mxm commented on code in PR #15996: URL: https://github.com/apache/iceberg/pull/15996#discussion_r3099913991
########## flink/v2.1/flink/src/main/java/org/apache/iceberg/flink/maintenance/operator/ReadCommand.java: ########## @@ -0,0 +1,87 @@ +/* + * 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.iceberg.flink.maintenance.operator; + +import java.io.Serializable; +import java.util.List; +import org.apache.flink.annotation.Internal; +import org.apache.iceberg.FileFormat; +import org.apache.iceberg.relocated.com.google.common.collect.Lists; + +/** + * Command from the {@link EqualityConvertPlanner} to the {@link EqualityConvertReader}. Describes a + * single file to read and index. + */ +@Internal +public class ReadCommand implements Serializable { + + public enum Type { + DATA_FILE, + EQ_DELETE_FILE Review Comment: Good point. We handle positional delete files, i.e. we add them as-is to the plan result and merge them back to the main branch alongside with the data files and the DVs (see [planner](https://github.com/apache/iceberg/blob/76ab1d996fc02bacb4d102568252bf6d82c89286/flink/v2.1/flink/src/main/java/org/apache/iceberg/flink/maintenance/operator/EqualityConvertPlanner.java#L376)). While this may be allowed in V3 (not 100% sure), it would be preferable to merge them with the DVs. >Or not even concurrent position deletes, but just position deletes created by the job itself I thought we wouldn't typically write positional deletes, but you are right that we create them even in a CDC job, e.g. when a writer receives an insert + delete. ########## flink/v2.1/flink/src/main/java/org/apache/iceberg/flink/maintenance/api/ConvertEqualityDeletes.java: ########## @@ -0,0 +1,251 @@ +/* + * 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.iceberg.flink.maintenance.api; + +import java.io.IOException; +import java.io.UncheckedIOException; +import org.apache.flink.api.common.typeinfo.TypeInformation; +import org.apache.flink.streaming.api.datastream.DataStream; +import org.apache.flink.streaming.api.datastream.SingleOutputStreamOperator; +import org.apache.iceberg.HasTableOperations; +import org.apache.iceberg.SnapshotRef; +import org.apache.iceberg.Table; +import org.apache.iceberg.flink.TableLoader; +import org.apache.iceberg.flink.maintenance.operator.DVMergeCommand; +import org.apache.iceberg.flink.maintenance.operator.DVMergeResult; +import org.apache.iceberg.flink.maintenance.operator.DVPosition; +import org.apache.iceberg.flink.maintenance.operator.EqualityConvertCommitter; +import org.apache.iceberg.flink.maintenance.operator.EqualityConvertDVMerger; +import org.apache.iceberg.flink.maintenance.operator.EqualityConvertDVResolver; +import org.apache.iceberg.flink.maintenance.operator.EqualityConvertPlanResult; +import org.apache.iceberg.flink.maintenance.operator.EqualityConvertPlanner; +import org.apache.iceberg.flink.maintenance.operator.EqualityConvertReader; +import org.apache.iceberg.flink.maintenance.operator.EqualityConvertWorker; +import org.apache.iceberg.flink.maintenance.operator.IndexCommand; +import org.apache.iceberg.flink.maintenance.operator.ReadCommand; +import org.apache.iceberg.flink.maintenance.operator.SerializedEqualityValues; +import org.apache.iceberg.flink.maintenance.operator.TaskResultAggregator; +import org.apache.iceberg.relocated.com.google.common.base.Preconditions; + +/** + * Creates the equality delete to DV conversion data stream. Runs a single iteration of the + * conversion for every {@link Trigger} event. + * + * <p>The pipeline reads equality delete files from a staging branch, converts them to deletion + * vectors (DVs) using a primary key index stored in Flink state, and commits the data files and DVs + * to the main branch. + * + * <p>The conversion is split into parallel stages: + * + * <ol> + * <li>Planner (p=1): scans staging branch, emits file-level ReadCommands with phase timestamps + * <li>Reader (p=N): reads files, emits row-level IndexCommands + * <li>Worker (p=N): maintains PK index shards, resolves equality deletes to DV positions + * <li>DVResolver (p=1): groups positions by data file, resolves partition info and existing DVs + * <li>DVMerger (p=N): writes merged deletion vector files + * <li>Committer (p=1): commits data files and DVs to main branch + * </ol> + * + * <p>Mutual exclusion with concurrent maintenance tasks (e.g. compaction) is enforced by the Flink + * maintenance framework lock. The lock is acquired before the maintenance task runs and released + * after each cycle completes, ensuring no conflicting commits occur on the target branch. + */ +public class ConvertEqualityDeletes { + static final String PLANNER_TASK_NAME = "EqConvert Planner"; + static final String READER_TASK_NAME = "EqConvert Reader"; + static final String WORKER_TASK_NAME = "EqConvert Worker"; + static final String DV_RESOLVER_TASK_NAME = "EqConvert DVResolver"; + static final String DV_MERGER_TASK_NAME = "EqConvert DVMerger"; + static final String COMMIT_TASK_NAME = "EqConvert Commit"; + static final String AGGREGATOR_TASK_NAME = "EqConvert Aggregator"; + + private ConvertEqualityDeletes() {} + + public static Builder builder() { + return new Builder(); + } + + public static class Builder extends MaintenanceTaskBuilder<Builder> { + private String stagingBranch; + private String targetBranch = SnapshotRef.MAIN_BRANCH; + private int maxSnapshotsPerTrigger = Integer.MAX_VALUE; Review Comment: Yes, when `maxSnapshotsPerTrigger > 1` (the default is `Integer.MAX_VALUE`) multiple staging snapshots are collapsed into one main commit. This works because the resulting deletes are positional. We can handle multiple delete+insert operations. You're right that we must resolve equality deletes one by one for each snapshot to build the correct deletion vectors. To do that, we first build the index from the main branch, then resolve the equality deletes of the first staging snapshot, then continue building the index from the data files of the first snapshot on the staging branch. We use incrementing watermarks to separate these phases. I realize handling multiple snapshots adds some complexity, so it might be worth to drop this feature. We still need the watermark logic to separate building the index and resolving the deletes. -- 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]
