morningman commented on code in PR #69194: URL: https://github.com/apache/doris/pull/69194#discussion_r4304750461
########## fe/fe-core/src/main/java/org/apache/doris/datasource/VariantWritePlanValidator.java: ########## @@ -0,0 +1,529 @@ +// 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.doris.datasource; + +import org.apache.doris.catalog.Column; +import org.apache.doris.common.Pair; +import org.apache.doris.nereids.CTEContext; +import org.apache.doris.nereids.exceptions.AnalysisException; +import org.apache.doris.nereids.trees.expressions.CTEId; +import org.apache.doris.nereids.trees.expressions.Cast; +import org.apache.doris.nereids.trees.expressions.ExprId; +import org.apache.doris.nereids.trees.expressions.Expression; +import org.apache.doris.nereids.trees.expressions.NamedExpression; +import org.apache.doris.nereids.trees.expressions.OrderExpression; +import org.apache.doris.nereids.trees.expressions.ScalarSubquery; +import org.apache.doris.nereids.trees.expressions.Slot; +import org.apache.doris.nereids.trees.expressions.SlotReference; +import org.apache.doris.nereids.trees.expressions.WhenClause; +import org.apache.doris.nereids.trees.expressions.WindowExpression; +import org.apache.doris.nereids.trees.expressions.functions.scalar.If; +import org.apache.doris.nereids.trees.plans.Plan; +import org.apache.doris.nereids.trees.plans.algebra.RecursiveCte; +import org.apache.doris.nereids.trees.plans.algebra.SetOperation; +import org.apache.doris.nereids.trees.plans.logical.LogicalCTEConsumer; +import org.apache.doris.nereids.trees.plans.logical.LogicalCTEProducer; +import org.apache.doris.nereids.trees.plans.logical.LogicalExcept; +import org.apache.doris.nereids.trees.plans.logical.LogicalGenerate; +import org.apache.doris.nereids.trees.plans.logical.LogicalIntersect; +import org.apache.doris.nereids.trees.plans.logical.LogicalRecursiveUnionAnchor; +import org.apache.doris.nereids.trees.plans.logical.LogicalUnion; +import org.apache.doris.nereids.trees.plans.logical.LogicalWorkTableReference; +import org.apache.doris.nereids.types.ArrayType; +import org.apache.doris.nereids.types.DataType; +import org.apache.doris.nereids.types.MapType; +import org.apache.doris.nereids.types.StructField; +import org.apache.doris.nereids.types.StructType; +import org.apache.doris.nereids.types.VariantType; + +import com.google.common.collect.ImmutableSet; + +import java.util.ArrayList; +import java.util.Collections; +import java.util.HashMap; +import java.util.HashSet; +import java.util.List; +import java.util.Map; +import java.util.Optional; +import java.util.Set; + +/** Shared source-plan validation for external Variant sinks. */ +public final class VariantWritePlanValidator { + private VariantWritePlanValidator() { + } + + /** + * Rejects an implicit Variant-to-non-Variant cast in the lineage of a Variant target column. + * + * <p>Common-type analysis for UNION, IF and CASE runs before sink binding. Without this + * check, an object or array Variant can become SQL NULL while being cast to a scalar, and the + * sink will only see that scalar/NULL and encode it back as Variant. Explicit casts remain an + * intentional user conversion and are not rejected. Only the expressions whose values can reach + * the target are inspected: the condition of an IF or of a CASE branch only chooses a value, so a + * Variant path compared with a scalar there loses nothing that is written. Likewise the later + * inputs of EXCEPT and INTERSECT only decide which rows of the first input stay, and the partition + * and order keys of a window and the ORDER BY of an aggregate function only arrange the rows whose + * values are used.</p> + */ + public static void validateNoLossyCoercion(List<Column> targetColumns, Plan sourcePlan) { + validateNoLossyCoercion(targetColumns, sourcePlan, Optional.empty()); + } + + /** Validates a sink while its enclosing CTE producers still live in the analyzer context. */ + public static void validateNoLossyCoercion( + List<Column> targetColumns, Plan sourcePlan, CTEContext cteContext) { + validateNoLossyCoercion(targetColumns, sourcePlan, Optional.of(cteContext)); + } + + private static void validateNoLossyCoercion( + List<Column> targetColumns, Plan sourcePlan, Optional<CTEContext> cteContext) { + if (targetColumns.size() != sourcePlan.getOutput().size()) { + throw new AnalysisException("VARIANT write target and source columns are not aligned"); + } + List<Integer> sourceOrdinals = new ArrayList<>(targetColumns.size()); + for (int i = 0; i < targetColumns.size(); i++) { + sourceOrdinals.add(i); + } + validateNoLossyCoercion(targetColumns, sourcePlan, sourceOrdinals, cteContext); + } + + /** + * Validates target columns that are embedded in a wider source output, such as row-level DML + * where the operation and row-id routing columns precede the table data columns. + */ + public static void validateNoLossyCoercion( + List<Column> targetColumns, Plan sourcePlan, List<Integer> sourceOrdinals) { + validateNoLossyCoercion(targetColumns, sourcePlan, sourceOrdinals, Optional.empty()); + } + + /** Validates mapped source ordinals with analyzer-time CTE producer visibility. */ + public static void validateNoLossyCoercion( + List<Column> targetColumns, Plan sourcePlan, List<Integer> sourceOrdinals, + CTEContext cteContext) { + validateNoLossyCoercion(targetColumns, sourcePlan, sourceOrdinals, Optional.of(cteContext)); + } + + private static void validateNoLossyCoercion( + List<Column> targetColumns, Plan sourcePlan, List<Integer> sourceOrdinals, + Optional<CTEContext> cteContext) { + if (targetColumns.size() != sourceOrdinals.size()) { + throw new AnalysisException("VARIANT write target and source columns are not aligned"); + } + + List<Integer> variantTargetOrdinals = new ArrayList<>(); + for (int i = 0; i < targetColumns.size(); i++) { + DataType targetType = DataType.fromCatalogType(targetColumns.get(i).getType()); + if (VariantType.containsVariant(targetType)) { + variantTargetOrdinals.add(i); + } + } + if (variantTargetOrdinals.isEmpty()) { + return; + } + + TraceContext traceContext = new TraceContext(sourcePlan, cteContext); + List<Slot> sourceOutputs = sourcePlan.getOutput(); + for (int targetOrdinal : variantTargetOrdinals) { + int sourceOrdinal = sourceOrdinals.get(targetOrdinal); + if (sourceOrdinal < 0 || sourceOrdinal >= sourceOutputs.size()) { + throw new AnalysisException("VARIANT write target and source columns are not aligned"); + } + Column targetColumn = targetColumns.get(targetOrdinal); + traceOutputLineage( + sourcePlan, + Collections.singleton(sourceOutputs.get(sourceOrdinal).getExprId()), + targetColumn.getName(), + traceContext); + } + } + + /** + * Validates the requested outputs and returns lineage ExprIds supplied by an outer scope. + * Scalar subqueries use the returned IDs to continue tracing correlated columns in the + * enclosing plan. + */ + private static Set<ExprId> traceOutputLineage( + Plan plan, Set<ExprId> requiredExprIds, String targetColumn, TraceContext context) { + if (requiredExprIds.isEmpty()) { + return Collections.emptySet(); + } + if (plan instanceof SetOperation) { + return traceSetOperation(plan, (SetOperation) plan, requiredExprIds, targetColumn, context); + } + if (plan instanceof RecursiveCte) { + return traceRecursiveCte(plan, (RecursiveCte) plan, requiredExprIds, targetColumn, context); + } + if (plan instanceof LogicalGenerate) { + return traceGenerate((LogicalGenerate<?>) plan, requiredExprIds, targetColumn, context); + } + if (plan instanceof LogicalCTEConsumer) { + return traceCteConsumer((LogicalCTEConsumer) plan, requiredExprIds, targetColumn, context); + } + + Set<ExprId> unresolvedExprIds = new HashSet<>(requiredExprIds); + Set<ExprId> inputExprIds = new HashSet<>(); + for (Expression expression : plan.getExpressions()) { + if (!(expression instanceof NamedExpression)) { + continue; + } + NamedExpression namedExpression = (NamedExpression) expression; + if (!requiredExprIds.contains(namedExpression.getExprId())) { + continue; + } + inputExprIds.addAll(validateExpression(namedExpression, targetColumn, context)); + unresolvedExprIds.remove(namedExpression.getExprId()); + } + inputExprIds.addAll(unresolvedExprIds); + + Set<ExprId> externalExprIds = new HashSet<>(inputExprIds); + for (Plan child : plan.children()) { + Set<ExprId> childOutputExprIds = child.getOutputExprIdSet(); + Set<ExprId> childRequiredExprIds = new HashSet<>(inputExprIds); + childRequiredExprIds.retainAll(childOutputExprIds); + externalExprIds.removeAll(childOutputExprIds); + externalExprIds.addAll(traceOutputLineage(child, childRequiredExprIds, targetColumn, context)); + } + return externalExprIds; + } + + private static Set<ExprId> traceRecursiveCte( + Plan plan, RecursiveCte recursiveCte, Set<ExprId> requiredExprIds, + String targetColumn, TraceContext context) { + // The recursive term reads the previous iteration through the work table, so a column it writes + // can carry the value of another column. Trace those columns as well until no new one is reached. + Map<ExprId, Integer> workTableOrdinals = workTableOrdinals(plan); + Set<ExprId> externalExprIds = new HashSet<>(); + Set<Integer> tracedOrdinals = new HashSet<>(); + Set<Integer> pendingOrdinals = requiredOutputOrdinals(plan, requiredExprIds); + while (!pendingOrdinals.isEmpty()) { + tracedOrdinals.addAll(pendingOrdinals); + Set<ExprId> childExternalExprIds = new HashSet<>(); + for (int childIndex = 0; childIndex < plan.arity(); childIndex++) { + List<SlotReference> childOutputs = recursiveCte.getRegularChildOutput(childIndex); + Set<ExprId> childRequiredExprIds = new HashSet<>(); + for (int ordinal : pendingOrdinals) { + if (ordinal < childOutputs.size()) { + childRequiredExprIds.add(childOutputs.get(ordinal).getExprId()); + } + } + childExternalExprIds.addAll(traceOutputLineage( + plan.child(childIndex), childRequiredExprIds, targetColumn, context)); + } + pendingOrdinals = new HashSet<>(); + for (ExprId exprId : childExternalExprIds) { + Integer ordinal = workTableOrdinals.get(exprId); + if (ordinal == null) { + externalExprIds.add(exprId); + } else if (!tracedOrdinals.contains(ordinal)) { + pendingOrdinals.add(ordinal); + } + } + } + return externalExprIds; + } + + /** Maps each column of the work table the recursive term reads to the CTE output ordinal it holds. */ + private static Map<ExprId, Integer> workTableOrdinals(Plan recursiveCte) { + CTEId cteId = ((LogicalRecursiveUnionAnchor<?>) recursiveCte.child(0)).getCteId(); + List<LogicalWorkTableReference> workTables = recursiveCte.child(1).collectToList( + node -> node instanceof LogicalWorkTableReference + && ((LogicalWorkTableReference) node).getCteId().equals(cteId)); + Map<ExprId, Integer> ordinals = new HashMap<>(); + for (LogicalWorkTableReference workTable : workTables) { + List<Slot> outputs = workTable.getOutput(); + for (int ordinal = 0; ordinal < outputs.size(); ordinal++) { + ordinals.put(outputs.get(ordinal).getExprId(), ordinal); + } + } + return ordinals; + } + + private static Set<ExprId> traceGenerate( + LogicalGenerate<?> generate, Set<ExprId> requiredExprIds, + String targetColumn, TraceContext context) { + Set<ExprId> childRequiredExprIds = new HashSet<>(requiredExprIds); + List<Slot> generatorOutputs = generate.getGeneratorOutput(); + List<? extends Expression> generators = generate.getGenerators(); + for (int ordinal = 0; ordinal < generatorOutputs.size(); ordinal++) { + Slot generatorOutput = generatorOutputs.get(ordinal); + if (!requiredExprIds.contains(generatorOutput.getExprId())) { + continue; + } + childRequiredExprIds.remove(generatorOutput.getExprId()); + childRequiredExprIds.addAll(validateExpression(generators.get(ordinal), targetColumn, context)); + } + Set<ExprId> childOutputExprIds = generate.child().getOutputExprIdSet(); + Set<ExprId> externalExprIds = new HashSet<>(childRequiredExprIds); + externalExprIds.removeAll(childOutputExprIds); + childRequiredExprIds.retainAll(childOutputExprIds); + externalExprIds.addAll(traceOutputLineage( + generate.child(), childRequiredExprIds, targetColumn, context)); + return externalExprIds; + } + + private static Set<ExprId> traceSetOperation( + Plan plan, SetOperation setOperation, Set<ExprId> requiredExprIds, + String targetColumn, TraceContext context) { + Set<ExprId> externalExprIds = new HashSet<>(); + Set<Integer> requiredOrdinals = requiredOutputOrdinals(plan, requiredExprIds); + + if (plan instanceof LogicalUnion) { + for (List<NamedExpression> constantRow : ((LogicalUnion) plan).getConstantExprsList()) { + for (int ordinal : requiredOrdinals) { + if (ordinal < constantRow.size()) { + externalExprIds.addAll(validateExpression(constantRow.get(ordinal), targetColumn, context)); + } + } + } + } + + // EXCEPT and INTERSECT return rows of their first input; the later inputs only decide which of + // those rows stay, so none of their values is written. + int valueInputs = plan instanceof LogicalExcept || plan instanceof LogicalIntersect + ? 1 + : setOperation.getArity(); + for (int childIndex = 0; childIndex < valueInputs; childIndex++) { + List<SlotReference> childOutputs = setOperation.getRegularChildOutput(childIndex); + Set<ExprId> childRequiredExprIds = new HashSet<>(); + for (int ordinal : requiredOrdinals) { + if (ordinal < childOutputs.size()) { + childRequiredExprIds.add(childOutputs.get(ordinal).getExprId()); + } + } + externalExprIds.addAll(traceOutputLineage( + plan.child(childIndex), childRequiredExprIds, targetColumn, context)); + } + return externalExprIds; + } + + private static Set<Integer> requiredOutputOrdinals(Plan plan, Set<ExprId> requiredExprIds) { + Set<Integer> requiredOrdinals = new HashSet<>(); + List<Slot> outputs = plan.getOutput(); + for (int i = 0; i < outputs.size(); i++) { + if (requiredExprIds.contains(outputs.get(i).getExprId())) { + requiredOrdinals.add(i); + } + } + return requiredOrdinals; + } + + private static Set<ExprId> traceCteConsumer( + LogicalCTEConsumer consumer, Set<ExprId> requiredExprIds, + String targetColumn, TraceContext context) { + // A producer shared by several consumers is traced once per output, whichever sets of its outputs + // the consumers request, so the work stays proportional to the plan rather than to the number of + // paths through it or of the sets of outputs requested along them. + Set<ExprId> externalExprIds = new HashSet<>(); + for (Map.Entry<Slot, Slot> mapping : consumer.getConsumerToProducerOutputMap().entrySet()) { + if (requiredExprIds.contains(mapping.getKey().getExprId())) { + externalExprIds.addAll(traceCteOutput( + consumer, mapping.getValue().getExprId(), targetColumn, context)); + } + } + return externalExprIds; + } + + private static Set<ExprId> traceCteOutput( + LogicalCTEConsumer consumer, ExprId producerExprId, String targetColumn, TraceContext context) { + Pair<CTEId, ExprId> traceKey = Pair.of(consumer.getCteId(), producerExprId); + Set<ExprId> completedTrace = context.completedCteTraces.get(traceKey); + if (completedTrace != null) { + return completedTrace; + } + LogicalCTEProducer<?> producer = context.cteProducers.get(consumer.getCteId()); + Optional<Plan> producerPlan; + if (producer == null) { + // A producer taken from the analyzer context is not part of the traced plan, so the + // producers of the WITH clauses nested in it have not been collected yet. + producerPlan = context.findAnalyzerCteProducer(consumer); + producerPlan.ifPresent(context::collectCteProducers); + } else { + producerPlan = Optional.of(producer.child()); + } + if (!producerPlan.isPresent() || !context.activeCteIds.add(consumer.getCteId())) { + return Collections.emptySet(); + } + try { + Set<ExprId> externalExprIds = ImmutableSet.copyOf(traceOutputLineage( + producerPlan.get(), Collections.singleton(producerExprId), targetColumn, context)); + context.completedCteTraces.put(traceKey, externalExprIds); + return externalExprIds; + } finally { + context.activeCteIds.remove(consumer.getCteId()); + } + } + + /** + * Validates the casts on the values {@code expression} writes and returns the ExprIds those + * values come from: the slots they read, and the correlated slots of their scalar subqueries, + * which an outer scope may supply. + */ + private static Set<ExprId> validateExpression( + Expression expression, String targetColumn, TraceContext context) { + List<Expression> valueNodes = valueNodes(expression); + validateLossyCast(valueNodes, targetColumn); + Set<ExprId> inputExprIds = new HashSet<>(); + for (Expression node : valueNodes) { + if (node instanceof Slot) { + inputExprIds.add(((Slot) node).getExprId()); + } else if (node instanceof ScalarSubquery) { + inputExprIds.addAll(traceScalarSubquery((ScalarSubquery) node, targetColumn, context)); + } + } + return inputExprIds; + } + + /** Traces a scalar subquery and returns its correlated slots that the subquery itself does not supply. */ + private static Set<ExprId> traceScalarSubquery( + ScalarSubquery scalarSubquery, String targetColumn, TraceContext context) { + scalarSubquery.getTypeCoercionExpr() + .ifPresent(coercion -> validateLossyCast(valueNodes(coercion), targetColumn)); + Plan queryPlan = scalarSubquery.getQueryPlan(); + context.collectCteProducers(queryPlan); + Set<ExprId> unresolvedExprIds = new HashSet<>(traceOutputLineage( + queryPlan, + Collections.singleton(queryPlan.getOutput().get(0).getExprId()), + targetColumn, + context)); + Set<ExprId> correlatedExprIds = new HashSet<>(); + for (Slot correlateSlot : scalarSubquery.getCorrelateSlots()) { + correlatedExprIds.add(correlateSlot.getExprId()); + } + unresolvedExprIds.retainAll(correlatedExprIds); + return unresolvedExprIds; + } + + /** + * The nodes of {@code expression} whose values can reach its result, in pre-order: every node + * except those under the condition of an IF or of a CASE branch, which only chooses a value, and + * those under the partition and order keys of a window or the ORDER BY of an aggregate function + * such as GROUP_CONCAT, which only arrange the rows whose values are used. + */ + private static List<Expression> valueNodes(Expression expression) { + List<Expression> nodes = new ArrayList<>(); + collectValueNodes(expression, nodes); + return nodes; + } + + private static void collectValueNodes(Expression expression, List<Expression> nodes) { + nodes.add(expression); + if (expression instanceof If) { + collectValueNodes(((If) expression).getTrueValue(), nodes); + collectValueNodes(((If) expression).getFalseValue(), nodes); + } else if (expression instanceof WhenClause) { + collectValueNodes(((WhenClause) expression).getResult(), nodes); + } else if (expression instanceof WindowExpression) { + collectValueNodes(((WindowExpression) expression).getFunction(), nodes); + } else { + for (Expression child : expression.children()) { + if (!(child instanceof OrderExpression)) { + collectValueNodes(child, nodes); Review Comment: The mechanism is right, and it reproduces on a local cluster. Into a format-version 3 Iceberg table `(id INT, payload VARIANT)`, `NULLIF(1, IF(TRUE, v, 1))`, `ELT(v['n'], 'a', 'b')`, `MAX_BY(1, IF(TRUE, v, 1))` and `ARRAY_FILTER(ARRAY(1, 2), ARRAY(v['n'] > 0, TRUE))` each fail with `VARIANT write cannot safely convert input column 'payload': source expression implicitly casts VARIANT to ...`. So do `SUBSTRING('abc', v['n'])` and a comparison written as the value, `v['n'] > 0`. We are keeping this rule in this PR, for the reasons given for the [complex-target case](https://github.com/apache/doris/pull/69194#discussion_r4293116118): - It rejects; it never writes a wrong value. The error names the column and the type of the cast, and the release note asks for an explicit CAST in that case. With one, the statements go through: `ELT(CAST(v['n'] AS INT), 'a', 'b')`, `NULLIF(1, IF(TRUE, CAST(v AS INT), 1))`, `SUBSTRING('abc', CAST(v['n'] AS INT))` and `CAST(v['n'] AS INT) > 0` write `a`, `1`, `bc` and `true` on the same cluster. - What this PR leaves out is decided by the kind of node: the condition of an IF or of a CASE branch, the partition and order keys of a window, the ORDER BY of an aggregate function, and the later inputs of EXCEPT and INTERSECT. The cases here turn on what one argument of one function means: the second argument of NULLIF, the index of ELT, the key of MIN_BY and MAX_BY, the mask of ARRAY_FILTER, and by the same reasoning the position of SUBSTRING and the operands of a comparison. `ExpressionUtils.getCaseWhenLikeBranchResults` models NULLIF and NVL but none of the others. Such a list would have to follow every function added later, and an argument listed by mistake would let a lossy cast through without an error, which is the failure this check exists to prevent. - Branch-4.1 and branch-4.2 search the whole written expression with `collectFirst`, so they reject all of these as well, and also the conditions and keys this PR leaves out. Tracing only the arguments whose values can be returned belongs with tracing only the VARIANT leaves of a complex target, so we will take it up in the follow-up for that thread. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
