-
Notifications
You must be signed in to change notification settings - Fork 4.7k
[Iceberg CDC sink] Split late data and create commit windows #40006
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: master
Are you sure you want to change the base?
Changes from all commits
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,224 @@ | ||
| /* | ||
| * 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.beam.sdk.io.iceberg.cdc.sink; | ||
|
|
||
| import static org.apache.beam.sdk.util.Preconditions.checkStateNotNull; | ||
| import static org.apache.beam.vendor.guava.v32_1_2_jre.com.google.common.base.Preconditions.checkArgument; | ||
|
|
||
| import java.util.Map; | ||
| import org.apache.beam.sdk.Pipeline; | ||
| import org.apache.beam.sdk.coders.Coder; | ||
| import org.apache.beam.sdk.coders.KvCoder; | ||
| import org.apache.beam.sdk.coders.RowCoder; | ||
| import org.apache.beam.sdk.extensions.sorter.BufferedExternalSorter; | ||
| import org.apache.beam.sdk.extensions.sorter.SortValues; | ||
| import org.apache.beam.sdk.schemas.Schema; | ||
| import org.apache.beam.sdk.transforms.GroupByKey; | ||
| import org.apache.beam.sdk.transforms.PTransform; | ||
| import org.apache.beam.sdk.transforms.ParDo; | ||
| import org.apache.beam.sdk.transforms.windowing.AfterPane; | ||
| import org.apache.beam.sdk.transforms.windowing.AfterWatermark; | ||
| import org.apache.beam.sdk.transforms.windowing.DefaultTrigger; | ||
| import org.apache.beam.sdk.transforms.windowing.FixedWindows; | ||
| import org.apache.beam.sdk.transforms.windowing.GlobalWindows; | ||
| import org.apache.beam.sdk.transforms.windowing.Window; | ||
| import org.apache.beam.sdk.values.KV; | ||
| import org.apache.beam.sdk.values.PCollection; | ||
| import org.apache.beam.sdk.values.PCollection.IsBounded; | ||
| import org.apache.beam.sdk.values.PCollectionTuple; | ||
| import org.apache.beam.sdk.values.PInput; | ||
| import org.apache.beam.sdk.values.POutput; | ||
| import org.apache.beam.sdk.values.PValue; | ||
| import org.apache.beam.sdk.values.Row; | ||
| import org.apache.beam.sdk.values.TupleTag; | ||
| import org.apache.beam.sdk.values.TupleTagList; | ||
| import org.apache.beam.vendor.guava.v32_1_2_jre.com.google.common.collect.ImmutableMap; | ||
| import org.checkerframework.checker.nullness.qual.Nullable; | ||
| import org.joda.time.Duration; | ||
|
|
||
| /** | ||
| * Windows sharded, sort-keyed records into event-time commit windows, grouped by the {@code | ||
| * KV<destination, shard>} key. Each {@code (destination, shard, window)} becomes one commit unit | ||
| * group and the event-time watermark is the commit barrier. Each group is sorted by the byte sort | ||
| * key applied from {@link AssignCdcKeys}: one primary key's records contiguous, in {@code (seq, | ||
| * kind)} order within the key. | ||
| * | ||
| * <p>Late panes are routed to the DLQ (via {@link SplitLateData}) before any ordering happens. This | ||
| * is necessary because the downstream commit step skips panes if their window token is present in | ||
| * an already committed snapshot. If late panes are let through, their contents will never get to | ||
| * the table. | ||
| * | ||
| * <p>Windowing by input mode: | ||
| * | ||
| * <ul> | ||
| * <li><b>Bounded</b>: a single {@link GlobalWindows} window. One commit per destination, no late | ||
| * data possible. | ||
| * <li><b>Unbounded</b>: event-time {@link FixedWindows} of {@code triggeringFrequency}, firing on | ||
| * the watermark with late firings per element, discarding fired panes. | ||
| * </ul> | ||
| */ | ||
| final class CommitWindows | ||
| extends PTransform< | ||
| PCollection<KV<KV<String, Integer>, KV<byte[], CdcRecord>>>, CommitWindows.Result> { | ||
|
|
||
| private static final TupleTag<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> | ||
| ON_TIME_TAG = new TupleTag<>("onTime"); | ||
| private static final TupleTag<Row> DEAD_LETTER_TAG = new TupleTag<>("deadLetter"); | ||
|
|
||
| private final CdcWriteConfig config; | ||
| private final @Nullable Duration triggeringFrequency; | ||
| private final Duration allowedLateness; | ||
|
|
||
| CommitWindows( | ||
| CdcWriteConfig config, @Nullable Duration triggeringFrequency, Duration allowedLateness) { | ||
| this.config = config; | ||
| this.triggeringFrequency = triggeringFrequency; | ||
| this.allowedLateness = allowedLateness; | ||
| } | ||
|
|
||
| @Override | ||
| public Result expand(PCollection<KV<KV<String, Integer>, KV<byte[], CdcRecord>>> input) { | ||
| Schema deadLetterSchema = SplitLateData.deadLetterSchema(dataSchemaOf(input.getCoder())); | ||
|
|
||
| PCollection<KV<KV<String, Integer>, KV<byte[], CdcRecord>>> windowed = applyCommitWindow(input); | ||
|
|
||
| // Exactly one group per (destination, shard, window) | ||
| PCollection<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> grouped = | ||
| windowed.apply("GroupByShardKey", GroupByKey.create()); | ||
|
|
||
| // Late-data split before sorting anything | ||
| PCollectionTuple split = | ||
| grouped.apply( | ||
| "SplitLateData", | ||
| ParDo.of(new SplitLateData(deadLetterSchema, ON_TIME_TAG, DEAD_LETTER_TAG)) | ||
| .withOutputTags(ON_TIME_TAG, TupleTagList.of(DEAD_LETTER_TAG))); | ||
| PCollection<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> onTimeUnsorted = | ||
| split.get(ON_TIME_TAG).setCoder(grouped.getCoder()); | ||
| PCollection<Row> deadLetter = | ||
| split.get(DEAD_LETTER_TAG).setCoder(RowCoder.of(deadLetterSchema)); | ||
|
|
||
| // Sort each surviving group's records by the byte sort key. The secondary key is byte[] + | ||
| // ByteArrayCoder, so SortValues compares the raw CdcSortKey bytes (no coder framing): | ||
| // each primary key's records come out contiguous, in (seq, kind) order within the key. | ||
| PCollection<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> sorted = | ||
| onTimeUnsorted.apply( | ||
| "SortBySeqKind", | ||
| SortValues.create( | ||
| BufferedExternalSorter.options().withMemoryMB(config.getSorterMemoryMB()))); | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This could lead to OOMs since we are sorting entire windows per shard. Let's make sure there's enough memory for the default case. |
||
|
|
||
| return new Result(input.getPipeline(), sorted, deadLetter, deadLetterSchema); | ||
| } | ||
|
|
||
| /** Applies the commit-window assignment for the input's boundedness; see the class Javadoc. */ | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more.
|
||
| private PCollection<KV<KV<String, Integer>, KV<byte[], CdcRecord>>> applyCommitWindow( | ||
| PCollection<KV<KV<String, Integer>, KV<byte[], CdcRecord>>> input) { | ||
| if (input.isBounded() == IsBounded.BOUNDED) { | ||
| return input.apply( | ||
| "GlobalWindows", | ||
| Window.<KV<KV<String, Integer>, KV<byte[], CdcRecord>>>into(new GlobalWindows()) | ||
| .triggering(DefaultTrigger.of()) | ||
| .discardingFiredPanes()); | ||
| } | ||
| return input.apply( | ||
| "EventTimeWindows", | ||
| Window.<KV<KV<String, Integer>, KV<byte[], CdcRecord>>>into( | ||
| FixedWindows.of( | ||
| checkStateNotNull( | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Do this check early ? |
||
| triggeringFrequency, | ||
| "triggeringFrequency is required for unbounded input"))) | ||
| .triggering( | ||
| AfterWatermark.pastEndOfWindow().withLateFirings(AfterPane.elementCountAtLeast(1))) | ||
| .withAllowedLateness(allowedLateness) | ||
| .discardingFiredPanes()); | ||
| } | ||
|
|
||
| /** Extracts the CDC data schema carried by the input's nested {@link CdcRecordCoder}. */ | ||
| private static Schema dataSchemaOf(Coder<?> inputCoder) { | ||
| checkArgument( | ||
| inputCoder instanceof KvCoder, | ||
| "expected a KvCoder input element coder, got %s", | ||
| inputCoder); | ||
| Coder<?> valueCoder = ((KvCoder<?, ?>) inputCoder).getValueCoder(); | ||
| checkArgument( | ||
| valueCoder instanceof KvCoder, "expected a KvCoder input value coder, got %s", valueCoder); | ||
| Coder<?> recordCoder = ((KvCoder<?, ?>) valueCoder).getValueCoder(); | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This kind of casts can be brittel and can be avoided by introducing new types as mentioned in the other comment I believe. |
||
| checkArgument( | ||
| recordCoder instanceof CdcRecordCoder, | ||
| "expected a CdcRecordCoder input record coder, got %s", | ||
| recordCoder); | ||
| return ((CdcRecordCoder) recordCoder).getDataSchema(); | ||
| } | ||
|
|
||
| /** | ||
| * The output of {@link CommitWindows}: the surviving sorted groups ready for the delta writer, | ||
| * and the replayable dead-letter {@link Row}s from late panes. | ||
| */ | ||
| public static final class Result implements POutput { | ||
|
|
||
| private final Pipeline pipeline; | ||
| private final PCollection<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> | ||
| sortedGroups; | ||
| private final PCollection<Row> deadLetterRows; | ||
| private final Schema deadLetterSchema; | ||
|
|
||
| private Result( | ||
| Pipeline pipeline, | ||
| PCollection<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> sortedGroups, | ||
| PCollection<Row> deadLetterRows, | ||
| Schema deadLetterSchema) { | ||
| this.pipeline = pipeline; | ||
| this.sortedGroups = sortedGroups; | ||
| this.deadLetterRows = deadLetterRows; | ||
| this.deadLetterSchema = deadLetterSchema; | ||
| } | ||
|
|
||
| /** The surviving sorted groups: one per {@code (destination, shard, window)}. */ | ||
| public PCollection<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> getSortedGroups() { | ||
| return sortedGroups; | ||
| } | ||
|
|
||
| /** Replayable dead-letter rows from late panes; {@link SplitLateData} describes the shape. */ | ||
| public PCollection<Row> getDeadLetterRows() { | ||
| return deadLetterRows; | ||
| } | ||
|
|
||
| /** The schema of {@link #getDeadLetterRows()}. */ | ||
| public Schema getDeadLetterSchema() { | ||
| return deadLetterSchema; | ||
| } | ||
|
|
||
| @Override | ||
| public Pipeline getPipeline() { | ||
| return pipeline; | ||
| } | ||
|
|
||
| @Override | ||
| public Map<TupleTag<?>, PValue> expand() { | ||
| return ImmutableMap.<TupleTag<?>, PValue>builder() | ||
| .put(ON_TIME_TAG, sortedGroups) | ||
| .put(DEAD_LETTER_TAG, deadLetterRows) | ||
| .build(); | ||
| } | ||
|
|
||
| @Override | ||
| public void finishSpecifyingOutput( | ||
| String transformName, PInput input, PTransform<?, ?> transform) { | ||
| // no-op | ||
| } | ||
| } | ||
| } | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,102 @@ | ||
| /* | ||
| * 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.beam.sdk.io.iceberg.cdc.sink; | ||
|
|
||
| import org.apache.beam.sdk.metrics.Counter; | ||
| import org.apache.beam.sdk.metrics.Metrics; | ||
| import org.apache.beam.sdk.schemas.Schema; | ||
| import org.apache.beam.sdk.transforms.DoFn; | ||
| import org.apache.beam.sdk.transforms.windowing.PaneInfo; | ||
| import org.apache.beam.sdk.values.KV; | ||
| import org.apache.beam.sdk.values.Row; | ||
| import org.apache.beam.sdk.values.TupleTag; | ||
|
|
||
| /** | ||
| * Splits late data in {@link CommitWindows}: every {@linkplain PaneInfo.Timing#LATE late} pane is | ||
| * diverted to a DLQ side output; on-time and early panes pass through unchanged. | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. We can't actually commit late data instead of sending to DLQ ? (probably reasonable but just to double check). |
||
| * | ||
| * <p>Pane history is per {@code (destination, shard, window)} while the committer's skip is per | ||
| * {@code (destination, window)}: a late pane on a shard with no on-time data cannot prove its | ||
| * destination-window uncommitted, and a record let through into a committed window would reach | ||
| * neither the table nor the dead-letter output. | ||
| * | ||
| * <p>Each dead letter nests the untouched data row under {@value #DL_RECORD} beside {@value | ||
| * #DL_CHANGE_TYPE}, {@value #DL_SEQ}, and {@value #DL_DEST}; to replay, unnest {@value #DL_RECORD} | ||
| * and map {@value #DL_CHANGE_TYPE}/{@value #DL_SEQ} as the sink's control columns. Replaying is | ||
| * only safe while no newer change for those keys has committed; a stale replay's equality delete | ||
| * removes the newer row. | ||
| */ | ||
| final class SplitLateData | ||
| extends DoFn< | ||
| KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>, | ||
| KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> { | ||
|
|
||
| static final String DL_RECORD = "record"; | ||
| static final String DL_CHANGE_TYPE = "change_type"; | ||
| static final String DL_SEQ = "sequence_number"; | ||
| static final String DL_DEST = "destination"; | ||
|
|
||
| private final Counter deadLetterRecords = | ||
| Metrics.counter(SplitLateData.class, "deadLetterRecords"); | ||
|
|
||
| private final Schema deadLetterSchema; | ||
| private final TupleTag<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> onTimeTag; | ||
| private final TupleTag<Row> deadLetterTag; | ||
|
|
||
| SplitLateData( | ||
| Schema deadLetterSchema, | ||
| TupleTag<KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>>> onTimeTag, | ||
| TupleTag<Row> deadLetterTag) { | ||
| this.deadLetterSchema = deadLetterSchema; | ||
| this.onTimeTag = onTimeTag; | ||
| this.deadLetterTag = deadLetterTag; | ||
| } | ||
|
|
||
| static Schema deadLetterSchema(Schema cdcDataSchema) { | ||
| return Schema.builder() | ||
| .addRowField(DL_RECORD, cdcDataSchema) | ||
| .addStringField(DL_CHANGE_TYPE) | ||
| .addInt64Field(DL_SEQ) | ||
| .addStringField(DL_DEST) | ||
| .build(); | ||
| } | ||
|
|
||
| @ProcessElement | ||
| public void process( | ||
| @Element KV<KV<String, Integer>, Iterable<KV<byte[], CdcRecord>>> group, | ||
| PaneInfo pane, | ||
| MultiOutputReceiver out) { | ||
| if (pane.getTiming() == PaneInfo.Timing.LATE) { | ||
| String dest = group.getKey().getKey(); | ||
| for (KV<byte[], CdcRecord> kv : group.getValue()) { | ||
| CdcRecord record = kv.getValue(); | ||
| Row deadLetter = | ||
| Row.withSchema(deadLetterSchema) | ||
| .addValue(record.getData()) | ||
| .addValue(record.getKind().name()) | ||
| .addValue(record.getSequenceNumber()) | ||
| .addValue(dest) | ||
| .build(); | ||
| out.get(deadLetterTag).output(deadLetter); | ||
| deadLetterRecords.inc(); | ||
| } | ||
| } else { | ||
| out.get(onTimeTag).output(group); | ||
| } | ||
| } | ||
| } | ||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
PCollection<KV<KV<String, Integer>, KV<byte[], CdcRecord>>>is pretty hard to follow.Shall we introduce new types to simplify this ? For example,
ShardIdentifier->KV<String, Integer>SortKeyAndRecord ->
KV<byte[], CdcRecord>Alternatively we can use Beam Rows or AutoValue classes with schema here.