SanJSp commented on code in PR #55508:
URL: https://github.com/apache/spark/pull/55508#discussion_r3137491014
##########
common/utils/src/main/resources/error/error-conditions.json:
##########
@@ -3278,6 +3278,26 @@
"message" : [
"`startingVersion` is required when `endingVersion` is specified for
CDC queries."
]
+ },
+ "NET_CHANGES_NOT_YET_SUPPORTED" : {
+ "message" : [
+ "The `deduplicationMode = netChanges` option on connector
`<changelogName>` is not yet supported. Use `deduplicationMode =
dropCarryovers` (default) or `deduplicationMode = none` instead."
+ ]
+ },
+ "STREAMING_POST_PROCESSING_NOT_SUPPORTED" : {
+ "message" : [
+ "Change Data Capture (CDC) streaming reads on connector
`<changelogName>` do not yet support post-processing (carry-over removal,
update detection, or net change computation). The requested combination of
options would require post-processing, which is currently only available for
batch reads. Use a batch read, or set `deduplicationMode = none` and
`computeUpdates = false` to receive raw change rows in streaming."
+ ]
+ },
+ "UPDATE_DETECTION_REQUIRES_CARRY_OVER_REMOVAL" : {
+ "message" : [
+ "`computeUpdates` cannot be used with `deduplicationMode=none` on
connector `<changelogName>` because the connector emits copy-on-write
carry-over pairs (`containsCarryoverRows()` returns true) that would be
silently mislabeled as updates. Set `deduplicationMode` to `dropCarryovers` or
`netChanges`."
+ ]
+ },
+ "UNEXPECTED_MULTIPLE_CHANGES_PER_ROW_VERSION" : {
+ "message" : [
+ "Connector emitted multiple delete or insert rows for the same
`(rowId, _commit_version)` partition. The `Changelog` contract requires at most
one logical change per row identity per commit when
`containsIntermediateChanges() = false`. Either fix the connector to
deduplicate intermediate states, or set `containsIntermediateChanges() = true`
and use `deduplicationMode = netChanges`."
+ ]
}
Review Comment:
Done. Split `UNEXPECTED_MULTIPLE_CHANGES_PER_ROW_VERSION` into new
`CHANGELOG_CONTRACT_VIOLATION` class with `sqlState = "XX000"` (matches
`CHECKPOINT_FILE_CHECKSUM_VERIFICATION_FAILED` and similar "upstream source
violated an invariant" errors). The `RaiseError` literal in the rule and the
test `checkError` conditions are updated.
##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/analysis/ResolveChangelogTable.scala:
##########
@@ -0,0 +1,315 @@
+/*
+ * 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.analysis
+
+import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.expressions.aggregate.{Count, Max, Min}
+import org.apache.spark.sql.catalyst.plans.logical._
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.catalyst.streaming.StreamingRelationV2
+import org.apache.spark.sql.catalyst.trees.TreeNodeTag
+import org.apache.spark.sql.connector.catalog.{Changelog, ChangelogInfo}
+import org.apache.spark.sql.errors.QueryCompilationErrors
+import org.apache.spark.sql.execution.datasources.v2.{ChangelogTable,
DataSourceV2Relation}
+import org.apache.spark.sql.types.{IntegerType, StringType}
+
+/**
+ * Post-processes a resolved [[ChangelogTable]] read to apply CDC option
semantics
+ * (carry-over removal, update detection) and to enforce supported option
combinations.
+ *
+ * Fires after [[ResolveRelations]] has wrapped the connector's [[Changelog]]
in a
+ * [[ChangelogTable]]. Both batch ([[DataSourceV2Relation]]) and streaming
+ * ([[StreamingRelationV2]]) reads are handled:
+ * - Batch: the requested post-processing passes are injected as logical
operators on top
+ * of the relation. Carry-over removal and update detection are fused into
a single
+ * pass over a (rowId, _commit_version)-partitioned Window: the Filter
drops CoW
+ * carry-over pairs (same rowVersion on both sides) and the subsequent
Project relabels
+ * real delete+insert pairs as update_preimage / update_postimage.
+ * - Streaming: post-processing is not yet supported. If the requested
options would
+ * require any post-processing, the rule throws an explicit
[[AnalysisException]] to
+ * prevent silent wrong results. Streams that don't require
post-processing pass
+ * through unchanged.
+ *
+ * Net change computation (`deduplicationMode = netChanges`) is not yet
implemented and
+ * is rejected up-front for both batch and streaming.
+ */
+object ResolveChangelogTable extends Rule[LogicalPlan] {
+
+ private val CHANGELOG_TRANSFORMED_TAG =
+ TreeNodeTag[Boolean]("changelog_transformed")
+
+ private object HelperColumn {
+ final val DelCnt = "_del_cnt"
+ final val InsCnt = "_ins_cnt"
+ final val MinRv = "_min_rv"
+ final val MaxRv = "_max_rv"
+
+ val all: Set[String] = Set(DelCnt, InsCnt, MinRv, MaxRv)
+ }
+
+ override def apply(plan: LogicalPlan): LogicalPlan = {
+ if (isAlreadyTransformed(plan)) return plan
+ var updatedPlan = plan
+ updatedPlan = plan.resolveOperatorsUp {
+ case rel @ DataSourceV2Relation(table: ChangelogTable, _, _, _, _, _) =>
+ val changelog = table.changelog
+ val req = evaluateRequirements(changelog, table.changelogInfo)
+
+ var updatedRel: LogicalPlan = rel
+ if (req.requiresCarryOverRemoval || req.requiresUpdateDetection) {
+ updatedRel = addRowLevelPostProcessing(
+ rel, changelog, req.requiresCarryOverRemoval,
req.requiresUpdateDetection)
+ }
+ if (req.requiresNetChanges) {
+ updatedRel = injectNetChangeComputation(updatedRel, changelog)
+ }
+ updatedRel
+
+ case rel @ StreamingRelationV2(_, _, table: ChangelogTable, _, _, _, _,
_, _) =>
+ // Streaming CDC reads do not yet apply post-processing. Run the same
option /
+ // capability validation as the batch path so silent wrong results are
impossible:
+ // either no post-processing would be required (fall through, return
raw stream),
+ // or we throw an explicit AnalysisException.
+ val changelog = table.changelog
+ val req = evaluateRequirements(changelog, table.changelogInfo)
+ if (req.needsAny) {
+ throw
QueryCompilationErrors.cdcStreamingPostProcessingNotSupported(changelog.name())
+ }
+ rel
+ }
+ if (updatedPlan ne plan) {
+ updatedPlan.setTagValue(CHANGELOG_TRANSFORMED_TAG, true)
+ }
+ updatedPlan
+ }
+
+ //
---------------------------------------------------------------------------
+ // Option validation & Requirement Computation
+ //
---------------------------------------------------------------------------
+
+ /**
+ * Captures which post-processing passes a CDC query requires, derived from
the
+ * user-provided [[ChangelogInfo]] options and the connector-declared
[[Changelog]]
+ * capability flags.
+ */
+ private case class PostProcessingRequirements(
+ requiresCarryOverRemoval: Boolean,
+ requiresUpdateDetection: Boolean,
+ requiresNetChanges: Boolean) {
+ def needsAny: Boolean =
+ requiresCarryOverRemoval || requiresUpdateDetection || requiresNetChanges
+ }
+
+ /**
+ * Validates CDC option/capability combinations and computes which
post-processing
+ * passes are required. Throws an [[org.apache.spark.sql.AnalysisException]]
for
+ * unsupported or contradictory combinations (currently: `netChanges`
deduplication,
+ * and `computeUpdates` with surfaced carry-overs but no carry-over removal).
+ */
+ private def evaluateRequirements(
+ changelog: Changelog,
+ options: ChangelogInfo): PostProcessingRequirements = {
+ // Net change computation is not yet implemented.
+ if (options.deduplicationMode() ==
ChangelogInfo.DeduplicationMode.NET_CHANGES) {
+ throw
QueryCompilationErrors.cdcNetChangesNotYetSupported(changelog.name())
+ }
+
+ val requiresCarryOverRemoval =
+ options.deduplicationMode() != ChangelogInfo.DeduplicationMode.NONE &&
+ changelog.containsCarryoverRows()
+ val requiresUpdateDetection =
+ options.computeUpdates() && changelog.representsUpdateAsDeleteAndInsert()
+ val requiresNetChanges =
+ options.deduplicationMode() ==
ChangelogInfo.DeduplicationMode.NET_CHANGES &&
+ changelog.containsIntermediateChanges()
+
+ // If carry-overs are surfaced and update detection is enabled without
carry-over
+ // removal, carry-overs would be falsely classified as updates, leading to
wrong
+ // results. Hence we throw.
+ if (requiresUpdateDetection &&
+ changelog.containsCarryoverRows() &&
+ options.deduplicationMode() == ChangelogInfo.DeduplicationMode.NONE) {
+ throw QueryCompilationErrors.cdcUpdateDetectionRequiresCarryOverRemoval(
+ changelog.name())
+ }
+
+ PostProcessingRequirements(
+ requiresCarryOverRemoval, requiresUpdateDetection, requiresNetChanges)
+ }
+
+ //
---------------------------------------------------------------------------
+ // Row Level Post Processing (Update Detection & Carry-over Removal)
+ //
---------------------------------------------------------------------------
+
+ /**
+ * Adds row-level post-processing (carry-over removal and/or update
detection) on top of
+ * the given plan:
+ * - both active -> Window(counts + rv bounds) -> Filter ->
Project(relabel) -> Drop helpers
+ * - carry-over only -> Window(counts + rv bounds) -> Filter -> Drop
helpers
+ * - update only -> Window(counts only) -> Project(relabel) -> Drop
helpers
+ * - neither -> not invoked (caller guards this case)
+ */
+ private def addRowLevelPostProcessing(
+ plan: LogicalPlan,
+ cl: Changelog,
+ requiresCarryOverRemoval: Boolean,
+ requiresUpdateDetection: Boolean): LogicalPlan = {
+ // Row-version bounds in the Window are needed iff we filter carry-over
pairs.
+ var modifiedPlan = addPostProcessingWindow(plan, cl,
+ includeRowVersionBounds = requiresCarryOverRemoval)
+ if (requiresCarryOverRemoval) modifiedPlan =
addCarryOverPairFilter(modifiedPlan)
+ if (requiresUpdateDetection) modifiedPlan =
addUpdateRelabelProjection(modifiedPlan)
+ removeHelperColumns(modifiedPlan)
+ }
+
+ /**
+ * Adds a Window node partitioned by (rowId, _commit_version) that computes
+ * `_del_cnt` and `_ins_cnt` per partition, and, when
`includeRowVersionBounds`
+ * is true, additionally `_min_rv` / `_max_rv` (min/max of
`Changelog.rowVersion()`).
+ *
+ * `_del_cnt` / `_ins_cnt` drive update detection (1 each -> relabel as
+ * update_preimage / update_postimage). `_min_rv` / `_max_rv` drive
carry-over
+ * detection (within a delete+insert pair, equal bounds signal a CoW
carry-over).
+ */
+ private def addPostProcessingWindow(
+ plan: LogicalPlan,
+ cl: Changelog,
+ includeRowVersionBounds: Boolean): LogicalPlan = {
+ val changeTypeAttr = getAttribute(plan, "_change_type")
+ val rowIdExprs =
V2ExpressionUtils.resolveRefs[NamedExpression](cl.rowId().toSeq, plan)
Review Comment:
Handled in #55507 via `ChangelogTable.validateSchema`: if any
post-processing capability is advertised, `rowId()` must return a non-empty
array (otherwise `INVALID_CHANGELOG_SCHEMA.MISSING_ROW_ID`), and the default
`UnsupportedOperationException` on `rowId()` / `rowVersion()` is surfaced
directly. Validation fires in the `ChangelogTable` constructor, so the rule
always sees a validated table. Tests for both modes are in #55507.
##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/analysis/ResolveChangelogTable.scala:
##########
@@ -0,0 +1,315 @@
+/*
+ * 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.analysis
+
+import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.expressions.aggregate.{Count, Max, Min}
+import org.apache.spark.sql.catalyst.plans.logical._
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.catalyst.streaming.StreamingRelationV2
+import org.apache.spark.sql.catalyst.trees.TreeNodeTag
+import org.apache.spark.sql.connector.catalog.{Changelog, ChangelogInfo}
+import org.apache.spark.sql.errors.QueryCompilationErrors
+import org.apache.spark.sql.execution.datasources.v2.{ChangelogTable,
DataSourceV2Relation}
+import org.apache.spark.sql.types.{IntegerType, StringType}
+
+/**
+ * Post-processes a resolved [[ChangelogTable]] read to apply CDC option
semantics
+ * (carry-over removal, update detection) and to enforce supported option
combinations.
+ *
+ * Fires after [[ResolveRelations]] has wrapped the connector's [[Changelog]]
in a
+ * [[ChangelogTable]]. Both batch ([[DataSourceV2Relation]]) and streaming
+ * ([[StreamingRelationV2]]) reads are handled:
+ * - Batch: the requested post-processing passes are injected as logical
operators on top
+ * of the relation. Carry-over removal and update detection are fused into
a single
+ * pass over a (rowId, _commit_version)-partitioned Window: the Filter
drops CoW
+ * carry-over pairs (same rowVersion on both sides) and the subsequent
Project relabels
+ * real delete+insert pairs as update_preimage / update_postimage.
+ * - Streaming: post-processing is not yet supported. If the requested
options would
+ * require any post-processing, the rule throws an explicit
[[AnalysisException]] to
+ * prevent silent wrong results. Streams that don't require
post-processing pass
+ * through unchanged.
+ *
+ * Net change computation (`deduplicationMode = netChanges`) is not yet
implemented and
+ * is rejected up-front for both batch and streaming.
+ */
+object ResolveChangelogTable extends Rule[LogicalPlan] {
+
+ private val CHANGELOG_TRANSFORMED_TAG =
+ TreeNodeTag[Boolean]("changelog_transformed")
+
+ private object HelperColumn {
+ final val DelCnt = "_del_cnt"
+ final val InsCnt = "_ins_cnt"
+ final val MinRv = "_min_rv"
+ final val MaxRv = "_max_rv"
+
+ val all: Set[String] = Set(DelCnt, InsCnt, MinRv, MaxRv)
+ }
+
+ override def apply(plan: LogicalPlan): LogicalPlan = {
+ if (isAlreadyTransformed(plan)) return plan
+ var updatedPlan = plan
+ updatedPlan = plan.resolveOperatorsUp {
+ case rel @ DataSourceV2Relation(table: ChangelogTable, _, _, _, _, _) =>
+ val changelog = table.changelog
+ val req = evaluateRequirements(changelog, table.changelogInfo)
+
+ var updatedRel: LogicalPlan = rel
+ if (req.requiresCarryOverRemoval || req.requiresUpdateDetection) {
+ updatedRel = addRowLevelPostProcessing(
+ rel, changelog, req.requiresCarryOverRemoval,
req.requiresUpdateDetection)
+ }
+ if (req.requiresNetChanges) {
+ updatedRel = injectNetChangeComputation(updatedRel, changelog)
+ }
+ updatedRel
+
+ case rel @ StreamingRelationV2(_, _, table: ChangelogTable, _, _, _, _,
_, _) =>
+ // Streaming CDC reads do not yet apply post-processing. Run the same
option /
+ // capability validation as the batch path so silent wrong results are
impossible:
+ // either no post-processing would be required (fall through, return
raw stream),
+ // or we throw an explicit AnalysisException.
+ val changelog = table.changelog
+ val req = evaluateRequirements(changelog, table.changelogInfo)
+ if (req.needsAny) {
+ throw
QueryCompilationErrors.cdcStreamingPostProcessingNotSupported(changelog.name())
+ }
+ rel
+ }
+ if (updatedPlan ne plan) {
+ updatedPlan.setTagValue(CHANGELOG_TRANSFORMED_TAG, true)
Review Comment:
Done. Added `resolved: Boolean = false` on `ChangelogTable` (field lives in
#55507 so both PRs stay compile-compatible). The rule uses `!table.resolved` as
pattern guard, `rel.copy(table = table.copy(resolved = true))` to mark after
transformation. Plan-tag mechanism, `isAlreadyTransformed`, and `TreeNodeTag`
import are removed.
--
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]
