Modelo de pesquisa de vetor do Spanner para a Vertex AI

O modelo do Spanner para arquivos de pesquisa de vetor da Vertex AI no Cloud Storage cria um pipeline em lote que exporta dados de embeddings vetoriais de uma tabela do Spanner para o Cloud Storage no formato JSON. Use os parâmetros do modelo para especificar a pasta do Cloud Storage para exportar os embeddings vetoriais. A pasta do Cloud Storage contém a lista de arquivos .json exportados, que representam os embeddings vetoriais em um formato aceito pelo índice Vector Search da Vertex AI.

Para mais informações, consulte Formato e estrutura dos dados de entrada.

Requisitos de pipeline

  • O banco de dados do Spanner precisa existir.
  • É necessário ter um bucket do Cloud Storage para a exibição de dados.
  • Além dos papéis do Identity and Access Management (IAM) necessários para executar jobs do Dataflow, você precisa dos papéis do IAM necessários para ler os dados do Spanner e gravar no bucket do Cloud Storage.

Parâmetros do modelo

Parâmetros obrigatórios

  • spannerProjectId: o ID do projeto da instância do Spanner.
  • spannerInstanceId: o ID da instância do Spanner de onde os embeddings vetoriais serão exportados.
  • spannerDatabaseId: o ID do banco de dados do Spanner para exportar os embeddings vetoriais.
  • spannerTable: a tabela do Spanner para leitura.
  • spannerColumnsToExport: uma lista separada por vírgulas de colunas obrigatórias para o índice de pesquisa de vetor da Vertex AI. As colunas de ID e de incorporação são obrigatórias para a pesquisa vetorial. Se os nomes das colunas não corresponderem à estrutura de entrada do índice da Pesquisa de vetor da Vertex AI, crie mapeamentos de coluna usando aliases. Se os nomes das colunas não corresponderem ao formato esperado pela Vertex AI, use a notação de:a. Por exemplo, se você tiver colunas com os nomes "id" e "my_embedding", especifique "id, my_embedding:embedding".
  • gcsOutputFolder: a pasta do Cloud Storage para gravar arquivos de saída. Esse caminho precisa terminar com uma barra. Por exemplo, gs://your-bucket/folder1/.
  • gcsOutputFilePrefix: o prefixo do nome de arquivo para gravar arquivos de saída. Por exemplo, vector-embeddings.

Parâmetros opcionais

  • spannerHost: o endpoint do Spanner a ser chamado no modelo. O valor padrão é https://batch-spanner.googleapis.com. Por exemplo, https://batch-spanner.googleapis.com.
  • spannerVersionTime: se definido, especifica o horário em que a versão do banco de dados precisa ser usada. O valor é uma string no formato de data RFC-3339 em tempo de época Unix. Por exemplo, 1990-12-31T23:59:60Z. O carimbo de data/hora precisa estar no passado, e a inatividade máxima (https://cloud.google.com/spanner/docs/timestamp-bounds#maximum_timestamp_staleness) é aplicável. Se não for definido, um limite forte (https://cloud.google.com/spanner/docs/timestamp-bounds#strong) será usado para ler os dados mais recentes. O padrão é empty. Por exemplo, 1990-12-31T23:59:60Z.
  • spannerDataBoostEnabled: quando definido como true, o modelo usa a computação sob demanda do Spanner. O job de exportação é executado em recursos de computação independentes que não afetam as cargas de trabalho atuais do Spanner. O uso dessa opção gera cobranças adicionais no Spanner. Para mais informações, consulte a visão geral do Data Boost do Spanner (https://cloud.google.com/spanner/docs/databoost/databoost-overview). O padrão é false.
  • spannerPriority: a prioridade da solicitação para chamadas do Spanner. Os valores permitidos são HIGH, MEDIUM e LOW. O valor padrão é MEDIUM.

Executar o modelo

  1. Acesse a página Criar job usando um modelo do Dataflow.
  2. Acesse Criar job usando um modelo
  3. No campo Nome do job, insira um nome exclusivo.
  4. Opcional: em Endpoint regional, selecione um valor no menu suspenso. A região padrão é us-central1.

    Para ver uma lista de regiões em que é possível executar um job do Dataflow, consulte Locais do Dataflow.

  5. No menu suspenso Modelo do Dataflow, selecione the Spanner to Vertex AI Vector Search files on Cloud Storage template.
  6. Nos campos de parâmetro fornecidos, insira os valores de parâmetro.
  7. Cliquem em Executar job.

No shell ou no terminal, execute o modelo:

gcloud dataflow jobs run JOB_NAME \
    --gcs-location=gs://dataflow-templates-REGION_NAME/VERSION/Cloud_Spanner_vectors_to_Cloud_Storage \
    --project=PROJECT_ID \
    --region=REGION_NAME \
    --parameters \
       spannerProjectId=SPANNER_PROJECT_ID,\
       spannerInstanceId=SPANNER_INSTANCE_ID,\
       spannerDatabaseId=SPANNER_DATABASE_ID,\
       spannerTable=SPANNER_TABLE,\
       spannerColumnsToExport=SPANNER_COLUMNS_TO_EXPORT,\
       gcsOutputFolder=GCS_OUTPUT_FOLDER,\
       gcsOutputFilePrefix=GCS_OUTPUT_FILE_PREFIX,\

Substitua:

  • JOB_NAME: um nome de job de sua escolha
  • VERSION: a versão do modelo que você quer usar

    Use estes valores:

  • REGION_NAME: a região em que você quer implantar o job do Dataflow, por exemplo, us-central1
  • SPANNER_PROJECT_ID: o ID do projeto do Spanner
  • SPANNER_INSTANCE_ID: o ID da instância do Spanner
  • SPANNER_DATABASE_ID: o ID do banco de dados do Spanner
  • SPANNER_TABLE: a tabela do Spanner
  • SPANNER_COLUMNS_TO_EXPORT: as colunas a serem exportadas da tabela do Spanner
  • GCS_OUTPUT_FOLDER: a pasta do Cloud Storage para enviar arquivos
  • GCS_OUTPUT_FILE_PREFIX: o prefixo dos arquivos de saída no Cloud Storage

Para executar o modelo usando a API REST, envie uma solicitação HTTP POST. Para mais informações sobre a API e os respectivos escopos de autorização, consulte projects.templates.launch.

POST https://dataflow.googleapis.com/v1b3/projects/PROJECT_ID/locations/LOCATION/templates:launch?gcsPath=gs://dataflow-templates-LOCATION/VERSION/Cloud_Spanner_vectors_to_Cloud_Storage
{
   "jobName": "JOB_NAME",
   "parameters": {
     "spannerProjectId": "SPANNER_PROJECT_ID",
     "spannerInstanceId": "SPANNER_INSTANCE_ID",
     "spannerDatabaseId": "SPANNER_DATABASE_ID",
     "spannerTable": "SPANNER_TABLE",
     "spannerColumnsToExport": "SPANNER_COLUMNS_TO_EXPORT",
     "gcsOutputFolder": "GCS_OUTPUT_FOLDER",
     "gcsOutputFilePrefix": "GCS_OUTPUT_FILE_PREFIX",
   },
   "environment": { "maxWorkers": "10" }
}

Substitua:

  • PROJECT_ID: o ID do projeto do Google Cloud em que você quer executar o job do Dataflow
  • JOB_NAME: um nome de job de sua escolha
  • VERSION: a versão do modelo que você quer usar

    Use estes valores:

  • LOCATION: a região em que você quer implantar o job do Dataflow, por exemplo, us-central1
  • SPANNER_PROJECT_ID: o ID do projeto do Spanner
  • SPANNER_INSTANCE_ID: o ID da instância do Spanner
  • SPANNER_DATABASE_ID: o ID do banco de dados do Spanner
  • SPANNER_TABLE: a tabela do Spanner
  • SPANNER_COLUMNS_TO_EXPORT: as colunas a serem exportadas da tabela do Spanner
  • GCS_OUTPUT_FOLDER: a pasta do Cloud Storage para enviar arquivos
  • GCS_OUTPUT_FILE_PREFIX: o prefixo dos arquivos de saída no Cloud Storage
Java
/*
 * Copyright (C) 2023 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.templates;

import com.google.cloud.spanner.Options.RpcPriority;
import com.google.cloud.teleport.metadata.Template;
import com.google.cloud.teleport.metadata.TemplateCategory;
import com.google.cloud.teleport.metadata.TemplateParameter;
import com.google.cloud.teleport.metadata.TemplateParameter.TemplateEnumOption;
import com.google.cloud.teleport.templates.SpannerVectorEmbeddingExport.SpannerToVectorEmbeddingJsonOptions;
import com.google.cloud.teleport.templates.common.SpannerConverters;
import com.google.cloud.teleport.templates.common.SpannerConverters.CreateTransactionFnWithTimestamp;
import com.google.cloud.teleport.templates.common.SpannerConverters.VectorSearchStructValidator;
import org.apache.beam.sdk.Pipeline;
import org.apache.beam.sdk.io.FileSystems;
import org.apache.beam.sdk.io.TextIO;
import org.apache.beam.sdk.io.gcp.spanner.LocalSpannerIO;
import org.apache.beam.sdk.io.gcp.spanner.ReadOperation;
import org.apache.beam.sdk.io.gcp.spanner.SpannerConfig;
import org.apache.beam.sdk.io.gcp.spanner.Transaction;
import org.apache.beam.sdk.options.Default;
import org.apache.beam.sdk.options.PipelineOptions;
import org.apache.beam.sdk.options.PipelineOptionsFactory;
import org.apache.beam.sdk.options.ValueProvider;
import org.apache.beam.sdk.transforms.Create;
import org.apache.beam.sdk.transforms.MapElements;
import org.apache.beam.sdk.transforms.PTransform;
import org.apache.beam.sdk.transforms.ParDo;
import org.apache.beam.sdk.transforms.SerializableFunction;
import org.apache.beam.sdk.transforms.View;
import org.apache.beam.sdk.values.PBegin;
import org.apache.beam.sdk.values.PCollection;
import org.apache.beam.sdk.values.PCollectionView;
import org.apache.beam.sdk.values.TypeDescriptors;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

/**
 * Dataflow template which export vector embeddings from Spanner to GCS in json format. It exports a
 * Spanner table using <a
 * href="https://cloud.google.com/spanner/docs/reads#read_data_in_parallel">Batch API</a>, which
 * creates multiple workers in parallel for better performance. The result is written to a JSON file
 * in Google Cloud Storage.
 *
 * <p>Check out <a
 * href="https://github.com/GoogleCloudPlatform/DataflowTemplates/blob/main/v1/README_Cloud_Spanner_to_Vector_Embedding.md">README</a>
 * for instructions on how to use or modify this template.
 */
@Template(
    name = "Cloud_Spanner_vectors_to_Cloud_Storage",
    category = TemplateCategory.BATCH,
    displayName = "Cloud Spanner vectors to Cloud Storage for Vertex Vector Search",
    optionsClass = SpannerToVectorEmbeddingJsonOptions.class,
    description = {
      "The Cloud Spanner to Vector Embeddings on Cloud Storage template is a batch pipeline that exports vector embeddings data from Cloud Spanner's table to Cloud Storage in JSON format. "
          + "Vector embeddings are exported to a Cloud Storage folder specified by the user in the template parameters."
          + " The Cloud Storage folder will contain the list of exported `.json` files representing vector embeddings in a format supported by Vertex AI Vector Search Index.\n",
      "Check <a href=\"https://cloud.google.com/vertex-ai/docs/vector-search/setup/format-structure#json\">Vector Search Format Structure</a> for additional details."
    },
    documentation =
        "https://cloud.google.com/dataflow/docs/guides/templates/provided/cloud-spanner-to-vertex-vector-search",
    contactInformation = "https://cloud.google.com/support",
    requirements = {
      "The Cloud Spanner database must exist.",
      "The output Cloud Storage bucket must exist.",
      "In addition to the Identity and Access Management (IAM) roles necessary to run Dataflow jobs, you must also have the <a href=\"https://cloud.google.com/spanner/docs/export#iam\">appropriate IAM roles</a> for reading your Cloud Spanner data and writing to your Cloud Storage bucket."
    })
@SuppressWarnings("unused")
public class SpannerVectorEmbeddingExport {

  private static final Logger LOG = LoggerFactory.getLogger(SpannerVectorEmbeddingExport.class);

  /** Custom PipelineOptions. */
  public interface SpannerToVectorEmbeddingJsonOptions extends PipelineOptions {
    @TemplateParameter.ProjectId(
        order = 10,
        groupName = "Source",
        description = "Cloud Spanner Project Id",
        helpText = "The project ID of the Spanner instance.")
    ValueProvider<String> getSpannerProjectId();

    void setSpannerProjectId(ValueProvider<String> value);

    @TemplateParameter.Text(
        order = 20,
        groupName = "Source",
        regexes = {"[a-z][a-z0-9\\-]*[a-z0-9]"},
        description = "Cloud Spanner instance ID",
        helpText = "The ID of the Spanner instance to export the vector embeddings from.")
    ValueProvider<String> getSpannerInstanceId();

    void setSpannerInstanceId(ValueProvider<String> spannerInstanceId);

    @TemplateParameter.Text(
        order = 30,
        groupName = "Source",
        regexes = {"[a-z][a-z0-9_\\-]*[a-z0-9]"},
        description = "Cloud Spanner database ID",
        helpText = "The ID of the Spanner database to export the vector embeddings from.")
    ValueProvider<String> getSpannerDatabaseId();

    void setSpannerDatabaseId(ValueProvider<String> spannerDatabaseId);

    @TemplateParameter.Text(
        order = 40,
        groupName = "Source",
        regexes = {"^.+$"},
        description = "Spanner Table",
        helpText = "The Spanner table to read from.")
    ValueProvider<String> getSpannerTable();

    void setSpannerTable(ValueProvider<String> table);

    @TemplateParameter.Text(
        order = 50,
        groupName = "Source",
        description = "Columns to Export from Spanner Table",
        helpText =
            "A comma-separated list of required columns for the Vertex AI Vector Search index. The ID and embedding columns are required by Vector Search. If your column names don't match the Vertex AI Vector Search index input structure, create column mappings by using aliases. If the column names don't match the format expected by Vertex AI, use the notation from:to. For example, if you have columns named id and my_embedding, specify id, my_embedding:embedding.")
    ValueProvider<String> getSpannerColumnsToExport();

    void setSpannerColumnsToExport(ValueProvider<String> value);

    @TemplateParameter.GcsWriteFolder(
        order = 60,
        groupName = "Target",
        description = "Output files folder in Cloud Storage",
        helpText =
            "The Cloud Storage folder to write output files to. The path must end with a slash.",
        example = "gs://your-bucket/folder1/")
    ValueProvider<String> getGcsOutputFolder();

    void setGcsOutputFolder(ValueProvider<String> value);

    @TemplateParameter.Text(
        order = 70,
        groupName = "Target",
        description = "Output files prefix in Cloud Storage",
        helpText = "The filename prefix for writing output files.",
        example = "vector-embeddings")
    ValueProvider<String> getGcsOutputFilePrefix();

    void setGcsOutputFilePrefix(ValueProvider<String> textWritePrefix);

    @TemplateParameter.Text(
        order = 80,
        groupName = "Source",
        optional = true,
        description = "Cloud Spanner Endpoint to call",
        helpText =
            "The Spanner endpoint to call in the template. The default value is https://batch-spanner.googleapis.com.",
        example = "https://batch-spanner.googleapis.com")
    @Default.String("https://batch-spanner.googleapis.com")
    ValueProvider<String> getSpannerHost();

    void setSpannerHost(ValueProvider<String> value);

    @TemplateParameter.Text(
        order = 90,
        groupName = "Source",
        optional = true,
        regexes = {
          "^([0-9]{4})-([0-9]{2})-([0-9]{2})T([0-9]{2}):([0-9]{2}):(([0-9]{2})(\\.[0-9]+)?)Z$"
        },
        description = "Timestamp to read stale data from a version in the past.",
        helpText =
            "If set, specifies the time when the database version must be taken. The value is a string in the RFC-3339 date format in Unix epoch time. For example: `1990-12-31T23:59:60Z`. The timestamp must be in the past, and maximum timestamp staleness (https://cloud.google.com/spanner/docs/timestamp-bounds#maximum_timestamp_staleness) applies. If not set, a strong bound (https://cloud.google.com/spanner/docs/timestamp-bounds#strong) is used to read the latest data. Defaults to `empty`.",
        example = "1990-12-31T23:59:60Z")
    @Default.String(value = "")
    ValueProvider<String> getSpannerVersionTime();

    void setSpannerVersionTime(ValueProvider<String> value);

    @TemplateParameter.Boolean(
        order = 100,
        groupName = "Source",
        optional = true,
        description = "Use independent compute resource (Spanner DataBoost).",
        helpText =
            "When set to `true`, the template uses Spanner on-demand compute. The export job runs on independent compute resources that don't impact current Spanner workloads. Using this option incurs additional charges in Spanner. For more information, see Spanner Data Boost overview (https://cloud.google.com/spanner/docs/databoost/databoost-overview). Defaults to: `false`.")
    @Default.Boolean(false)
    ValueProvider<Boolean> getSpannerDataBoostEnabled();

    void setSpannerDataBoostEnabled(ValueProvider<Boolean> value);

    @TemplateParameter.Enum(
        order = 110,
        groupName = "Source",
        enumOptions = {
          @TemplateEnumOption("LOW"),
          @TemplateEnumOption("MEDIUM"),
          @TemplateEnumOption("HIGH")
        },
        optional = true,
        description = "Priority for Spanner RPC invocations",
        helpText =
            "The request priority for Spanner calls. The allowed values are `HIGH`, `MEDIUM`, and `LOW`. The default value is `MEDIUM`.")
    ValueProvider<RpcPriority> getSpannerPriority();

    void setSpannerPriority(ValueProvider<RpcPriority> value);
  }

  /**
   * Runs a pipeline which reads in vector embeddings records from Spanner, and writes the JSON to
   * TextIO sink.
   *
   * @param args arguments to the pipeline
   */
  public static void main(String[] args) {
    LOG.info("Starting pipeline setup");
    PipelineOptionsFactory.register(SpannerToVectorEmbeddingJsonOptions.class);

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

    FileSystems.setDefaultPipelineOptions(options);
    Pipeline pipeline = Pipeline.create(options);

    SpannerConfig spannerConfig =
        SpannerConfig.create()
            .withHost(options.getSpannerHost())
            .withProjectId(options.getSpannerProjectId())
            .withInstanceId(options.getSpannerInstanceId())
            .withDatabaseId(options.getSpannerDatabaseId())
            .withRpcPriority(options.getSpannerPriority())
            .withDataBoostEnabled(options.getSpannerDataBoostEnabled());

    ValueProvider<String> gcsOutputFilePrefix = options.getGcsOutputFilePrefix();

    // Concatenating cloud storage folder with file prefix to get complete path
    ValueProvider<String> gcsOutputFilePathWithPrefix =
        ValueProvider.NestedValueProvider.of(
            options.getGcsOutputFolder(),
            (SerializableFunction<String, String>)
                folder -> {
                  if (!folder.endsWith("/")) {
                    // Appending the slash if not provided by user
                    folder = folder + "/";
                  }
                  return folder + gcsOutputFilePrefix.get();
                });

    PTransform<PBegin, PCollection<ReadOperation>> spannerExport =
        SpannerConverters.ExportTransformFactory.create(
            options.getSpannerTable(),
            spannerConfig,
            gcsOutputFilePathWithPrefix,
            options.getSpannerVersionTime(),
            options.getSpannerColumnsToExport(),
            ValueProvider.StaticValueProvider.of(/* disable_schema_export= */ false));

    /* CreateTransaction and CreateTransactionFn classes in LocalSpannerIO
     * only take a timestamp object for exact staleness which works when
     * parameters are provided during template compile time. They do not work with
     * a Timestamp valueProvider which can take parameters at runtime. Hence a new
     * ParDo class CreateTransactionFnWithTimestamp had to be created for this
     * purpose.
     */
    PCollectionView<Transaction> tx =
        pipeline
            .apply("Setup for Transaction", Create.of(1))
            .apply(
                "Create transaction",
                ParDo.of(
                    new CreateTransactionFnWithTimestamp(
                        spannerConfig, options.getSpannerVersionTime())))
            .apply("As PCollectionView", View.asSingleton());

    PCollection<String> json =
        pipeline
            .apply("Create export", spannerExport)
            // We need to use LocalSpannerIO.readAll() instead of LocalSpannerIO.read()
            // because ValueProvider parameters such as table name required for
            // LocalSpannerIO.read() can be read only inside DoFn but LocalSpannerIO.read() is of
            // type PTransform<PBegin, Struct>, which prevents prepending it with DoFn that reads
            // these parameters at the pipeline execution time.
            .apply(
                "Read all records",
                LocalSpannerIO.readAll().withTransaction(tx).withSpannerConfig(spannerConfig))
            .apply(
                "Struct To JSON",
                MapElements.into(TypeDescriptors.strings())
                    .via(
                        struct ->
                            (new SpannerConverters.StructJSONPrinter(
                                    new VectorSearchStructValidator()))
                                .print(struct)));

    json.apply(
        "Write to storage", TextIO.write().to(gcsOutputFilePathWithPrefix).withSuffix(".json"));

    pipeline.run();
    LOG.info("Completed pipeline setup");
  }
}

A seguir