使用 Dataflow 建立變更串流連線

透過集合功能整理內容 你可以依據偏好儲存及分類內容。

<dependency>
  <groupId>org.apache.beam</groupId>
  <artifactId>beam-sdks-java-io-google-cloud-platform</artifactId>
  <version>${beam-version}</version> <!-- available from version 2.38.0 -->
</dependency>

SpannerConfig spannerConfig = SpannerConfig
  .create()
  .withProjectId("my-project-id")
  .withInstanceId("my-instance-id")
  .withDatabaseId("my-database-id")
  .withDatabaseRole("my-database-role");    // Needed for fine-grained access control only

Timestamp startTime = Timestamp.now();
Timestamp endTime = Timestamp.ofTimeSecondsAndNanos(
   startTime.getSeconds() + (10 * 60),
   startTime.getNanos()
);

SpannerIO
  .readChangeStream()
  .withSpannerConfig(spannerConfig)
  .withChangeStreamName("my-change-stream")
  .withMetadataInstance("my-meta-instance-id")
  .withMetadataDatabase("my-meta-database-id")
  .withMetadataTable("my-meta-table-name")
  .withRpcPriority(RpcPriority.MEDIUM)
  .withInclusiveStartAt(startTime)
  .withInclusiveEndAt(endTime);

pipeline
  .apply(SpannerIO
    .readChangeStream()
    .withSpannerConfig(SpannerConfig
      .create()
      .withProjectId("my-project-id")
      .withInstanceId("my-instance-id")
      .withDatabaseId("my-database-id")
      .withDatabaseRole("my-database-role"))    // Needed for fine-grained access control only
    .withChangeStreamName("my-change-stream")
    .withMetadataInstance("my-metadata-instance-id")
    .withMetadataDatabase("my-metadata-database-id")
    .withInclusiveStartAt(Timestamp.now()))
  .apply(ParDo.of(new BreakRecordByModFn()))
  .apply(ParDo.of(new KeyByIdFn()))
  .apply(ParDo.of(new BufferKeyUntilOutputTimestamp()))
  // Subsequent processing goes here

private static class BreakRecordByModFn extends DoFn<DataChangeRecord,
                                                     DataChangeRecord>  {
  @ProcessElement
  public void processElement(
      @Element DataChangeRecord record, OutputReceiver<DataChangeRecord>
    outputReceiver) {
    record.getMods().stream()
      .map(
          mod ->
              new DataChangeRecord(
                  record.getPartitionToken(),
                  record.getCommitTimestamp(),
                  record.getServerTransactionId(),
                  record.isLastRecordInTransactionInPartition(),
                  record.getRecordSequence(),
                  record.getTableName(),
                  record.getRowType(),
                  Collections.singletonList(mod),
                  record.getModType(),
                  record.getValueCaptureType(),
                  record.getNumberOfRecordsInTransaction(),
                  record.getNumberOfPartitionsInTransaction(),
                  record.getTransactionTag(),
                  record.isSystemTransaction(),
                  record.getMetadata()))
      .forEach(outputReceiver::output);
  }
}

private static class KeyByIdFn extends DoFn<DataChangeRecord, KV<String, DataChangeRecord>>  {
  // NUMBER_OF_BUCKETS should be configured by the user to match their key cardinality
  // Here, we are choosing to hash the Spanner primary keys to a bucket index, in order to have a deterministic number
  // of states and timers for performance purposes.
  // Note that having too many buckets might have undesirable effects if it results in a low number of records per bucket
  // On the other hand, having too few buckets might also be problematic, since many keys will be contained within them.
  private static final int NUMBER_OF_BUCKETS = 1000;

  @ProcessElement
  public void processElement(
      @Element DataChangeRecord record,
      OutputReceiver<KV<String, DataChangeRecord>> outputReceiver) {
    int hashCode = (int) record.getMods().get(0).getKeysJson().hashCode();
    // Hash the received keys into a bucket in order to have a
    // deterministic number of buffers and timers.
    String bucketIndex = String.valueOf(hashCode % NUMBER_OF_BUCKETS);

    outputReceiver.output(KV.of(bucketIndex, record));
  }
}

private static class BufferKeyUntilOutputTimestamp extends
    DoFn<KV<String, DataChangeRecord>, KV<String, Iterable<DataChangeRecord>>>  {
  private static final Logger LOG =
      LoggerFactory.getLogger(BufferKeyUntilOutputTimestamp.class);

  private final long incrementIntervalInSeconds = 2;

  private BufferKeyUntilOutputTimestamp(long incrementIntervalInSeconds) {
    this.incrementIntervalInSeconds = incrementIntervalInSeconds;
  }

  @SuppressWarnings("unused")
  @TimerId("timer")
  private final TimerSpec timerSpec = TimerSpecs.timer(TimeDomain.EVENT_TIME);

  @StateId("buffer")
  private final StateSpec<BagState<DataChangeRecord>> buffer = StateSpecs.bag();

  @StateId("keyString")
  private final StateSpec<ValueState<String>> keyString =
      StateSpecs.value(StringUtf8Coder.of());

  @ProcessElement
  public void process(
      @Element KV<String, DataChangeRecord> element,
      @StateId("buffer") BagState<DataChangeRecord> buffer,
      @TimerId("timer") Timer timer,
      @StateId("keyString") ValueState<String> keyString) {
    buffer.add(element.getValue());

    // Only set the timer if this is the first time we are receiving a data change
    // record with this key.
    String elementKey = keyString.read();
    if (elementKey == null) {
      Instant commitTimestamp =
          new Instant(element.getValue().getCommitTimestamp().toSqlTimestamp());
      Instant outputTimestamp =
          commitTimestamp.plus(Duration.standardSeconds(incrementIntervalInSeconds));
      timer.set(outputTimestamp);
      keyString.write(element.getKey());
    }
  }

  @OnTimer("timer")
  public void onExpiry(
      OnTimerContext context,
      @StateId("buffer") BagState<DataChangeRecord> buffer,
      @TimerId("timer") Timer timer,
      @StateId("keyString") ValueState<String> keyString) {
    if (!buffer.isEmpty().read()) {
      String elementKey = keyString.read();

      final List<DataChangeRecord> records =
          StreamSupport.stream(buffer.read().spliterator(), false)
              .collect(Collectors.toList());
      buffer.clear();

      List<DataChangeRecord> recordsToOutput = new ArrayList<>();
      for (DataChangeRecord record : records) {
        Instant recordCommitTimestamp =
            new Instant(record.getCommitTimestamp().toSqlTimestamp());
        final String recordString =
            record.getMods().get(0).getNewValuesJson().isEmpty()
                ? "Deleted record"
                : record.getMods().get(0).getNewValuesJson();
        // When the watermark passes time T, this means that all records with
        // event time < T have been processed and successfully committed. Since the
        // timer fires when the watermark passes the expiration time, we should
        // only output records with event time < expiration time.
        if (recordCommitTimestamp.isBefore(context.timestamp())) {
          LOG.info(
             "Outputting record with key {} and value {} at expiration " +
             "timestamp {}",
              elementKey,
              recordString,
              context.timestamp().toString());
          recordsToOutput.add(record);
        } else {
          LOG.info(
              "Expired at {} but adding record with key {} and value {} back to " +
              "buffer due to commit timestamp {}",
              context.timestamp().toString(),
              elementKey,
              recordString,
              recordCommitTimestamp.toString());
          buffer.add(record);
        }
      }

      // Output records, if there are any to output.
      if (!recordsToOutput.isEmpty()) {
        // Order the records in place, and output them. The user would need
        // to implement DataChangeRecordComparator class that sorts the
        // data change records by commit timestamp and transaction ID.
        Collections.sort(recordsToOutput, new DataChangeRecordComparator());
        context.outputWithTimestamp(
            KV.of(elementKey, recordsToOutput), context.timestamp());
        LOG.info(
            "Expired at {}, outputting records for key {}",
            context.timestamp().toString(),
            elementKey);
      } else {
        LOG.info("Expired at {} with no records", context.timestamp().toString());
      }
    }

    Instant nextTimer = context.timestamp().plus(Duration.standardSeconds(incrementIntervalInSeconds));
    if (buffer.isEmpty() != null && !buffer.isEmpty().read()) {
      LOG.info("Setting next timer to {}", nextTimer.toString());
      timer.set(nextTimer);
    } else {
      LOG.info(
          "Timer not being set since the buffer is empty: ");
      keyString.clear();
    }
  }
}

pipeline
  .apply(SpannerIO
    .readChangeStream()
    .withSpannerConfig(SpannerConfig
      .create()
      .withProjectId("my-project-id")
      .withInstanceId("my-instance-id")
      .withDatabaseId("my-database-id")
      .withDatabaseRole("my-database-role"))    // Needed for fine-grained access control only
    .withChangeStreamName("my-change-stream")
    .withMetadataInstance("my-metadata-instance-id")
    .withMetadataDatabase("my-metadata-database-id")
    .withInclusiveStartAt(Timestamp.now()))
  .apply(ParDo.of(new KeyByTransactionIdFn()))
  .apply(ParDo.of(new TransactionBoundaryFn()))
  // Subsequent processing goes here

private static class KeyByTransactionIdFn extends DoFn<DataChangeRecord, KV<String, DataChangeRecord>>  {
  @ProcessElement
  public void processElement(
      @Element DataChangeRecord record,
      OutputReceiver<KV<String, DataChangeRecord>> outputReceiver) {
    outputReceiver.output(KV.of(record.getServerTransactionId(), record));
  }
}

private static class TransactionBoundaryFn extends DoFn<KV<String, DataChangeRecord>, KV<SortKey, Iterable<DataChangeRecord>>>  {
  @StateId("buffer")
  private final StateSpec<BagState<DataChangeRecord>> buffer = StateSpecs.bag();

  @StateId("count")
  private final StateSpec<ValueState<Integer>> countState = StateSpecs.value();

  @ProcessElement
  public void process(
      ProcessContext context,
      @StateId("buffer") BagState<DataChangeRecord> buffer,
      @StateId("count") ValueState<Integer> countState) {
    final KV<String, DataChangeRecord> element = context.element();
    final DataChangeRecord record = element.getValue();

    buffer.add(record);
    int count = (countState.read() != null ? countState.read() : 0);
    count = count + 1;
    countState.write(count);

    if (count == record.getNumberOfRecordsInTransaction()) {
      final List<DataChangeRecord> sortedRecords =
          StreamSupport.stream(buffer.read().spliterator(), false)
              .sorted(Comparator.comparing(DataChangeRecord::getRecordSequence))
              .collect(Collectors.toList());

      final Instant commitInstant =
          new Instant(sortedRecords.get(0).getCommitTimestamp().toSqlTimestamp()
              .getTime());
      context.outputWithTimestamp(
          KV.of(
              new SortKey(sortedRecords.get(0).getCommitTimestamp(),
                          sortedRecords.get(0).getServerTransactionId()),
              sortedRecords),
          commitInstant);
      buffer.clear();
      countState.clear();
    }
  }
}

pipeline
  .apply(SpannerIO
    .readChangeStream()
    .withSpannerConfig(SpannerConfig
      .create()
      .withProjectId("my-project-id")
      .withInstanceId("my-instance-id")
      .withDatabaseId("my-database-id")
      .withDatabaseRole("my-database-role"))    // Needed for fine-grained access control only
    .withChangeStreamName("my-change-stream")
    .withMetadataInstance("my-metadata-instance-id")
    .withMetadataDatabase("my-metadata-database-id")
    .withInclusiveStartAt(Timestamp.now()))
  .apply(Filter.by(record ->
           !record.isSystemTransaction()
           && record.getTransactionTag().equalsIgnoreCase("my-tx-tag")))
  // Subsequent processing goes here

pipeline
  .apply(SpannerIO
    .readChangeStream()
    .withSpannerConfig(SpannerConfig
      .create()
      .withProjectId("my-project-id")
      .withInstanceId("my-instance-id")
      .withDatabaseId("my-database-id")
      .withDatabaseRole("my-database-role"))    // Needed for fine-grained access control only
    .withChangeStreamName("my-change-stream")
    .withMetadataInstance("my-metadata-instance-id")
    .withMetadataDatabase("my-metadata-database-id")
    .withInclusiveStartAt(Timestamp.now()))
  .apply(Filter.by(record ->
           record.isSystemTransaction()
           && record.getTransactionTag().equals("RowDeletionPolicy")))
  // Subsequent processing goes here

CREATE TABLE Singers (
  SingerId INT64 NOT NULL,
  FirstName STRING(1024),
  LastName STRING(1024)
) PRIMARY KEY (SingerId);

SpannerConfig spannerConfig = SpannerConfig
   .create()
   .withProjectId("my-project-id")
   .withInstanceId("my-instance-id")
   .withDatabaseId("my-database-id")
   .withDatabaseRole("my-database-role");   // Needed for fine-grained access control only

pipeline
   .apply(SpannerIO
       .readChangeStream()
       .withSpannerConfig(spannerConfig)
       // Assume we have a change stream "my-change-stream" that watches Singers table.
       .withChangeStreamName("my-change-stream")
       .withMetadataInstance("my-metadata-instance-id")
       .withMetadataDatabase("my-metadata-database-id")
       .withInclusiveStartAt(Timestamp.now()))
   .apply(ParDo.of(new ToFullRowJsonFn(spannerConfig)))
   // Subsequent processing goes here

public class ToFullRowJsonFn extends DoFn<DataChangeRecord, String> {
 // Since each instance of this DoFn will create its own session pool and will
 // perform calls to Spanner sequentially, we keep the number of sessions in
 // the pool small. This way, we avoid wasting resources.
 private static final int MIN_SESSIONS = 1;
 private static final int MAX_SESSIONS = 5;
 private final String projectId;
 private final String instanceId;
 private final String databaseId;

 private transient DatabaseClient client;
 private transient Spanner spanner;

 public ToFullRowJsonFn(SpannerConfig spannerConfig) {
   this.projectId = spannerConfig.getProjectId().get();
   this.instanceId = spannerConfig.getInstanceId().get();
   this.databaseId = spannerConfig.getDatabaseId().get();
 }

 @Setup
 public void setup() {
   SessionPoolOptions sessionPoolOptions = SessionPoolOptions
      .newBuilder()
      .setMinSessions(MIN_SESSIONS)
      .setMaxSessions(MAX_SESSIONS)
      .build();
   SpannerOptions options = SpannerOptions
       .newBuilder()
       .setProjectId(projectId)
       .setSessionPoolOption(sessionPoolOptions)
       .build();
   DatabaseId id = DatabaseId.of(projectId, instanceId, databaseId);
   spanner = options.getService();
   client = spanner.getDatabaseClient(id);
 }

 @Teardown
 public void teardown() {
   spanner.close();
 }

 @ProcessElement
 public void process(
   @Element DataChangeRecord element,
   OutputReceiver<String> output) {
   com.google.cloud.Timestamp commitTimestamp = element.getCommitTimestamp();
   element.getMods().forEach(mod -> {
     JSONObject keysJson = new JSONObject(mod.getKeysJson());
     JSONObject newValuesJson = new JSONObject(mod.getNewValuesJson());
     ModType modType = element.getModType();
     JSONObject jsonRow = new JSONObject();
     long singerId = keysJson.getLong("SingerId");
     jsonRow.put("SingerId", singerId);
     if (modType == ModType.INSERT) {
       // For INSERT mod, get non-primary key columns from mod.
       jsonRow.put("FirstName", newValuesJson.get("FirstName"));
       jsonRow.put("LastName", newValuesJson.get("LastName"));
     } else if (modType == ModType.UPDATE) {
       // For UPDATE mod, get non-primary key columns by doing a snapshot read using the primary key column from mod.
       try (ResultSet resultSet = client
         .singleUse(TimestampBound.ofReadTimestamp(commitTimestamp))
         .read(
           "Singers",
           KeySet.singleKey(com.google.cloud.spanner.Key.of(singerId)),
             Arrays.asList("FirstName", "LastName"))) {
         if (resultSet.next()) {
           jsonRow.put("FirstName", resultSet.isNull("FirstName") ?
             JSONObject.NULL : resultSet.getString("FirstName"));
           jsonRow.put("LastName", resultSet.isNull("LastName") ?
             JSONObject.NULL : resultSet.getString("LastName"));
         }
       }
     } else {
       // For DELETE mod, there is nothing to do, as we already set SingerId.
     }

     output.output(jsonRow.toString());
   });
 }
}

pipeline
  .apply(SpannerIO
    .readChangeStream()
    .withSpannerConfig(SpannerConfig
      .create()
      .withProjectId("my-project-id")
      .withInstanceId("my-instance-id")
      .withDatabaseId("my-database-id")
      .withDatabaseRole("my-database-role"))    // Needed for fine-grained access control only
    .withChangeStreamName("my-change-stream")
    .withMetadataInstance("my-metadata-instance-id")
    .withMetadataDatabase("my-metadata-database-id")
    .withInclusiveStartAt(Timestamp.now()))
  .apply(MapElements.into(TypeDescriptors.strings()).via(Object::toString))
  .apply(PubsubIO.writeStrings().to("my-topic"));

pipeline
  .apply(SpannerIO
    .readChangeStream()
    .withSpannerConfig(SpannerConfig
      .create()
      .withProjectId("my-project-id")
      .withInstanceId("my-instance-id")
      .withDatabaseId("my-database-id")
      .withDatabaseRole("my-database-role"))    // Needed for fine-grained access control only
    .withChangeStreamName("my-change-stream")
    .withMetadataInstance("my-metadata-instance-id")
    .withMetadataDatabase("my-metadata-database-id")
    .withInclusiveStartAt(Timestamp.now()))
  .apply(MapElements.into(TypeDescriptors.strings()).via(Object::toString))
  .apply(Window.into(FixedWindows.of(Duration.standardMinutes(1))))
  .apply(TextIO
    .write()
    .to("gs://my-bucket/change-stream-results-")
    .withSuffix(".txt")
    .withWindowedWrites()
    .withNumShards(1));

SpannerConfig spannerConfig = SpannerConfig
  .create()
  .withProjectId("my-project-id")
  .withInstanceId("my-instance-id")
  .withDatabaseId("my-database-id")
  .withDatabaseRole("my-database-role");   // Needed for fine-grained access control only

pipeline
  .apply(SpannerIO
    .readChangeStream()
    .withSpannerConfig(spannerConfig)
    .withChangeStreamName("my-change-stream")
    .withMetadataInstance("my-metadata-instance-id")
    .withMetadataDatabase("my-metadata-database-id")
    .withInclusiveStartAt(Timestamp.now()))
  .apply(ParDo.of(new ToFullRowJsonFn(spannerConfig)))
  .apply(BigQueryIO
    .<String>write()
    .to("my-bigquery-table")
    .withCreateDisposition(CreateDisposition.CREATE_IF_NEEDED)
    .withWriteDisposition(Write.WriteDisposition.WRITE_APPEND)
    .withSchema(new TableSchema().setFields(Arrays.asList(
      new TableFieldSchema()
        .setName("SingerId")
        .setType("INT64")
        .setMode("REQUIRED"),
      new TableFieldSchema()
        .setName("FirstName")
        .setType("STRING")
        .setMode("REQUIRED"),
      new TableFieldSchema()
        .setName("LastName")
        .setType("STRING")
        .setMode("REQUIRED")
    )))
    .withAutoSharding()
    .optimizedWrites()
    .withFormatFunction((String element) -> {
      ObjectMapper objectMapper = new ObjectMapper();
      JsonNode jsonNode = null;
      try {
        jsonNode = objectMapper.readTree(element);
      } catch (IOException e) {
        e.printStackTrace();
      }
      return new TableRow()
        .set("SingerId", jsonNode.get("SingerId").asInt())
        .set("FirstName", jsonNode.get("FirstName").asText())
        .set("LastName", jsonNode.get("LastName").asText());
    }
  )
);

java.lang.NullPointerException: null value in entry: Cloud Storage_PROJECT_ID=null

<dependency>
  <groupId>org.apache.beam</groupId>
  <artifactId>beam-sdks-java-core</artifactId>
  <version>2.37.0</version>
</dependency>