sigmod commented on code in PR #55629: URL: https://github.com/apache/spark/pull/55629#discussion_r3184233299
########## sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/RewriteNearestByJoin.scala: ########## @@ -0,0 +1,141 @@ +/* + * 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.spark.sql.catalyst.optimizer + +import org.apache.spark.sql.catalyst.expressions._ +import org.apache.spark.sql.catalyst.expressions.aggregate._ +import org.apache.spark.sql.catalyst.plans._ +import org.apache.spark.sql.catalyst.plans.logical._ +import org.apache.spark.sql.catalyst.rules._ + +/** + * Replaces a logical [[NearestByJoin]] operator with a `Generate(Inline(...))` over an + * `Aggregate` that tags each left row with a unique id, cross-joins with the right side, and + * groups by the unique id to compute the top-K matches via `MAX_BY`/`MIN_BY` (K-overload). + * + * Input Pseudo-Query: + * {{{ + * SELECT * FROM left [INNER | LEFT OUTER] JOIN right + * {APPROX | EXACT} NEAREST k BY {DISTANCE | SIMILARITY} expr + * }}} + * + * Rewritten Plan (SIMILARITY, INNER join type): + * {{{ + * Generate inline(_matches), [N], outer=false, [right.col1, right.col2, ...] + * +- Aggregate [__qid], + * [first(left.col0) AS left.col0, ..., first(left.colN-1) AS left.colN-1, + * max_by(struct(right.*), expr, k) AS _matches] + * +- Join LeftOuter + * :- Project [left.*, monotonically_increasing_id() AS __qid] + * : +- left + * +- right + * }}} + * + * For `DISTANCE`, `MIN_BY` is used instead of `MAX_BY`. For `LEFT OUTER`, the `Generate` is + * constructed with `outer = true` so left rows with no matches (empty/null `_matches`) are + * preserved with `NULL` right-side columns. + * + * If `rankingExpression` is nondeterministic (legal only under `APPROX`), an extra + * `Project` is inserted above the `Join` to materialize the value as `__ranking__`. The + * standard projection machinery runs `Nondeterministic.initialize(partitionIndex)` on every + * nondeterministic descendant before any value is evaluated, so `MaxMinByK` only ever sees a + * plain `AttributeReference` and never evaluates a nondeterministic expression directly. + * + * Unlike [[RewriteAsOfJoin]], which uses a correlated scalar subquery, this rule materializes + * the cross product directly. A scalar subquery returns a single value per left row, so it + * cannot carry K matches without an array-valued subquery + `Generate(Inline(...))` -- which + * collapses back to the same cross product after decorrelation. The aggregate-then-inline form + * makes the intended shape explicit and avoids round-tripping through subquery decorrelation. + */ +object RewriteNearestByJoin extends Rule[LogicalPlan] { + def apply(plan: LogicalPlan): LogicalPlan = plan.transformUpWithNewOutput { + case j @ NearestByJoin(left, right, joinType, _, numResults, rankingExpression, direction) => + // 1. Tag each left row with a unique id so that rows from the same left row can later be + // grouped together after the cross-join with `right`. + val qidAlias = Alias(MonotonicallyIncreasingID(), "__qid")() + val taggedLeft = Project(left.output :+ qidAlias, left) + val qidAttr = qidAlias.toAttribute + + // 2. LEFT OUTER-join the tagged left with right (no join condition). LEFT OUTER + // (rather than INNER) preserves left rows even when `right` is empty, so that a + // `LEFT OUTER NEAREST BY` query still returns those rows with `NULL` right-side + // columns after the aggregate + inline below. When `right` is non-empty every left + // row already has right-row pairings, so LEFT OUTER and INNER are equivalent. + // + // `CheckCartesianProducts` recognizes this synthetic join structurally (by its + // parent `Aggregate` containing a `MaxMinByK`) and skips it, so user queries + // written as `NEAREST BY` are not rejected when `spark.sql.crossJoin.enabled` is + // false. + val join = Join(taggedLeft, right, LeftOuter, None, JoinHint.NONE) + + val (aggInput, rankingForAgg) = if (!rankingExpression.deterministic) { + val rankingAlias = Alias(rankingExpression, "__ranking__")() + Project(join.output :+ rankingAlias, join) -> rankingAlias.toAttribute + } else { + join -> rankingExpression + } + + // 4. Aggregate grouped by `__qid`: + // - first(col) for every left column so it flows to the output. + // - max_by/min_by(struct(right.*), ranking, k) as `_matches`. + // The ranking expression references left and right columns directly; no outer + // reference is needed because both sides are present in the joined input. + val rightStruct = CreateStruct(right.output) + // reverse = true -> MIN_BY (smallest ranking value first, for DISTANCE) + // reverse = false -> MAX_BY (largest ranking value first, for SIMILARITY) + val reverse = direction match { + case NearestByDistance => true + case NearestBySimilarity => false + } + val topK = MaxMinByK( + rightStruct, + rankingForAgg, + Literal(numResults), + reverse = reverse).toAggregateExpression() + val matchesAlias = Alias(topK, "__nearest_matches__")() + + // Carry left columns through with `First`. Within a `__qid` group every row has the same + // left values (each group corresponds to one left row), so `First` is effectively a no-op. + // We use `First` rather than adding all left columns to the GROUP BY because grouping by + // `__qid` alone keeps the shuffle key small. + val firstLeftAggs = left.output.map { attr => + Alias( + First(attr, ignoreNulls = false).toAggregateExpression(), + attr.name)(exprId = attr.exprId, qualifier = attr.qualifier) + } + val aggregate = Aggregate(Seq(qidAttr), firstLeftAggs :+ matchesAlias, aggInput) + + // 4. Generate inline(_matches) expands the K-element array into K rows, exposing each + // struct field as a top-level column. `outer = true` for LEFT OUTER preserves the + // left row with NULL right columns when there are no matches. + val generatorOutput = right.output.map { a => + AttributeReference(a.name, a.dataType, nullable = true, a.metadata)( + qualifier = a.qualifier) + } + val generate = Generate( + Inline(matchesAlias.toAttribute), + unrequiredChildIndex = Seq(aggregate.output.indexOf(matchesAlias.toAttribute)), + outer = joinType == LeftOuter, + qualifier = None, + generatorOutput = generatorOutput, + child = aggregate) + + val attrMapping = j.output.zip(generate.output) + generate -> attrMapping Review Comment: We have to project away the `qidAttr` attribute? Otherwise, the output plan's schema is not equivalent to the input plan's schema and hence it might be a correctness issue when the NNJoin is the last step in a query? If we don't use `qidAttr` but just group by the struct, then we also need a `Project` to unfold the struct into ordinary columns (depending on whether we group by struct)? -- 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]
