Details

    • filterAttributes

      NamedReference[] filterAttributes()
      Returns attributes this scan can be filtered by at runtime.

      Spark will call filter(Filter[]) if it can derive a runtime predicate for any of the filter attributes.

      Each reference must resolve against the scan relation output when Spark builds it. Attributes pruned out of Scan.readSchema() fail to resolve.

      Specified by:
      filterAttributes in interface SupportsRuntimeV2Filtering
    • filter

      void filter(Filter[] filters)
      Filters this scan using runtime filters.

      The provided expressions must be interpreted as a set of filters that are ANDed together. Implementations may use the filters to prune initially planned InputPartitions.

      Spark currently tracks runtime-filter eligibility by root attribute. If filterAttributes() returns a nested reference, this method may receive a filter on another nested field under the same root. Implementations must inspect each filter and use only filters they can apply. Nested paths are encoded in a V1 Filter as dot-separated names, with each path part quoted as needed, such as parent.`child.with.dot`. A top-level column whose name contains a dot remains quoted, such as `parent.child`.

      If the scan also implements SupportsReportPartitioning, it must preserve the originally reported partitioning during runtime filtering. While applying runtime filters, the scan may detect that some InputPartitions have no matching data, in which case it can either replace the initially planned InputPartitions that have no matching data with empty InputPartitions, or report only a subset of the original partition values (omitting those with no data) via Batch.planInputPartitions(). The scan must not report new partition values that were not present in the original partitioning.

      Note that Spark will call Scan.toBatch() again after filtering the scan at runtime.

      Parameters:
      filters - data source filters used to filter the scan at runtime
    • filter

      default void filter(Predicate[] predicates)
      Description copied from interface: SupportsRuntimeV2Filtering
      Filters this scan using runtime predicates.

      The provided expressions must be interpreted as a set of predicates that are ANDed together. Implementations may use the predicates to prune initially planned InputPartitions.

      Spark currently tracks runtime-filter eligibility by root attribute. If SupportsRuntimeV2Filtering.filterAttributes() returns a nested reference, this method may receive a predicate on another nested field under the same root. Implementations must inspect each predicate and use only predicates they can apply.

      If the scan also implements SupportsReportPartitioning, it must preserve the originally reported partitioning during runtime filtering. While applying runtime predicates, the scan may detect that some InputPartitions have no matching data, in which case it can either replace the initially planned InputPartitions that have no matching data with empty InputPartitions, or report only a subset of the original partition values (omitting those with no data) via Batch.planInputPartitions(). The scan must not report new partition values that were not present in the original partitioning.

      This method may be called multiple times with additional predicates (e.g. PartitionPredicate) when SupportsRuntimeV2Filtering.supportsIterativePushdown() returns true. The implementation must accumulate state across all calls so that SupportsRuntimeV2Filtering.pushedPredicates() can return predicates from all of them.

      Independently of SupportsRuntimeV2Filtering.supportsIterativePushdown(), this method may also be called once per scan node when a plan holds several scan nodes sharing one Scan instance (e.g. the two branches of a group-based UPDATE). Implementations must accumulate state across those calls as well.

      Note that Spark will call Scan.toBatch() again after filtering the scan at runtime.

      Specified by:
      filter in interface SupportsRuntimeV2Filtering
      Parameters:
      predicates - data source V2 predicates used to filter the scan at runtime