nsivabalan commented on code in PR #20524: URL: https://github.com/apache/hudi/pull/20524#discussion_r4474744669
########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/dataskipping/RangeRule.scala: ########## @@ -0,0 +1,124 @@ +/* + * 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.hudi.dataskipping + +import org.apache.spark.sql.catalyst.expressions.{Expression, If, Literal} +import org.apache.spark.sql.hudi.dataskipping.rules.{CastRule, ConstantOperandRule, DateTimeFormatRule, MonotonicFunctionRule, PrefixRule, TimeZoneShiftRule} + +/** + * Stage 3 (Range), rules: how does a function of the child map the child's range? + * + * A [[Step]] says: `term` is a function of `child` with this [[Direction]], so the range of `term` + * is the range of `child` pushed through it. Invariant: a rule only emits a step when its guard + * holds for this `term`; a term no rule knows has no range (and the file is kept). + * + * Emitting a Step is a correctness claim: `term` is order-preserving in `child`. + * <ul> + * <li>[[Increasing]]: a <= b implies f(a) <= f(b)</li> + * <li>[[Decreasing]]: a <= b implies f(a) >= f(b)</li> + * </ul> + * The claim must hold for every value the file's range can hold (or wherever the step's + * `fileGuard` is true). If it is false, data skipping drops files that have matching rows. When + * unsure, return Nil: that only costs pruning, never correctness. + */ +sealed trait Direction + +/** + * Order-preserving (non-decreasing): a <= b implies f(a) <= f(b), so f(lo) is the low bound and + * f(hi) the high bound. Every step today is increasing. + */ +case object Increasing extends Direction + +/** + * Order-reversing (non-increasing): a <= b implies f(a) >= f(b), so f(hi) is the low bound and + * f(lo) the high bound. Unused today; negative factors would use it. + */ +case object Decreasing extends Direction + +/** + * One rule application. + * + * @param term the function application being ranged, for ex `x + 1` + * @param child the operand whose range is pushed through, for ex `x` + * @param direction whether `term` grows or shrinks with `child` + * @param fileGuard per-file condition under which the mapping holds (for ex, no overflow); outside it + * the bounds become NULL, i.e. unknown. `None` for every step today + * @param partial whether `term` can be NULL for a non-NULL `child` (for ex, `log` of a negative), in + * which case null / value counts of the child no longer describe `term`. No behavioral + * effect yet: IS NULL / IS NOT NULL only take bare columns + */ +case class Step(term: Expression, + child: Expression, + direction: Direction, + fileGuard: Option[Range => Expression], + partial: Boolean) { + + def apply(r: Range): Range = { + val (lo, hi) = direction match { + case Increasing => (r.lo, r.hi) + case Decreasing => (r.hi, r.lo) Review Comment: 💬 **SUGGESTION.** `Decreasing` and the `fileGuard` branch of `guarded` are unreachable today, which is probably most of the 54 uncovered lines codecov reports. The follow-ups will rely on them being correct, so a small `Step.apply` test would help now: swapped lo/hi for `Decreasing`, the `If(guard, bound, NULL)` shape, and `partial` dropping the counts. Otherwise the first PR that uses them is also the first to test them. ########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/dataskipping/rules/CastRule.scala: ########## @@ -77,4 +80,13 @@ object HoodieSparkTypeUtils { case _ => true } + + override def steps(term: Expression): Seq[Step] = term match { + // NOTE: Some casts (for ex, string to date in non-ANSI mode) return NULL on malformed input; + // `partial` is left false for now, as it has no behavioral effect yet + case cast @ Cast(child, _, _, _) if preservesOrder(cast.child.dataType, cast.dataType) => Review Comment: 💬 **SUGGESTION.** On master the matcher's `Cast(OrderPreservingTransformation(attrRef), ...)` ran the extractor first, so the guard never saw an unresolved child. Here `preservesOrder(cast.child.dataType, ...)` runs first, and with `strictResolution = false` (expression index) an unresolved child throws from `dataType`. This is the 'one unreachable case' from the equivalence run. `SparkBaseIndexSupport` catches `NonFatal`, so the worst case is losing pruning. Still, `if cast.child.resolved && preservesOrder(...)` makes it impossible rather than unreachable, and costs nothing. ########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/dataskipping/Range.scala: ########## @@ -0,0 +1,99 @@ +/* + * 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.hudi.dataskipping + +import org.apache.hudi.ColumnStatsIndexSupport.{getMaxColumnNameFor, getMinColumnNameFor, getNullCountColumnNameFor, getValueCountColumnNameFor} +import org.apache.hudi.SparkAdapterSupport.sparkAdapter + +import org.apache.spark.sql.catalyst.expressions.{AttributeReference, Expression} +import org.apache.spark.sql.functions.col +import org.apache.spark.sql.hudi.DataSkippingUtils.getTargetIndexedColumnName + +/** + * Stage 3 (Range): what values can the term take in a file? + * + * `lo` / `hi` are expressions over the stats table bounding the term within one file; `nulls` / `rows` + * are its null and value counts, when known. A range is read from a [[RangeSource]] for a column, + * or derived through a chain of [[Step]]s from the range of the step's child. + * + * Invariant: no range means unknown, and unknown keeps the file. + */ +case class Range(lo: Expression, hi: Expression, nulls: Option[Expression], rows: Option[Expression]) + +/** + * Where ranges come from. Returns `None` for anything it does not hold stats for. + */ +sealed trait RangeSource { + def rangeOf(term: Expression): Option[Range] +} + +/** + * Column stats index: a column maps to its min / max / null-count / value-count columns. + * + * NOTE: Expression indexes go through this source too, but only with the bare column: + * `ExpressionIndexSupport` strips the indexed f(col) down to col before translating, so the + * stored f(col) bounds are compared as is + */ +case class ColumnStatsSource(indexedCols: Seq[String]) extends RangeSource { + override def rangeOf(term: Expression): Option[Range] = term match { + case column: AttributeReference => + getTargetIndexedColumnName(column, indexedCols).map { colName => + Range( + lo = statsColumn(getMinColumnNameFor(colName)), + hi = statsColumn(getMaxColumnNameFor(colName)), + nulls = Some(statsColumn(getNullCountColumnNameFor(colName))), + rows = Some(statsColumn(getValueCountColumnNameFor))) + } + case _ => None + } + + /** A column of the transposed column stats table, for ex `qty_minValue`. */ + private def statsColumn(name: String): Expression = sparkAdapter.getExpressionFromColumn(col(name)) +} + +object Range { + + def derive(term: Expression, source: RangeSource): Option[Range] = + viaRules(term, source).map(rebindOffChainReferences(term, source, _)) + + /** + * The source first; otherwise the first rule step whose child derives a range. Several steps + * only arise for operators accepted on either side (for ex, `BitwiseOr`), tried in today's order. + */ + private def viaRules(term: Expression, source: RangeSource): Option[Range] = + source.rangeOf(term).orElse( + RangeRules.of(term).view.flatMap(step => viaRules(step.child, source).map(step.apply)).headOption) + + /** + * Re-binds EVERY reference to the column, including ones in operands off the rule chain (for ex, + * the days operand of `date_add(d, dayofmonth(d))`), matching the pre-stage translator. A [[Step]] + * only rebuilds its chain slot, so the leftover references are re-bound to the same bound here. + * Such terms are not order-preserving in general; kept as is so the output stays identical. + */ + private def rebindOffChainReferences(term: Expression, source: RangeSource, r: Range): Range = { Review Comment: ⚠️ **IMPORTANT.** This re-binds every off-chain reference to the column's **lo** for the low bound and **hi** for the high bound. That only lines up while every step is `Increasing`. With a `Decreasing` step, `Step.apply` builds the low bound from `r.hi`, but this function then re-binds the off-chain operands of that same expression to `r.lo`. The result mixes bounds and isn't a bound of anything. Today nothing emits `Decreasing`, so there's no impact. But the plan has 'order-reversing ranges for negative factors and datediff' and 'constant operands for DateAdd / DateSub / ShiftLeft / ShiftRight / DateDiff' as separate follow-ups. If the datediff `start` one lands first, `datediff(<expr of d>, d)` hits exactly this path. Could you make the dependency explicit in code, not just in the plan? For example: - skip the re-binding, and return no range at all, when any step in the chain is `Decreasing`; or - require a constant other operand in any rule that emits `Decreasing`, and add a `require` in `Step` saying so. Either way, please add a test that pins it. ########## hudi-spark-datasource/hudi-spark/src/test/scala/org/apache/hudi/TestHoodieCatalystExpressionUtils.scala: ########## @@ -41,8 +42,14 @@ class TestHoodieCatalystExpressionUtils extends SparkAdapterSupport { private val boolAttr = AttributeReference("b", BooleanType)() private val binAttr = AttributeReference("bin", BinaryType)() - private def matched(expr: Expression): Option[AttributeReference] = - sparkAdapter.getCatalystExpressionUtils.tryMatchAttributeOrderingPreservingTransformation(expr) + /** + * Walks the rule table down to a bare column: the attribute itself, else the first rule step + * whose child resolves to one. Like the stage-3 range walk, it does not check reference count. + */ + private def matched(expr: Expression): Option[AttributeReference] = expr match { Review Comment: 💬 **SUGGESTION.** This helper re-implements the range walk instead of calling it, so the test now pins the rule table but not `Range.derive` (the source-first order, the either-side fall-through, the re-binding). A stub `RangeSource` that returns a marker `Range` for one attribute, used with `Range.derive(expr, stub)`, would keep the assertions and exercise the production path. -- 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]
