Bigtable change streams to Vector Search template

This template creates a streaming pipeline to stream Bigtable data change records and write them to Vertex AI Vector Search using Dataflow Runner V2.

Pipeline requirements

• The Bigtable source instance must exist.
• The Bigtable source table must exist, and the table must have change streams enabled.
• The Bigtable application profile must exist.
• The Vector Search index path must exist.

Template parameters

Required parameters

• embeddingColumn: The fully qualified column name where the embeddings are stored. In the format cf:col.
• embeddingByteSize: The byte size of each entry in the embeddings array. Use 4 for Float, and 8 for Double. Defaults to: 4.
• vectorSearchIndex: The Vector Search Index where changes will be streamed, in the format 'projects/{projectID}/locations/{region}/indexes/{indexID}' (no leading or trailing spaces) For example, projects/123/locations/us-east1/indexes/456.
• bigtableChangeStreamAppProfile: The Bigtable application profile ID. The application profile must use single-cluster routing and allow single-row transactions.
• bigtableReadInstanceId: The source Bigtable instance ID.
• bigtableReadTableId: The source Bigtable table ID.

Run the template

Console

1. Go to the Dataflow Create job from template page.
Go to Create job from template
2. In the Job name field, enter a unique job name.
3. Optional: For Regional endpoint, select a value from the drop-down menu. The default region is us-central1.

   For a list of regions where you can run a Dataflow job, see Dataflow locations.

4. From the Dataflow template drop-down menu, select the Bigtable Change Streams to Vector Search template.
5. In the provided parameter fields, enter your parameter values.
6. Click Run job.

gcloud CLI

Note: To use the Google Cloud CLI to run flex templates, you must have Google Cloud CLI version 284.0.0 or later.

In your shell or terminal, run the template:

gcloud dataflow flex-template run JOB_NAME \
    --template-file-gcs-location=gs://dataflow-templates-REGION_NAME/VERSION/flex/ \
    --project=PROJECT_ID \
    --region=REGION_NAME \
    --parameters \
       embeddingColumn=EMBEDDING_COLUMN,\
       embeddingByteSize=EMBEDDING_BYTE_SIZE,\
       vectorSearchIndex=VECTOR_SEARCH_INDEX,\
       bigtableChangeStreamAppProfile=BIGTABLE_CHANGE_STREAM_APP_PROFILE,\
       bigtableReadInstanceId=BIGTABLE_READ_INSTANCE_ID,\
       bigtableReadTableId=BIGTABLE_READ_TABLE_ID,\

Replace the following:

• JOB_NAME: a unique job name of your choice
• VERSION: the version of the template that you want to use

  You can use the following values:

  • latest to use the latest version of the template, which is available in the non-dated parent folder in the bucket— gs://dataflow-templates-REGION_NAME/latest/
  • the version name, like 2023-09-12-00_RC00, to use a specific version of the template, which can be found nested in the respective dated parent folder in the bucket— gs://dataflow-templates-REGION_NAME/
  Caution: The latest version of templates might update with breaking changes. Your production environments should use templates kept in the most recent dated parent folder to prevent these breaking changes from affecting your production workflows.
• REGION_NAME: the region where you want to deploy your Dataflow job—for example, us-central1
• EMBEDDING_COLUMN: the Embedding column
• EMBEDDING_BYTE_SIZE: the The byte size of the embeddings array. Can be 4 or 8.
• VECTOR_SEARCH_INDEX: the Vector Search index Path
• BIGTABLE_CHANGE_STREAM_APP_PROFILE: the Bigtable application profile ID
• BIGTABLE_READ_INSTANCE_ID: the source Bigtable Instance ID
• BIGTABLE_READ_TABLE_ID: the source Bigtable table ID

API

To run the template using the REST API, send an HTTP POST request. For more information on the API and its authorization scopes, see projects.templates.launch.

POST https://dataflow.googleapis.com/v1b3/projects/PROJECT_ID/locations/LOCATION/flexTemplates:launch
{
   "launchParameter": {
     "jobName": "JOB_NAME",
     "parameters": {
       "embeddingColumn": "EMBEDDING_COLUMN",
       "embeddingByteSize": "EMBEDDING_BYTE_SIZE",
       "vectorSearchIndex": "VECTOR_SEARCH_INDEX",
       "bigtableChangeStreamAppProfile": "BIGTABLE_CHANGE_STREAM_APP_PROFILE",
       "bigtableReadInstanceId": "BIGTABLE_READ_INSTANCE_ID",
       "bigtableReadTableId": "BIGTABLE_READ_TABLE_ID",
     },
     "containerSpecGcsPath": "gs://dataflow-templates-LOCATION/VERSION/flex/",
     "environment": { "maxWorkers": "10" }
  }
}

Replace the following:

• PROJECT_ID: the Google Cloud project ID where you want to run the Dataflow job
• JOB_NAME: a unique job name of your choice
• VERSION: the version of the template that you want to use

  You can use the following values:

  • latest to use the latest version of the template, which is available in the non-dated parent folder in the bucket— gs://dataflow-templates-REGION_NAME/latest/
  • the version name, like 2023-09-12-00_RC00, to use a specific version of the template, which can be found nested in the respective dated parent folder in the bucket— gs://dataflow-templates-REGION_NAME/
  Caution: The latest version of templates might update with breaking changes. Your production environments should use templates kept in the most recent dated parent folder to prevent these breaking changes from affecting your production workflows.
• LOCATION: the region where you want to deploy your Dataflow job—for example, us-central1
• EMBEDDING_COLUMN: the Embedding column
• EMBEDDING_BYTE_SIZE: the The byte size of the embeddings array. Can be 4 or 8.
• VECTOR_SEARCH_INDEX: the Vector Search index Path
• BIGTABLE_CHANGE_STREAM_APP_PROFILE: the Bigtable application profile ID
• BIGTABLE_READ_INSTANCE_ID: the source Bigtable Instance ID
• BIGTABLE_READ_TABLE_ID: the source Bigtable table ID

Template source code

Java

/*
 * Copyright (C) 2024 Google LLC
 *
 * Licensed 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 com.google.cloud.teleport.v2.templates.bigtablechangestreamstovectorsearch;

import com.google.cloud.Timestamp;
import com.google.cloud.aiplatform.v1.IndexDatapoint;
import com.google.cloud.teleport.metadata.Template;
import com.google.cloud.teleport.metadata.TemplateCategory;
import com.google.cloud.teleport.v2.bigtable.options.BigtableCommonOptions.ReadChangeStreamOptions;
import com.google.cloud.teleport.v2.bigtable.options.BigtableCommonOptions.ReadOptions;
import com.google.cloud.teleport.v2.cdc.dlq.DeadLetterQueueManager;
import com.google.cloud.teleport.v2.options.BigtableChangeStreamsToVectorSearchOptions;
import com.google.cloud.teleport.v2.transforms.DLQWriteTransform;
import com.google.cloud.teleport.v2.utils.DurationUtils;
import java.io.IOException;
import java.util.ArrayList;
import java.util.List;
import org.apache.beam.runners.dataflow.options.DataflowPipelineOptions;
import org.apache.beam.sdk.Pipeline;
import org.apache.beam.sdk.PipelineResult;
import org.apache.beam.sdk.io.gcp.bigtable.BigtableIO;
import org.apache.beam.sdk.options.PipelineOptionsFactory;
import org.apache.beam.sdk.transforms.GroupIntoBatches;
import org.apache.beam.sdk.transforms.ParDo;
import org.apache.beam.sdk.transforms.Values;
import org.apache.beam.sdk.transforms.WithKeys;
import org.apache.beam.sdk.values.PCollectionTuple;
import org.apache.beam.sdk.values.TupleTagList;
import org.apache.commons.lang3.StringUtils;
import org.joda.time.Duration;
import org.joda.time.Instant;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

@Template(
    name = "Bigtable_Change_Streams_to_Vector_Search",
    category = TemplateCategory.STREAMING,
    displayName = "Bigtable Change Streams to Vector Search",
    description =
        "Streaming pipeline. Streams Bigtable data change records and writes them into Vertex AI Vector Search using Dataflow Runner V2.",
    optionsClass = BigtableChangeStreamsToVectorSearchOptions.class,
    optionsOrder = {
      BigtableChangeStreamsToVectorSearchOptions.class,
      ReadChangeStreamOptions.class,
      ReadOptions.class
    },
    skipOptions = {
      "bigtableReadAppProfile",
      "bigtableAdditionalRetryCodes",
      "bigtableRpcAttemptTimeoutMs",
      "bigtableRpcTimeoutMs"
    },
    documentation =
        "https://cloud.google.com/dataflow/docs/guides/templates/provided/bigtable-change-streams-to-vector-search",
    flexContainerName = "googlecloud-to-googlecloud",
    contactInformation = "https://cloud.google.com/support",
    streaming = true)
public final class BigtableChangeStreamsToVectorSearch {
  private static final Logger LOG =
      LoggerFactory.getLogger(BigtableChangeStreamsToVectorSearch.class);

  private static final String USE_RUNNER_V2_EXPERIMENT = "use_runner_v2";

  /**
   * Main entry point for executing the pipeline.
   *
   * @param args The command-line arguments to the pipeline.
   */
  public static void main(String[] args) throws Exception {
    LOG.info("Starting replication from Cloud Bigtable Change Streams to Vector Search");

    BigtableChangeStreamsToVectorSearchOptions options =
        PipelineOptionsFactory.fromArgs(args)
            .withValidation()
            .as(BigtableChangeStreamsToVectorSearchOptions.class);

    run(options);
  }

  public static PipelineResult run(BigtableChangeStreamsToVectorSearchOptions options)
      throws IOException {
    options.setStreaming(true);
    options.setEnableStreamingEngine(true);

    List<String> experiments = options.getExperiments();
    if (experiments == null) {
      experiments = new ArrayList<>();
    }
    boolean hasUseRunnerV2 = false;
    for (String experiment : experiments) {
      if (experiment.equalsIgnoreCase(USE_RUNNER_V2_EXPERIMENT)) {
        hasUseRunnerV2 = true;
        break;
      }
    }
    if (!hasUseRunnerV2) {
      experiments.add(USE_RUNNER_V2_EXPERIMENT);
    }
    options.setExperiments(experiments);

    Instant startTimestamp =
        options.getBigtableChangeStreamStartTimestamp().isEmpty()
            ? Instant.now()
            : toInstant(Timestamp.parseTimestamp(options.getBigtableChangeStreamStartTimestamp()));

    String bigtableProjectId = getBigtableProjectId(options);

    LOG.info("  - startTimestamp {}", startTimestamp);
    LOG.info("  - bigtableReadInstanceId {}", options.getBigtableReadInstanceId());
    LOG.info("  - bigtableReadTableId {}", options.getBigtableReadTableId());
    LOG.info("  - bigtableChangeStreamAppProfile {}", options.getBigtableChangeStreamAppProfile());
    LOG.info("  - embeddingColumn {}", options.getEmbeddingColumn());
    LOG.info("  - crowdingTagColumn {}", options.getCrowdingTagColumn());
    LOG.info("  - project {}", options.getProject());
    LOG.info("  - indexName {}", options.getVectorSearchIndex());

    String indexName = options.getVectorSearchIndex();

    String vertexRegion = Utils.extractRegionFromIndexName(indexName);
    String vertexEndpoint = vertexRegion + "-aiplatform.googleapis.com:443";

    final Pipeline pipeline = Pipeline.create(options);

    DeadLetterQueueManager dlqManager = buildDlqManager(options);

    BigtableIO.ReadChangeStream readChangeStream =
        BigtableIO.readChangeStream()
            .withChangeStreamName(options.getBigtableChangeStreamName())
            .withExistingPipelineOptions(
                options.getBigtableChangeStreamResume()
                    ? BigtableIO.ExistingPipelineOptions.RESUME_OR_FAIL
                    : BigtableIO.ExistingPipelineOptions.FAIL_IF_EXISTS)
            .withProjectId(bigtableProjectId)
            .withAppProfileId(options.getBigtableChangeStreamAppProfile())
            .withInstanceId(options.getBigtableReadInstanceId())
            .withTableId(options.getBigtableReadTableId())
            .withMetadataTableInstanceId(options.getBigtableChangeStreamMetadataInstanceId())
            .withMetadataTableTableId(options.getBigtableMetadataTableTableId())
            .withStartTime(startTimestamp);

    PCollectionTuple results =
        pipeline
            .apply("Read from Cloud Bigtable Change Streams", readChangeStream)
            .apply("Create Values", Values.create())
            .apply(
                "Converting to Vector Search Datapoints",
                ParDo.of(
                        new ChangeStreamMutationToDatapointOperationFn(
                            options.getEmbeddingColumn(),
                            options.getEmbeddingByteSize(),
                            options.getCrowdingTagColumn(),
                            Utils.parseColumnMapping(options.getAllowRestrictsMappings()),
                            Utils.parseColumnMapping(options.getDenyRestrictsMappings()),
                            Utils.parseColumnMapping(options.getIntNumericRestrictsMappings()),
                            Utils.parseColumnMapping(options.getFloatNumericRestrictsMappings()),
                            Utils.parseColumnMapping(options.getDoubleNumericRestrictsMappings())))
                    .withOutputTags(
                        ChangeStreamMutationToDatapointOperationFn.UPSERT_DATAPOINT_TAG,
                        TupleTagList.of(
                            ChangeStreamMutationToDatapointOperationFn.REMOVE_DATAPOINT_TAG)));
    results
        .get(ChangeStreamMutationToDatapointOperationFn.UPSERT_DATAPOINT_TAG)
        .apply("Add placeholer keys", WithKeys.of("placeholder"))
        .apply(
            "Batch Contents",
            GroupIntoBatches.<String, IndexDatapoint>ofSize(
                    bufferSizeOption(options.getUpsertMaxBatchSize()))
                .withMaxBufferingDuration(
                    bufferDurationOption(options.getUpsertMaxBufferDuration())))
        .apply("Map to Values", Values.create())
        .apply(
            "Upsert Datapoints to VectorSearch",
            ParDo.of(new UpsertDatapointsFn(vertexEndpoint, indexName)))
        .apply(
            "Write errors to DLQ",
            DLQWriteTransform.WriteDLQ.newBuilder()
                .withDlqDirectory(dlqManager.getSevereDlqDirectory() + "YYYY/MM/dd/HH/mm/")
                .withTmpDirectory(dlqManager.getSevereDlqDirectory() + "tmp/")
                .setIncludePaneInfo(true)
                .build());

    results
        .get(ChangeStreamMutationToDatapointOperationFn.REMOVE_DATAPOINT_TAG)
        .apply("Add placeholder keys", WithKeys.of("placeholer"))
        .apply(
            "Batch Contents",
            GroupIntoBatches.<String, String>ofSize(
                    bufferSizeOption(options.getDeleteMaxBatchSize()))
                .withMaxBufferingDuration(
                    bufferDurationOption(options.getDeleteMaxBufferDuration())))
        .apply("Map to Values", Values.create())
        .apply(
            "Remove Datapoints From VectorSearch",
            ParDo.of(new RemoveDatapointsFn(vertexEndpoint, indexName)))
        .apply(
            "Write errors to DLQ",
            DLQWriteTransform.WriteDLQ.newBuilder()
                .withDlqDirectory(dlqManager.getSevereDlqDirectory() + "YYYY/MM/dd/HH/mm/")
                .withTmpDirectory(dlqManager.getSevereDlqDirectory() + "tmp/")
                .setIncludePaneInfo(true)
                .build());

    return pipeline.run();
  }

  private static String getBigtableProjectId(BigtableChangeStreamsToVectorSearchOptions options) {
    return StringUtils.isEmpty(options.getBigtableReadProjectId())
        ? options.getProject()
        : options.getBigtableReadProjectId();
  }

  private static Instant toInstant(Timestamp timestamp) {
    if (timestamp == null) {
      return null;
    } else {
      return Instant.ofEpochMilli(timestamp.getSeconds() * 1000 + timestamp.getNanos() / 1000000);
    }
  }

  private static int bufferSizeOption(int size) {
    if (size < 1) {
      size = 1;
    }

    return size;
  }

  private static Duration bufferDurationOption(String duration) {
    if (duration.isEmpty()) {
      return Duration.standardSeconds(1);
    }

    return DurationUtils.parseDuration(duration);
  }

  private static DeadLetterQueueManager buildDlqManager(
      BigtableChangeStreamsToVectorSearchOptions options) {
    String dlqDirectory = options.getDlqDirectory();
    if (dlqDirectory.isEmpty()) {
      LOG.info("Falling back to temp dir for DLQ");

      String tempLocation = options.as(DataflowPipelineOptions.class).getTempLocation();

      LOG.info("Have temp location {}", tempLocation);
      if (tempLocation == null || tempLocation.isEmpty()) {
        tempLocation = "/";
      } else if (!tempLocation.endsWith("/")) {
        tempLocation += "/";
      }

      dlqDirectory = tempLocation + "dlq";
    }

    LOG.info("Writing dead letter queue to: {}", dlqDirectory);

    return DeadLetterQueueManager.create(dlqDirectory, 1);
  }
}