/* * Licensed to the Apache Software Foundation (ASF) under one * or more contributor license agreements. See the NOTICE file * distributed with this work for additional information * regarding copyright ownership. The ASF licenses this file * to you under the Apache License, Version 2.0 (the * "License"); you may not use this file except in compliance * with the License. You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ package org.apache.flink.compiler; import java.util.ArrayDeque; import java.util.ArrayList; import java.util.Deque; import java.util.HashMap; import java.util.HashSet; import java.util.Iterator; import java.util.List; import java.util.Map; import java.util.Set; import org.slf4j.Logger; import org.slf4j.LoggerFactory; import org.apache.flink.api.common.InvalidProgramException; import org.apache.flink.api.common.Plan; import org.apache.flink.api.common.operators.Operator; import org.apache.flink.api.common.operators.Union; import org.apache.flink.api.common.operators.base.BulkIterationBase; import org.apache.flink.api.common.operators.base.CoGroupOperatorBase; import org.apache.flink.api.common.operators.base.CrossOperatorBase; import org.apache.flink.api.common.operators.base.DeltaIterationBase; import org.apache.flink.api.common.operators.base.FilterOperatorBase; import org.apache.flink.api.common.operators.base.FlatMapOperatorBase; import org.apache.flink.api.common.operators.base.GenericDataSinkBase; import org.apache.flink.api.common.operators.base.GenericDataSourceBase; import org.apache.flink.api.common.operators.base.GroupReduceOperatorBase; import org.apache.flink.api.common.operators.base.JoinOperatorBase; import org.apache.flink.api.common.operators.base.MapOperatorBase; import org.apache.flink.api.common.operators.base.MapPartitionOperatorBase; import org.apache.flink.api.common.operators.base.PartitionOperatorBase; import org.apache.flink.api.common.operators.base.ReduceOperatorBase; import org.apache.flink.api.common.operators.base.BulkIterationBase.PartialSolutionPlaceHolder; import org.apache.flink.api.common.operators.base.DeltaIterationBase.SolutionSetPlaceHolder; import org.apache.flink.api.common.operators.base.DeltaIterationBase.WorksetPlaceHolder; import org.apache.flink.compiler.costs.CostEstimator; import org.apache.flink.compiler.costs.DefaultCostEstimator; import org.apache.flink.compiler.dag.BinaryUnionNode; import org.apache.flink.compiler.dag.BulkIterationNode; import org.apache.flink.compiler.dag.BulkPartialSolutionNode; import org.apache.flink.compiler.dag.CoGroupNode; import org.apache.flink.compiler.dag.CollectorMapNode; import org.apache.flink.compiler.dag.CrossNode; import org.apache.flink.compiler.dag.DataSinkNode; import org.apache.flink.compiler.dag.DataSourceNode; import org.apache.flink.compiler.dag.FilterNode; import org.apache.flink.compiler.dag.FlatMapNode; import org.apache.flink.compiler.dag.GroupReduceNode; import org.apache.flink.compiler.dag.IterationNode; import org.apache.flink.compiler.dag.MapNode; import org.apache.flink.compiler.dag.MapPartitionNode; import org.apache.flink.compiler.dag.MatchNode; import org.apache.flink.compiler.dag.OptimizerNode; import org.apache.flink.compiler.dag.PactConnection; import org.apache.flink.compiler.dag.PartitionNode; import org.apache.flink.compiler.dag.ReduceNode; import org.apache.flink.compiler.dag.SinkJoiner; import org.apache.flink.compiler.dag.SolutionSetNode; import org.apache.flink.compiler.dag.TempMode; import org.apache.flink.compiler.dag.WorksetIterationNode; import org.apache.flink.compiler.dag.WorksetNode; import org.apache.flink.compiler.deadlockdetect.DeadlockPreventer; import org.apache.flink.compiler.plan.BinaryUnionPlanNode; import org.apache.flink.compiler.plan.BulkIterationPlanNode; import org.apache.flink.compiler.plan.BulkPartialSolutionPlanNode; import org.apache.flink.compiler.plan.Channel; import org.apache.flink.compiler.plan.IterationPlanNode; import org.apache.flink.compiler.plan.NAryUnionPlanNode; import org.apache.flink.compiler.plan.OptimizedPlan; import org.apache.flink.compiler.plan.PlanNode; import org.apache.flink.compiler.plan.SinkJoinerPlanNode; import org.apache.flink.compiler.plan.SinkPlanNode; import org.apache.flink.compiler.plan.SolutionSetPlanNode; import org.apache.flink.compiler.plan.SourcePlanNode; import org.apache.flink.compiler.plan.WorksetIterationPlanNode; import org.apache.flink.compiler.plan.WorksetPlanNode; import org.apache.flink.compiler.postpass.OptimizerPostPass; import org.apache.flink.configuration.ConfigConstants; import org.apache.flink.configuration.GlobalConfiguration; import org.apache.flink.runtime.operators.shipping.ShipStrategyType; import org.apache.flink.runtime.operators.util.LocalStrategy; import org.apache.flink.util.InstantiationUtil; import org.apache.flink.util.Visitor; /** * The optimizer that takes the user specified program plan and creates an optimized plan that contains * exact descriptions about how the physical execution will take place. It first translates the user * program into an internal optimizer representation and then chooses between different alternatives * for shipping strategies and local strategies. *

* The basic principle is taken from optimizer works in systems such as Volcano/Cascades and Selinger/System-R/DB2. The * optimizer walks from the sinks down, generating interesting properties, and ascends from the sources generating * alternative plans, pruning against the interesting properties. *

* The optimizer also assigns the memory to the individual tasks. This is currently done in a very simple fashion: All * sub-tasks that need memory (e.g. reduce or join) are given an equal share of memory. */ public class PactCompiler { // ------------------------------------------------------------------------ // Constants // ------------------------------------------------------------------------ /** * Compiler hint key for the input channel's shipping strategy. This String is a key to the operator's stub * parameters. The corresponding value tells the compiler which shipping strategy to use for the input channel. * If the operator has two input channels, the shipping strategy is applied to both input channels. */ public static final String HINT_SHIP_STRATEGY = "INPUT_SHIP_STRATEGY"; /** * Compiler hint key for the first input channel's shipping strategy. This String is a key to * the operator's stub parameters. The corresponding value tells the compiler which shipping strategy * to use for the first input channel. Only applicable to operators with two inputs. */ public static final String HINT_SHIP_STRATEGY_FIRST_INPUT = "INPUT_LEFT_SHIP_STRATEGY"; /** * Compiler hint key for the second input channel's shipping strategy. This String is a key to * the operator's stub parameters. The corresponding value tells the compiler which shipping strategy * to use for the second input channel. Only applicable to operators with two inputs. */ public static final String HINT_SHIP_STRATEGY_SECOND_INPUT = "INPUT_RIGHT_SHIP_STRATEGY"; /** * Value for the shipping strategy compiler hint that enforces a Forward strategy on the * input channel, i.e. no redistribution of any kind. * * @see #HINT_SHIP_STRATEGY * @see #HINT_SHIP_STRATEGY_FIRST_INPUT * @see #HINT_SHIP_STRATEGY_SECOND_INPUT */ public static final String HINT_SHIP_STRATEGY_FORWARD = "SHIP_FORWARD"; /** * Value for the shipping strategy compiler hint that enforces a random repartition strategy. * * @see #HINT_SHIP_STRATEGY * @see #HINT_SHIP_STRATEGY_FIRST_INPUT * @see #HINT_SHIP_STRATEGY_SECOND_INPUT */ public static final String HINT_SHIP_STRATEGY_REPARTITION= "SHIP_REPARTITION"; /** * Value for the shipping strategy compiler hint that enforces a hash-partition strategy. * * @see #HINT_SHIP_STRATEGY * @see #HINT_SHIP_STRATEGY_FIRST_INPUT * @see #HINT_SHIP_STRATEGY_SECOND_INPUT */ public static final String HINT_SHIP_STRATEGY_REPARTITION_HASH = "SHIP_REPARTITION_HASH"; /** * Value for the shipping strategy compiler hint that enforces a range-partition strategy. * * @see #HINT_SHIP_STRATEGY * @see #HINT_SHIP_STRATEGY_FIRST_INPUT * @see #HINT_SHIP_STRATEGY_SECOND_INPUT */ public static final String HINT_SHIP_STRATEGY_REPARTITION_RANGE = "SHIP_REPARTITION_RANGE"; /** * Value for the shipping strategy compiler hint that enforces a broadcast strategy on the * input channel. * * @see #HINT_SHIP_STRATEGY * @see #HINT_SHIP_STRATEGY_FIRST_INPUT * @see #HINT_SHIP_STRATEGY_SECOND_INPUT */ public static final String HINT_SHIP_STRATEGY_BROADCAST = "SHIP_BROADCAST"; /** * Compiler hint key for the operator's local strategy. This String is a key to the operator's stub * parameters. The corresponding value tells the compiler which local strategy to use to process the * data inside one partition. *

* This hint is ignored by operators that do not have a local strategy (such as Map), or by operators that * have no choice in their local strategy (such as Cross). */ public static final String HINT_LOCAL_STRATEGY = "LOCAL_STRATEGY"; /** * Value for the local strategy compiler hint that enforces a sort based local strategy. * For example, a Reduce operator will sort the data to group it. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_SORT = "LOCAL_STRATEGY_SORT"; /** * Value for the local strategy compiler hint that enforces a sort based local strategy. * During sorting a combine method is repeatedly applied to reduce the data volume. * For example, a Reduce operator will sort the data to group it. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_COMBINING_SORT = "LOCAL_STRATEGY_COMBINING_SORT"; /** * Value for the local strategy compiler hint that enforces a sort merge based local strategy on both * inputs with subsequent merging of inputs. * For example, a Match or CoGroup operator will use a sort-merge strategy to find pairs * of matching keys. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_SORT_BOTH_MERGE = "LOCAL_STRATEGY_SORT_BOTH_MERGE"; /** * Value for the local strategy compiler hint that enforces a sort merge based local strategy. * The the first input is sorted, the second input is assumed to be sorted. After sorting both inputs are merged. * For example, a Match or CoGroup operator will use a sort-merge strategy to find pairs * of matching keys. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_SORT_FIRST_MERGE = "LOCAL_STRATEGY_SORT_FIRST_MERGE"; /** * Value for the local strategy compiler hint that enforces a sort merge based local strategy. * The the second input is sorted, the first input is assumed to be sorted. After sorting both inputs are merged. * For example, a Match or CoGroup operator will use a sort-merge strategy to find pairs * of matching keys. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_SORT_SECOND_MERGE = "LOCAL_STRATEGY_SORT_SECOND_MERGE"; /** * Value for the local strategy compiler hint that enforces a merge based local strategy. * Both inputs are assumed to be sorted and are merged. * For example, a Match or CoGroup operator will use a merge strategy to find pairs * of matching keys. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_MERGE = "LOCAL_STRATEGY_MERGE"; /** * Value for the local strategy compiler hint that enforces a hash based local strategy. * For example, a Match operator will use a hybrid-hash-join strategy to find pairs of * matching keys. The first input will be used to build the hash table, the second input will be * used to probe the table. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_HASH_BUILD_FIRST = "LOCAL_STRATEGY_HASH_BUILD_FIRST"; /** * Value for the local strategy compiler hint that enforces a hash based local strategy. * For example, a Match operator will use a hybrid-hash-join strategy to find pairs of * matching keys. The second input will be used to build the hash table, the first input will be * used to probe the table. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_HASH_BUILD_SECOND = "LOCAL_STRATEGY_HASH_BUILD_SECOND"; /** * Value for the local strategy compiler hint that chooses the outer side of the nested-loop local strategy. * A Cross operator will process the data of the first input in the outer-loop of the nested loops. * Hence, the data of the first input will be is streamed though, while the data of the second input is stored on * disk * and repeatedly read. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_NESTEDLOOP_STREAMED_OUTER_FIRST = "LOCAL_STRATEGY_NESTEDLOOP_STREAMED_OUTER_FIRST"; /** * Value for the local strategy compiler hint that chooses the outer side of the nested-loop local strategy. * A Cross operator will process the data of the second input in the outer-loop of the nested loops. * Hence, the data of the second input will be is streamed though, while the data of the first input is stored on * disk * and repeatedly read. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_NESTEDLOOP_STREAMED_OUTER_SECOND = "LOCAL_STRATEGY_NESTEDLOOP_STREAMED_OUTER_SECOND"; /** * Value for the local strategy compiler hint that chooses the outer side of the nested-loop local strategy. * A Cross operator will process the data of the first input in the outer-loop of the nested loops. * Further more, the first input, being the outer side, will be processed in blocks, and for each block, the second * input, * being the inner side, will read repeatedly from disk. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_NESTEDLOOP_BLOCKED_OUTER_FIRST = "LOCAL_STRATEGY_NESTEDLOOP_BLOCKED_OUTER_FIRST"; /** * Value for the local strategy compiler hint that chooses the outer side of the nested-loop local strategy. * A Cross operator will process the data of the second input in the outer-loop of the nested loops. * Further more, the second input, being the outer side, will be processed in blocks, and for each block, the first * input, * being the inner side, will read repeatedly from disk. * * @see #HINT_LOCAL_STRATEGY */ public static final String HINT_LOCAL_STRATEGY_NESTEDLOOP_BLOCKED_OUTER_SECOND = "LOCAL_STRATEGY_NESTEDLOOP_BLOCKED_OUTER_SECOND"; /** * The log handle that is used by the compiler to log messages. */ public static final Logger LOG = LoggerFactory.getLogger(PactCompiler.class); // ------------------------------------------------------------------------ // Members // ------------------------------------------------------------------------ /** * The statistics object used to obtain statistics, such as input sizes, * for the cost estimation process. */ private final DataStatistics statistics; /** * The cost estimator used by the compiler. */ private final CostEstimator costEstimator; /** * The default degree of parallelism for jobs compiled by this compiler. */ private int defaultDegreeOfParallelism; // ------------------------------------------------------------------------ // Constructor & Setup // ------------------------------------------------------------------------ /** * Creates a new compiler instance. The compiler has no access to statistics about the * inputs and can hence not determine any properties. It will perform all optimization with * unknown sizes and default to the most robust execution strategies. The * compiler also uses conservative default estimates for the operator costs, since * it has no access to another cost estimator. *

* The address of the job manager (to obtain system characteristics) is determined via the global configuration. */ public PactCompiler() { this(null, new DefaultCostEstimator()); } /** * Creates a new compiler instance that uses the statistics object to determine properties about the input. * Given those statistics, the compiler can make better choices for the execution strategies. * as if no filesystem was given. The compiler uses conservative default estimates for the operator costs, since * it has no access to another cost estimator. *

* The address of the job manager (to obtain system characteristics) is determined via the global configuration. * * @param stats * The statistics to be used to determine the input properties. */ public PactCompiler(DataStatistics stats) { this(stats, new DefaultCostEstimator()); } /** * Creates a new compiler instance. The compiler has no access to statistics about the * inputs and can hence not determine any properties. It will perform all optimization with * unknown sizes and default to the most robust execution strategies. It uses * however the given cost estimator to compute the costs of the individual operations. *

* The address of the job manager (to obtain system characteristics) is determined via the global configuration. * * @param estimator * The CostEstimator to use to cost the individual operations. */ public PactCompiler(CostEstimator estimator) { this(null, estimator); } /** * Creates a new compiler instance that uses the statistics object to determine properties about the input. * Given those statistics, the compiler can make better choices for the execution strategies. * as if no filesystem was given. It uses the given cost estimator to compute the costs of the individual * operations. *

* The address of the job manager (to obtain system characteristics) is determined via the global configuration. * * @param stats * The statistics to be used to determine the input properties. * @param estimator * The CostEstimator to use to cost the individual operations. */ public PactCompiler(DataStatistics stats, CostEstimator estimator) { this.statistics = stats; this.costEstimator = estimator; // determine the default parallelization degree this.defaultDegreeOfParallelism = GlobalConfiguration.getInteger(ConfigConstants.DEFAULT_PARALLELIZATION_DEGREE_KEY, ConfigConstants.DEFAULT_PARALLELIZATION_DEGREE); if (defaultDegreeOfParallelism < 1) { LOG.warn("Config value " + defaultDegreeOfParallelism + " for option " + ConfigConstants.DEFAULT_PARALLELIZATION_DEGREE + " is invalid. Ignoring and using a value of 1."); this.defaultDegreeOfParallelism = 1; } } // ------------------------------------------------------------------------ // Getters / Setters // ------------------------------------------------------------------------ public int getDefaultDegreeOfParallelism() { return defaultDegreeOfParallelism; } public void setDefaultDegreeOfParallelism(int defaultDegreeOfParallelism) { if (defaultDegreeOfParallelism > 0) { this.defaultDegreeOfParallelism = defaultDegreeOfParallelism; } else { throw new IllegalArgumentException("Default parallelism cannot be zero or negative."); } } // ------------------------------------------------------------------------ // Compilation // ------------------------------------------------------------------------ /** * Translates the given plan in to an OptimizedPlan, where all nodes have their local strategy assigned * and all channels have a shipping strategy assigned. The compiler connects to the job manager to obtain information * about the available instances and their memory and then chooses an instance type to schedule the execution on. *

* The compilation process itself goes through several phases: *

    *
  1. Create an optimizer data flow representation of the program, assign parallelism and compute size estimates.
    2. *
  2. Compute interesting properties and auxiliary structures.
    3. *
  3. Enumerate plan alternatives. This cannot be done in the same step as the interesting property computation (as * opposed to the Database approaches), because we support plans that are not trees.
    4. *
* * @param program The program to be translated. * @return The optimized plan. * @throws CompilerException * Thrown, if the plan is invalid or the optimizer encountered an inconsistent * situation during the compilation process. */ public OptimizedPlan compile(Plan program) throws CompilerException { // -------------------- try to get the connection to the job manager ---------------------- // --------------------------to obtain instance information -------------------------------- final OptimizerPostPass postPasser = getPostPassFromPlan(program); return compile(program, postPasser); } /** * Translates the given pact plan in to an OptimizedPlan, where all nodes have their local strategy assigned * and all channels have a shipping strategy assigned. The process goes through several phases: *
    *
  1. Create OptimizerNode representations of the PACTs, assign parallelism and compute size estimates.
    2. *
  2. Compute interesting properties and auxiliary structures.
    3. *
  3. Enumerate plan alternatives. This cannot be done in the same step as the interesting property computation (as * opposed to the Database approaches), because we support plans that are not trees.
    4. *
* * @param program The program to be translated. * @param postPasser The function to be used for post passing the optimizer's plan and setting the * data type specific serialization routines. * @return The optimized plan. * * @throws CompilerException * Thrown, if the plan is invalid or the optimizer encountered an inconsistent * situation during the compilation process. */ private OptimizedPlan compile(Plan program, OptimizerPostPass postPasser) throws CompilerException { if (program == null || postPasser == null) { throw new NullPointerException(); } if (LOG.isDebugEnabled()) { LOG.debug("Beginning compilation of program '" + program.getJobName() + '\''); } // set the default degree of parallelism int defaultParallelism = program.getDefaultParallelism() > 0 ? program.getDefaultParallelism() : this.defaultDegreeOfParallelism; // log the output if (LOG.isDebugEnabled()) { LOG.debug("Using a default degree of parallelism of " + defaultParallelism + '.'); } // the first step in the compilation is to create the optimizer plan representation // this step does the following: // 1) It creates an optimizer plan node for each operator // 2) It connects them via channels // 3) It looks for hints about local strategies and channel types and // sets the types and strategies accordingly // 4) It makes estimates about the data volume of the data sources and // propagates those estimates through the plan GraphCreatingVisitor graphCreator = new GraphCreatingVisitor(defaultParallelism); program.accept(graphCreator); // if we have a plan with multiple data sinks, add logical optimizer nodes that have two data-sinks as children // each until we have only a single root node. This allows to transparently deal with the nodes with // multiple outputs OptimizerNode rootNode; if (graphCreator.sinks.size() == 1) { rootNode = graphCreator.sinks.get(0); } else if (graphCreator.sinks.size() > 1) { Iterator iter = graphCreator.sinks.iterator(); rootNode = iter.next(); while (iter.hasNext()) { rootNode = new SinkJoiner(rootNode, iter.next()); } } else { throw new CompilerException("Bug: The optimizer plan representation has no sinks."); } // now that we have all nodes created and recorded which ones consume memory, tell the nodes their minimal // guaranteed memory, for further cost estimations. we assume an equal distribution of memory among consumer tasks rootNode.accept(new IdAndEstimatesVisitor(this.statistics)); // Now that the previous step is done, the next step is to traverse the graph again for the two // steps that cannot directly be performed during the plan enumeration, because we are dealing with DAGs // rather than a trees. That requires us to deviate at some points from the classical DB optimizer algorithms. // // 1) propagate the interesting properties top-down through the graph // 2) Track information about nodes with multiple outputs that are later on reconnected in a node with // multiple inputs. InterestingPropertyVisitor propsVisitor = new InterestingPropertyVisitor(this.costEstimator); rootNode.accept(propsVisitor); BranchesVisitor branchingVisitor = new BranchesVisitor(); rootNode.accept(branchingVisitor); // perform a sanity check: the root may not have any unclosed branches if (rootNode.getOpenBranches() != null && rootNode.getOpenBranches().size() > 0) { throw new CompilerException("Bug: Logic for branching plans (non-tree plans) has an error, and does not " + "track the re-joining of branches correctly."); } // the final step is now to generate the actual plan alternatives List bestPlan = rootNode.getAlternativePlans(this.costEstimator); if (bestPlan.size() != 1) { throw new CompilerException("Error in compiler: more than one best plan was created!"); } // check if the best plan's root is a data sink (single sink plan) // if so, directly take it. if it is a sink joiner node, get its contained sinks PlanNode bestPlanRoot = bestPlan.get(0); List bestPlanSinks = new ArrayList(4); if (bestPlanRoot instanceof SinkPlanNode) { bestPlanSinks.add((SinkPlanNode) bestPlanRoot); } else if (bestPlanRoot instanceof SinkJoinerPlanNode) { ((SinkJoinerPlanNode) bestPlanRoot).getDataSinks(bestPlanSinks); } DeadlockPreventer dp = new DeadlockPreventer(); dp.resolveDeadlocks(bestPlanSinks); // finalize the plan OptimizedPlan plan = new PlanFinalizer().createFinalPlan(bestPlanSinks, program.getJobName(), program); // swap the binary unions for n-ary unions. this changes no strategies or memory consumers whatsoever, so // we can do this after the plan finalization plan.accept(new BinaryUnionReplacer()); // post pass the plan. this is the phase where the serialization and comparator code is set postPasser.postPass(plan); return plan; } /** * This function performs only the first step to the compilation process - the creation of the optimizer * representation of the plan. No estimations or enumerations of alternatives are done here. * * @param program The plan to generate the optimizer representation for. * @return The optimizer representation of the plan, as a collection of all data sinks * from the plan can be traversed. */ public static List createPreOptimizedPlan(Plan program) { GraphCreatingVisitor graphCreator = new GraphCreatingVisitor(1); program.accept(graphCreator); return graphCreator.sinks; } // ------------------------------------------------------------------------ // Visitors for Compilation Traversals // ------------------------------------------------------------------------ /** * This utility class performs the translation from the user specified program to the optimizer plan. * It works as a visitor that walks the user's job in a depth-first fashion. During the descend, it creates * an optimizer node for each operator, respectively data source or -sink. During the ascend, it connects * the nodes to the full graph. *

* This translator relies on the setInputs method in the nodes. As that method implements the size * estimation and the awareness for optimizer hints, the sizes will be properly estimated and the translated plan * already respects all optimizer hints. */ private static final class GraphCreatingVisitor implements Visitor> { private final Map, OptimizerNode> con2node; // map from the operator objects to their // corresponding optimizer nodes private final List sources; // all data source nodes in the optimizer plan private final List sinks; // all data sink nodes in the optimizer plan private final int defaultParallelism; // the default degree of parallelism private final GraphCreatingVisitor parent; // reference to enclosing creator, in case of a recursive translation private final boolean forceDOP; private GraphCreatingVisitor(int defaultParallelism) { this(null, false, defaultParallelism, null); } private GraphCreatingVisitor(GraphCreatingVisitor parent, boolean forceDOP, int defaultParallelism, HashMap, OptimizerNode> closure) { if (closure == null){ con2node = new HashMap, OptimizerNode>(); } else { con2node = closure; } this.sources = new ArrayList(4); this.sinks = new ArrayList(2); this.defaultParallelism = defaultParallelism; this.parent = parent; this.forceDOP = forceDOP; } @SuppressWarnings("deprecation") @Override public boolean preVisit(Operator c) { // check if we have been here before if (this.con2node.containsKey(c)) { return false; } final OptimizerNode n; // create a node for the operator (or sink or source) if we have not been here before if (c instanceof GenericDataSinkBase) { DataSinkNode dsn = new DataSinkNode((GenericDataSinkBase) c); this.sinks.add(dsn); n = dsn; } else if (c instanceof GenericDataSourceBase) { DataSourceNode dsn = new DataSourceNode((GenericDataSourceBase) c); this.sources.add(dsn); n = dsn; } else if (c instanceof MapOperatorBase) { n = new MapNode((MapOperatorBase) c); } else if (c instanceof MapPartitionOperatorBase) { n = new MapPartitionNode((MapPartitionOperatorBase) c); } else if (c instanceof org.apache.flink.api.common.operators.base.CollectorMapOperatorBase) { n = new CollectorMapNode((org.apache.flink.api.common.operators.base.CollectorMapOperatorBase) c); } else if (c instanceof FlatMapOperatorBase) { n = new FlatMapNode((FlatMapOperatorBase) c); } else if (c instanceof FilterOperatorBase) { n = new FilterNode((FilterOperatorBase) c); } else if (c instanceof ReduceOperatorBase) { n = new ReduceNode((ReduceOperatorBase) c); } else if (c instanceof GroupReduceOperatorBase) { n = new GroupReduceNode((GroupReduceOperatorBase) c); } else if (c instanceof JoinOperatorBase) { n = new MatchNode((JoinOperatorBase) c); } else if (c instanceof CoGroupOperatorBase) { n = new CoGroupNode((CoGroupOperatorBase) c); } else if (c instanceof CrossOperatorBase) { n = new CrossNode((CrossOperatorBase) c); } else if (c instanceof BulkIterationBase) { n = new BulkIterationNode((BulkIterationBase) c); } else if (c instanceof DeltaIterationBase) { n = new WorksetIterationNode((DeltaIterationBase) c); } else if (c instanceof Union){ n = new BinaryUnionNode((Union) c); } else if (c instanceof PartitionOperatorBase) { n = new PartitionNode((PartitionOperatorBase) c); } else if (c instanceof PartialSolutionPlaceHolder) { if (this.parent == null) { throw new InvalidProgramException("It is currently not supported to create data sinks inside iterations."); } final PartialSolutionPlaceHolder holder = (PartialSolutionPlaceHolder) c; final BulkIterationBase enclosingIteration = holder.getContainingBulkIteration(); final BulkIterationNode containingIterationNode = (BulkIterationNode) this.parent.con2node.get(enclosingIteration); // catch this for the recursive translation of step functions BulkPartialSolutionNode p = new BulkPartialSolutionNode(holder, containingIterationNode); p.setDegreeOfParallelism(containingIterationNode.getDegreeOfParallelism()); n = p; } else if (c instanceof WorksetPlaceHolder) { if (this.parent == null) { throw new InvalidProgramException("It is currently not supported to create data sinks inside iterations."); } final WorksetPlaceHolder holder = (WorksetPlaceHolder) c; final DeltaIterationBase enclosingIteration = holder.getContainingWorksetIteration(); final WorksetIterationNode containingIterationNode = (WorksetIterationNode) this.parent.con2node.get(enclosingIteration); // catch this for the recursive translation of step functions WorksetNode p = new WorksetNode(holder, containingIterationNode); p.setDegreeOfParallelism(containingIterationNode.getDegreeOfParallelism()); n = p; } else if (c instanceof SolutionSetPlaceHolder) { if (this.parent == null) { throw new InvalidProgramException("It is currently not supported to create data sinks inside iterations."); } final SolutionSetPlaceHolder holder = (SolutionSetPlaceHolder) c; final DeltaIterationBase enclosingIteration = holder.getContainingWorksetIteration(); final WorksetIterationNode containingIterationNode = (WorksetIterationNode) this.parent.con2node.get(enclosingIteration); // catch this for the recursive translation of step functions SolutionSetNode p = new SolutionSetNode(holder, containingIterationNode); p.setDegreeOfParallelism(containingIterationNode.getDegreeOfParallelism()); n = p; } else { throw new IllegalArgumentException("Unknown operator type: " + c); } this.con2node.put(c, n); // set the parallelism only if it has not been set before. some nodes have a fixed DOP, such as the // key-less reducer (all-reduce) if (n.getDegreeOfParallelism() < 1) { // set the degree of parallelism int par = c.getDegreeOfParallelism(); if (par > 0) { if (this.forceDOP && par != this.defaultParallelism) { par = this.defaultParallelism; LOG.warn("The degree-of-parallelism of nested Dataflows (such as step functions in iterations) is " + "currently fixed to the degree-of-parallelism of the surrounding operator (the iteration)."); } } else { par = this.defaultParallelism; } n.setDegreeOfParallelism(par); } return true; } @Override public void postVisit(Operator c) { OptimizerNode n = this.con2node.get(c); // first connect to the predecessors n.setInput(this.con2node); n.setBroadcastInputs(this.con2node); // if the node represents a bulk iteration, we recursively translate the data flow now if (n instanceof BulkIterationNode) { final BulkIterationNode iterNode = (BulkIterationNode) n; final BulkIterationBase iter = iterNode.getIterationContract(); // calculate closure of the anonymous function HashMap, OptimizerNode> closure = new HashMap, OptimizerNode>(con2node); // first, recursively build the data flow for the step function final GraphCreatingVisitor recursiveCreator = new GraphCreatingVisitor(this, true, iterNode.getDegreeOfParallelism(), closure); BulkPartialSolutionNode partialSolution = null; iter.getNextPartialSolution().accept(recursiveCreator); partialSolution = (BulkPartialSolutionNode) recursiveCreator.con2node.get(iter.getPartialSolution()); OptimizerNode rootOfStepFunction = recursiveCreator.con2node.get(iter.getNextPartialSolution()); if (partialSolution == null) { throw new CompilerException("Error: The step functions result does not depend on the partial solution."); } OptimizerNode terminationCriterion = null; if (iter.getTerminationCriterion() != null) { terminationCriterion = recursiveCreator.con2node.get(iter.getTerminationCriterion()); // no intermediate node yet, traverse from the termination criterion to build the missing parts if (terminationCriterion == null) { iter.getTerminationCriterion().accept(recursiveCreator); terminationCriterion = recursiveCreator.con2node.get(iter.getTerminationCriterion()); } } iterNode.setNextPartialSolution(rootOfStepFunction, terminationCriterion); iterNode.setPartialSolution(partialSolution); // go over the contained data flow and mark the dynamic path nodes StaticDynamicPathIdentifier identifier = new StaticDynamicPathIdentifier(iterNode.getCostWeight()); rootOfStepFunction.accept(identifier); if(terminationCriterion != null){ terminationCriterion.accept(identifier); } } else if (n instanceof WorksetIterationNode) { final WorksetIterationNode iterNode = (WorksetIterationNode) n; final DeltaIterationBase iter = iterNode.getIterationContract(); // calculate the closure of the anonymous function HashMap, OptimizerNode> closure = new HashMap, OptimizerNode>(con2node); // first, recursively build the data flow for the step function final GraphCreatingVisitor recursiveCreator = new GraphCreatingVisitor(this, true, iterNode.getDegreeOfParallelism(), closure); // descend from the solution set delta. check that it depends on both the workset // and the solution set. If it does depend on both, this descend should create both nodes iter.getSolutionSetDelta().accept(recursiveCreator); final SolutionSetNode solutionSetNode = (SolutionSetNode) recursiveCreator.con2node.get(iter.getSolutionSet()); final WorksetNode worksetNode = (WorksetNode) recursiveCreator.con2node.get(iter.getWorkset()); if (worksetNode == null) { throw new CompilerException("In the given plan, the solution set delta does not depend on the workset. This is a prerequisite in workset iterations."); } iter.getNextWorkset().accept(recursiveCreator); if (solutionSetNode == null || solutionSetNode.getOutgoingConnections() == null || solutionSetNode.getOutgoingConnections().isEmpty()) { throw new CompilerException("Error: The step function does not reference the solution set."); } else { for (PactConnection conn : solutionSetNode.getOutgoingConnections()) { OptimizerNode successor = conn.getTarget(); if (successor.getClass() == MatchNode.class) { // find out which input to the match the solution set is MatchNode mn = (MatchNode) successor; if (mn.getFirstPredecessorNode() == solutionSetNode) { mn.makeJoinWithSolutionSet(0); } else if (mn.getSecondPredecessorNode() == solutionSetNode) { mn.makeJoinWithSolutionSet(1); } else { throw new CompilerException(); } } else if (successor.getClass() == CoGroupNode.class) { CoGroupNode cg = (CoGroupNode) successor; if (cg.getFirstPredecessorNode() == solutionSetNode) { cg.makeCoGroupWithSolutionSet(0); } else if (cg.getSecondPredecessorNode() == solutionSetNode) { cg.makeCoGroupWithSolutionSet(1); } else { throw new CompilerException(); } } else { throw new CompilerException("Error: The only operations allowed on the solution set are Join and CoGroup."); } } } final OptimizerNode nextWorksetNode = recursiveCreator.con2node.get(iter.getNextWorkset()); final OptimizerNode solutionSetDeltaNode = recursiveCreator.con2node.get(iter.getSolutionSetDelta()); // set the step function nodes to the iteration node iterNode.setPartialSolution(solutionSetNode, worksetNode); iterNode.setNextPartialSolution(solutionSetDeltaNode, nextWorksetNode); // go over the contained data flow and mark the dynamic path nodes StaticDynamicPathIdentifier pathIdentifier = new StaticDynamicPathIdentifier(iterNode.getCostWeight()); nextWorksetNode.accept(pathIdentifier); iterNode.getSolutionSetDelta().accept(pathIdentifier); } } }; private static final class StaticDynamicPathIdentifier implements Visitor { private final Set seenBefore = new HashSet(); private final int costWeight; private StaticDynamicPathIdentifier(int costWeight) { this.costWeight = costWeight; } @Override public boolean preVisit(OptimizerNode visitable) { return this.seenBefore.add(visitable); } @Override public void postVisit(OptimizerNode visitable) { visitable.identifyDynamicPath(this.costWeight); } } /** * Simple visitor that sets the minimal guaranteed memory per task based on the amount of available memory, * the number of memory consumers, and on the task's degree of parallelism. */ private static final class IdAndEstimatesVisitor implements Visitor { private final DataStatistics statistics; private int id = 1; private IdAndEstimatesVisitor(DataStatistics statistics) { this.statistics = statistics; } @Override public boolean preVisit(OptimizerNode visitable) { if (visitable.getId() != -1) { // been here before return false; } return true; } @Override public void postVisit(OptimizerNode visitable) { // the node ids visitable.initId(this.id++); // connections need to figure out their maximum path depths for (PactConnection conn : visitable.getIncomingConnections()) { conn.initMaxDepth(); } for (PactConnection conn : visitable.getBroadcastConnections()) { conn.initMaxDepth(); } // the estimates visitable.computeOutputEstimates(this.statistics); // if required, recurse into the step function if (visitable instanceof IterationNode) { ((IterationNode) visitable).acceptForStepFunction(this); } } } /** * Visitor that computes the interesting properties for each node in the plan. On its recursive * depth-first descend, it propagates all interesting properties top-down. */ public static final class InterestingPropertyVisitor implements Visitor { private CostEstimator estimator; // the cost estimator for maximal costs of an interesting property /** * Creates a new visitor that computes the interesting properties for all nodes in the plan. * It uses the given cost estimator used to compute the maximal costs for an interesting property. * * @param estimator * The cost estimator to estimate the maximal costs for interesting properties. */ public InterestingPropertyVisitor(CostEstimator estimator) { this.estimator = estimator; } @Override public boolean preVisit(OptimizerNode node) { // The interesting properties must be computed on the descend. In case a node has multiple outputs, // that computation must happen during the last descend. if (node.getInterestingProperties() == null && node.haveAllOutputConnectionInterestingProperties()) { node.computeUnionOfInterestingPropertiesFromSuccessors(); node.computeInterestingPropertiesForInputs(this.estimator); return true; } else { return false; } } @Override public void postVisit(OptimizerNode visitable) {} } /** * On its re-ascend (post visit) this visitor, computes auxiliary maps that are needed to support plans * that are not a minimally connected DAG (Such plans are not trees, but at least one node feeds its * output into more than one other node). */ private static final class BranchesVisitor implements Visitor { @Override public boolean preVisit(OptimizerNode node) { return node.getOpenBranches() == null; } @Override public void postVisit(OptimizerNode node) { if (node instanceof IterationNode) { ((IterationNode) node).acceptForStepFunction(this); } node.computeUnclosedBranchStack(); } } /** * Utility class that traverses a plan to collect all nodes and add them to the OptimizedPlan. * Besides collecting all nodes, this traversal assigns the memory to the nodes. */ private static final class PlanFinalizer implements Visitor { private final Set allNodes; // a set of all nodes in the optimizer plan private final List sources; // all data source nodes in the optimizer plan private final List sinks; // all data sink nodes in the optimizer plan private final Deque stackOfIterationNodes; private int memoryConsumerWeights; // a counter of all memory consumers /** * Creates a new plan finalizer. */ private PlanFinalizer() { this.allNodes = new HashSet(); this.sources = new ArrayList(); this.sinks = new ArrayList(); this.stackOfIterationNodes = new ArrayDeque(); } private OptimizedPlan createFinalPlan(List sinks, String jobName, Plan originalPlan) { this.memoryConsumerWeights = 0; // traverse the graph for (SinkPlanNode node : sinks) { node.accept(this); } // assign the memory to each node if (this.memoryConsumerWeights > 0) { for (PlanNode node : this.allNodes) { // assign memory to the driver strategy of the node final int consumerWeight = node.getMemoryConsumerWeight(); if (consumerWeight > 0) { final double relativeMem = (double)consumerWeight / this.memoryConsumerWeights; node.setRelativeMemoryPerSubtask(relativeMem); if (LOG.isDebugEnabled()) { LOG.debug("Assigned " + relativeMem + " of total memory to each subtask of " + node.getPactContract().getName() + "."); } } // assign memory to the local and global strategies of the channels for (Channel c : node.getInputs()) { if (c.getLocalStrategy().dams()) { final double relativeMem = 1.0 / this.memoryConsumerWeights; c.setRelativeMemoryLocalStrategy(relativeMem); if (LOG.isDebugEnabled()) { LOG.debug("Assigned " + relativeMem + " of total memory to each local strategy " + "instance of " + c + "."); } } if (c.getTempMode() != TempMode.NONE) { final double relativeMem = 1.0/ this.memoryConsumerWeights; c.setRelativeTempMemory(relativeMem); if (LOG.isDebugEnabled()) { LOG.debug("Assigned " + relativeMem + " of total memory to each instance of the temp " + "table" + " " + "for " + c + "."); } } } } } return new OptimizedPlan(this.sources, this.sinks, this.allNodes, jobName, originalPlan); } @Override public boolean preVisit(PlanNode visitable) { // if we come here again, prevent a further descend if (!this.allNodes.add(visitable)) { return false; } if (visitable instanceof SinkPlanNode) { this.sinks.add((SinkPlanNode) visitable); } else if (visitable instanceof SourcePlanNode) { this.sources.add((SourcePlanNode) visitable); } else if (visitable instanceof BulkPartialSolutionPlanNode) { // tell the partial solution about the iteration node that contains it final BulkPartialSolutionPlanNode pspn = (BulkPartialSolutionPlanNode) visitable; final IterationPlanNode iteration = this.stackOfIterationNodes.peekLast(); // sanity check! if (iteration == null || !(iteration instanceof BulkIterationPlanNode)) { throw new CompilerException("Bug: Error finalizing the plan. " + "Cannot associate the node for a partial solutions with its containing iteration."); } pspn.setContainingIterationNode((BulkIterationPlanNode) iteration); } else if (visitable instanceof WorksetPlanNode) { // tell the partial solution about the iteration node that contains it final WorksetPlanNode wspn = (WorksetPlanNode) visitable; final IterationPlanNode iteration = this.stackOfIterationNodes.peekLast(); // sanity check! if (iteration == null || !(iteration instanceof WorksetIterationPlanNode)) { throw new CompilerException("Bug: Error finalizing the plan. " + "Cannot associate the node for a partial solutions with its containing iteration."); } wspn.setContainingIterationNode((WorksetIterationPlanNode) iteration); } else if (visitable instanceof SolutionSetPlanNode) { // tell the partial solution about the iteration node that contains it final SolutionSetPlanNode sspn = (SolutionSetPlanNode) visitable; final IterationPlanNode iteration = this.stackOfIterationNodes.peekLast(); // sanity check! if (iteration == null || !(iteration instanceof WorksetIterationPlanNode)) { throw new CompilerException("Bug: Error finalizing the plan. " + "Cannot associate the node for a partial solutions with its containing iteration."); } sspn.setContainingIterationNode((WorksetIterationPlanNode) iteration); } // double-connect the connections. previously, only parents knew their children, because // one child candidate could have been referenced by multiple parents. for (Channel conn : visitable.getInputs()) { conn.setTarget(visitable); conn.getSource().addOutgoingChannel(conn); } for (Channel c : visitable.getBroadcastInputs()) { c.setTarget(visitable); c.getSource().addOutgoingChannel(c); } // count the memory consumption this.memoryConsumerWeights += visitable.getMemoryConsumerWeight(); for (Channel c : visitable.getInputs()) { if (c.getLocalStrategy().dams()) { this.memoryConsumerWeights++; } if (c.getTempMode() != TempMode.NONE) { this.memoryConsumerWeights++; } } for (Channel c : visitable.getBroadcastInputs()) { if (c.getLocalStrategy().dams()) { this.memoryConsumerWeights++; } if (c.getTempMode() != TempMode.NONE) { this.memoryConsumerWeights++; } } // pass the visitor to the iteraton's step function if (visitable instanceof IterationPlanNode) { // push the iteration node onto the stack final IterationPlanNode iterNode = (IterationPlanNode) visitable; this.stackOfIterationNodes.addLast(iterNode); // recurse ((IterationPlanNode) visitable).acceptForStepFunction(this); // pop the iteration node from the stack this.stackOfIterationNodes.removeLast(); } return true; } @Override public void postVisit(PlanNode visitable) {} } /** * A visitor that traverses the graph and collects cascading binary unions into a single n-ary * union operator. The exception is, when on of the union inputs is materialized, such as in the * static-code-path-cache in iterations. */ private static final class BinaryUnionReplacer implements Visitor { private final Set seenBefore = new HashSet(); @Override public boolean preVisit(PlanNode visitable) { if (this.seenBefore.add(visitable)) { if (visitable instanceof IterationPlanNode) { ((IterationPlanNode) visitable).acceptForStepFunction(this); } return true; } else { return false; } } @Override public void postVisit(PlanNode visitable) { if (visitable instanceof BinaryUnionPlanNode) { final BinaryUnionPlanNode unionNode = (BinaryUnionPlanNode) visitable; final Channel in1 = unionNode.getInput1(); final Channel in2 = unionNode.getInput2(); PlanNode newUnionNode; List inputs = new ArrayList(); collect(in1, inputs); collect(in2, inputs); newUnionNode = new NAryUnionPlanNode(unionNode.getOptimizerNode(), inputs, unionNode.getGlobalProperties()); for (Channel c : inputs) { c.setTarget(newUnionNode); } for(Channel channel : unionNode.getOutgoingChannels()){ channel.swapUnionNodes(newUnionNode); } } } private void collect(Channel in, List inputs) { if (in.getSource() instanceof NAryUnionPlanNode) { // sanity check if (in.getShipStrategy() != ShipStrategyType.FORWARD) { throw new CompilerException("Bug: Plan generation for Unions picked a ship strategy between binary plan operators."); } if (!(in.getLocalStrategy() == null || in.getLocalStrategy() == LocalStrategy.NONE)) { throw new CompilerException("Bug: Plan generation for Unions picked a local strategy between binary plan operators."); } inputs.addAll(((NAryUnionPlanNode) in.getSource()).getListOfInputs()); } else { // is not a union node, so we take the channel directly inputs.add(in); } } } // ------------------------------------------------------------------------ // Miscellaneous // ------------------------------------------------------------------------ private OptimizerPostPass getPostPassFromPlan(Plan program) { final String className = program.getPostPassClassName(); if (className == null) { throw new CompilerException("Optimizer Post Pass class description is null"); } try { Class clazz = Class.forName(className).asSubclass(OptimizerPostPass.class); try { return InstantiationUtil.instantiate(clazz, OptimizerPostPass.class); } catch (RuntimeException rtex) { // unwrap the source exception if (rtex.getCause() != null) { throw new CompilerException("Cannot instantiate optimizer post pass: " + rtex.getMessage(), rtex.getCause()); } else { throw rtex; } } } catch (ClassNotFoundException cnfex) { throw new CompilerException("Cannot load Optimizer post-pass class '" + className + "'.", cnfex); } catch (ClassCastException ccex) { throw new CompilerException("Class '" + className + "' is not an optimizer post passer.", ccex); } } }