sigmod commented on code in PR #55629: URL: https://github.com/apache/spark/pull/55629#discussion_r3184256514
########## 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")() Review Comment: > would there be a concern in terms of performance and even distribution of keys? I think correctness is more important here - e.g., no duplicate query rows in the result set :-) We can always improve performance later, e.g., with index, or with a better implementation of such as an aggregate. -- 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]
