Package org.apache.druid.query.groupby

Class GroupByQueryQueryToolChest

java.lang.Object
org.apache.druid.query.QueryToolChest<ResultRow,GroupByQuery>
org.apache.druid.query.groupby.GroupByQueryQueryToolChest
public class GroupByQueryQueryToolChest extends QueryToolChest<ResultRow,GroupByQuery>
Toolchest for GroupBy queries

  • Constructor Details

  • Method Details

    • mergeResults

      public QueryRunner<ResultRow> mergeResults(QueryRunner<ResultRow> runner)
      Description copied from class: QueryToolChest
      This method wraps a QueryRunner. The input QueryRunner, by contract, will provide a series of ResultType objects in time order (ascending or descending). This method should return a new QueryRunner that merges the stream of ordered ResultType objects.

      A default implementation constructs a ResultMergeQueryRunner which creates a CombiningSequence using the supplied QueryRunner with QueryToolChest.createResultComparator(Query) and QueryToolChest.createMergeFn(Query)} supplied by this toolchest.

      Generally speaking, the logic that exists in makePostComputeManipulatorFn should actually exist in this method. Additionally, if a query supports PostAggregations, this method should take steps to ensure that it computes PostAggregations a minimum number of times. This is most commonly achieved by computing the PostAgg results during merge and also rewriting the query such that it has the minimum number of PostAggs (most often zero).

      Overrides:
      mergeResults in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      runner - A QueryRunner that provides a series of ResultType objects in time order (ascending or descending)
      Returns:
      a QueryRunner that merges the stream of ordered ResultType objects

    • mergeResults

      public QueryRunner<ResultRow> mergeResults(QueryRunner<ResultRow> runner, boolean willMergeRunner)
      Description copied from class: QueryToolChest
      Like QueryToolChest.mergeResults(QueryRunner), but with an additional flag that indicates the type of runner that is passeed to the call. willMergeRunner specifies that the input runner to the mergeResults would be the one created by the corresponding QueryRunnerFactory.mergeRunners(java.util.concurrent.ExecutorService, java.lang.Iterable<org.apache.druid.query.QueryRunner<T>>). While it depends on the input runner, it is usually true since most of the time the same server is generating a runner that it wants to merge. The notable deviation from this norm is when the broker is accumulating the results from the data servers and needs to merge them together. In this case willMergeRunner is false. Currently, the sole consumer of this parameter is GroupByQueryQueryToolChest, where it is used to determine if the mergeResults is called with GroupByMergingQueryRunner to estimate the number of merge buffers required for the query to succeed. It is set false on the brokers, because they (mostly) fetch the results from the historicals, while the data servers set it to false (because they call this method with QueryRunnerFactory.mergeRunners(java.util.concurrent.ExecutorService, java.lang.Iterable<org.apache.druid.query.QueryRunner<T>>). By default, the willMergeRunners is ignored, and the QueryToolChest.mergeResults(QueryRunner) is called. For the toolchests that override this method must ensure that QueryToolChest.mergeResults(QueryRunner) delegates to it (else it will use the default implementation for QueryToolChest.mergeResults(QueryRunner)) which would be undesirable.
      Overrides:
      mergeResults in class QueryToolChest<ResultRow,GroupByQuery>

    • createMergeFn

      public BinaryOperator<ResultRow> createMergeFn(Query<ResultRow> query)
      Description copied from class: QueryToolChest
      Creates a merge function that is used to merge intermediate aggregates from historicals in broker. This merge function is used in the default ResultMergeQueryRunner provided by QueryToolChest.mergeResults(QueryRunner) and also used in ParallelMergeCombiningSequence by 'CachingClusteredClient' if it does not return null.

      Returning null from this function means that a query does not support result merging, at least via the mechanisms that utilize this function.

      Overrides:
      createMergeFn in class QueryToolChest<ResultRow,GroupByQuery>

    • createResultComparator

      public Comparator<ResultRow> createResultComparator(Query<ResultRow> query)
      Description copied from class: QueryToolChest
      Creates an ordering comparator that is used to order results. This comparator is used in the default ResultMergeQueryRunner provided by QueryToolChest.mergeResults(QueryRunner)
      Overrides:
      createResultComparator in class QueryToolChest<ResultRow,GroupByQuery>

    • isNestedQueryPushDown

      public static boolean isNestedQueryPushDown(GroupByQuery q)

    • makeMetrics

      public GroupByQueryMetrics makeMetrics(GroupByQuery query)
      Description copied from class: QueryToolChest
      Creates a QueryMetrics object that is used to generate metrics for this specific query type. This exists to allow for query-specific dimensions and metrics. That is, the ToolChest is expected to set some meaningful dimensions for metrics given this query type. Examples might be the topN threshold for a TopN query or the number of dimensions included for a groupBy query.

      QueryToolChests for query types in core (druid-processing) and public extensions (belonging to the Druid source tree) should use delegate this method to GenericQueryMetricsFactory.makeMetrics(Query) on an injected instance of GenericQueryMetricsFactory, as long as they don't need to emit custom dimensions and/or metrics.

      If some custom dimensions and/or metrics should be emitted for a query type, a plan described in "Making subinterfaces of QueryMetrics" section in QueryMetrics's class-level Javadocs should be followed.

      One way or another, this method should ensure that QueryMetrics.query(Query) is called with the given query passed on the created QueryMetrics object before returning.

      Specified by:
      makeMetrics in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      query - The query that is being processed
      Returns:
      A QueryMetrics that can be used to make metrics for the provided query

    • makePreComputeManipulatorFn

      public com.google.common.base.Function<ResultRow,ResultRow> makePreComputeManipulatorFn(GroupByQuery query, MetricManipulationFn fn)
      Description copied from class: QueryToolChest
      Creates a Function that can take in a ResultType and return a new ResultType having applied the MetricManipulatorFn to each of the metrics.

      This function's primary purpose is to help work around some challenges that exist around deserializing results across the wire. Specifically, different aggregators will generate different object types in a result set, if we wanted jackson to be able to deserialize these directly, we'd need to generate a response class for each query that jackson could use to deserialize things. That is not what we do. Instead, we have jackson deserialize Object instances and then use a MetricManipulatorFn to convert from those object instances to the actual object that the aggregator expects. As such, this would be more effectively named "makeObjectDeserializingFn".

      It is safe and acceptable for implementations of this method to first validate that the MetricManipulationFn is MetricManipulatorFns.DESERIALIZING_INSTANCE and throw an exception if it is not. If such an exception is ever thrown, it is indicative of a bug in the caller which should be fixed by not calling this method with anything other than the deserializing manipulator function.

      There are some implementations where this was also tasked with computing PostAggregators, but this is actually not a good place to compute those as this function can be called in a number of cases when PostAggs are not really meaningful to compute. Instead, PostAggs should be computed in the mergeResults call and the mergeResults implementation should take care to ensure that PostAggs are only computed the minimum number of times necessary.

      This function is called very early in the processing pipeline on the Broker.

      Specified by:
      makePreComputeManipulatorFn in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      query - The Query that is currently being processed
      fn - The function that should be applied to all metrics in the results
      Returns:
      A function that will apply the provided fn to all metrics in the input ResultType object

    • makePostComputeManipulatorFn

      public com.google.common.base.Function<ResultRow,ResultRow> makePostComputeManipulatorFn(GroupByQuery query, MetricManipulationFn fn)
      Description copied from class: QueryToolChest
      This manipulator functions primary purpose is to conduct finalization of aggregator values. It would be better named "makeFinalizingManipulatorFn", even that should really be done as part of QueryToolChest.mergeResults(org.apache.druid.query.QueryRunner<ResultType>) instead of with this separate method.

      It is safe and acceptable for implementations of this method to first validate that the MetricManipulationFn is either MetricManipulatorFns.FINALIZING_INSTANCE or MetricManipulatorFns.IDENTITY_INSTANCE and throw an exception if it is not. If such an exception is ever thrown, it is indicative of a bug in the caller which should be fixed by not calling this method with unsupported manipulator functions.

      Overrides:
      makePostComputeManipulatorFn in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      query - The Query that is currently being processed
      fn - The function that should be applied to all metrics in the results
      Returns:
      A function that will apply the provided fn to all metrics in the input ResultType object

    • getResultTypeReference

      public com.fasterxml.jackson.core.type.TypeReference<ResultRow> getResultTypeReference()
      Description copied from class: QueryToolChest
      Returns a TypeReference object that is just passed through to Jackson in order to deserialize the results of this type of query.
      Specified by:
      getResultTypeReference in class QueryToolChest<ResultRow,GroupByQuery>
      Returns:
      A TypeReference to indicate to Jackson what type of data will exist for this query

    • decorateObjectMapper

      public com.fasterxml.jackson.databind.ObjectMapper decorateObjectMapper(com.fasterxml.jackson.databind.ObjectMapper objectMapper, GroupByQuery query)
      Description copied from class: QueryToolChest
      Perform any per-query decoration of an ObjectMapper that enables it to read and write objects of the query's QueryToolChest. It is used by QueryResource on the write side, and DirectDruidClient on the read side.

      For most queries, this is a no-op, but it can be useful for query types that support more than one result serialization format. Queries that implement this method must not modify the provided ObjectMapper, but instead must return a copy.

      Jackson's default implementation of deserialization is usually optimised and this method should be overriden only if there is a functional requirement of so. The method must be benchmarked in isolation, without other portions of the query engine executing as modifying this method can alter the performance of queries where deserializing is a major portion of the execution.

      Overrides:
      decorateObjectMapper in class QueryToolChest<ResultRow,GroupByQuery>

    • preMergeQueryDecoration

      public QueryRunner<ResultRow> preMergeQueryDecoration(QueryRunner<ResultRow> runner)
      Description copied from class: QueryToolChest
      Wraps a QueryRunner. The input QueryRunner is the QueryRunner as it exists *before* being passed to mergeResults().

      In fact, the return value of this method is always passed to mergeResults, so it is equivalent to just implement this functionality as extra decoration on the QueryRunner during mergeResults().

      In the interests of potentially simplifying these interfaces, the recommendation is to actually not override this method and instead apply anything that might be needed here in the mergeResults() call.

      Overrides:
      preMergeQueryDecoration in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      runner - The runner to be wrapped
      Returns:
      The wrapped runner

    • getCacheStrategy

      @Nullable public CacheStrategy<ResultRow,Object,GroupByQuery> getCacheStrategy(GroupByQuery query)
      Description copied from class: QueryToolChest
      Like QueryToolChest.getCacheStrategy(Query, ObjectMapper) but the caller doesn't supply the object mapper for deserializing and converting the cached data to desired type. It's upto the individual implementations to decide the appropriate action in that case. It can either throw an exception outright or decide if the query requires the object mapper for proper downstream processing and work with the generic java types if not.

      Overrides:
      getCacheStrategy in class QueryToolChest<ResultRow,GroupByQuery>

    • getCacheStrategy

      public CacheStrategy<ResultRow,Object,GroupByQuery> getCacheStrategy(GroupByQuery query, @Nullable com.fasterxml.jackson.databind.ObjectMapper mapper)
      Description copied from class: QueryToolChest
      Returns a CacheStrategy to be used to load data into the cache and remove it from the cache.

      This is optional. If it returns null, caching is effectively disabled for the query.

      Overrides:
      getCacheStrategy in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      query - The query whose results might be cached
      mapper - Object mapper to convert the deserialized generic java objects to desired types. It can be nullable to preserve backward compatibility.
      Returns:
      A CacheStrategy that can be used to populate and read from the Cache

    • canPerformSubquery

      public boolean canPerformSubquery(Query<?> subquery)
      Description copied from class: QueryToolChest
      Returns whether this toolchest is able to handle the provided subquery.

      When this method returns true, the core query stack will pass subquery datasources over to the toolchest and will assume they are properly handled.

      When this method returns false, the core query stack will throw an error if subqueries are present. In the future, instead of throwing an error, the core query stack will handle the subqueries on its own.

      Overrides:
      canPerformSubquery in class QueryToolChest<ResultRow,GroupByQuery>

    • resultArraySignature

      public RowSignature resultArraySignature(GroupByQuery query)
      Description copied from class: QueryToolChest
      Returns a RowSignature for the arrays returned by QueryToolChest.resultsAsArrays(QueryType, org.apache.druid.java.util.common.guava.Sequence<ResultType>). The returned signature will be the same length as each array returned by QueryToolChest.resultsAsArrays(QueryType, org.apache.druid.java.util.common.guava.Sequence<ResultType>).
      Overrides:
      resultArraySignature in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      query - same query passed to QueryToolChest.resultsAsArrays(QueryType, org.apache.druid.java.util.common.guava.Sequence<ResultType>)
      Returns:
      row signature

    • resultsAsArrays

      public Sequence<Object[]> resultsAsArrays(GroupByQuery query, Sequence<ResultRow> resultSequence)
      Description copied from class: QueryToolChest
      Converts a sequence of this query's ResultType into arrays. The array signature is given by QueryToolChest.resultArraySignature(QueryType). This functionality is useful because it allows higher-level processors to operate on the results of any query in a consistent way. This is useful for the SQL layer and for any algorithm that might operate on the results of an inner query.

      Not all query types support this method. They will throw UnsupportedOperationException, and they cannot be used by the SQL layer or by generic higher-level algorithms.

      Some query types return less information after translating their results into arrays, especially in situations where there is no clear way to translate fully rich results into flat arrays. For example, the scan query does not include the segmentId in its array-based results, because it could potentially conflict with a 'segmentId' field in the actual datasource being scanned.

      It is possible that there will be multiple arrays returned for a single result object. For example, in the topN query, each TopNResultValue will generate a separate array for each of its values.

      By convention, the array form should include the __time column, if present, as a long (milliseconds since epoch).

      Overrides:
      resultsAsArrays in class QueryToolChest<ResultRow,GroupByQuery>
      resultSequence - results of the form returned by QueryToolChest.mergeResults(org.apache.druid.query.QueryRunner<ResultType>)
      Returns:
      results in array form

    • resultsAsFrames

      public Optional<Sequence<FrameSignaturePair>> resultsAsFrames(GroupByQuery query, Sequence<ResultRow> resultSequence, MemoryAllocatorFactory memoryAllocatorFactory, boolean useNestedForUnknownTypes)
      This returns a single frame containing the results of the group by query.
      Overrides:
      resultsAsFrames in class QueryToolChest<ResultRow,GroupByQuery>
      Parameters:
      query - Query being executed by the toolchest. Used to determine the rowSignature of the Frames
      resultSequence - results of the form returned by QueryToolChest.mergeResults(QueryRunner)
      useNestedForUnknownTypes - true if the unknown types in the results can be serded using complex types