实时模式示例

重要

此功能目前以公共预览版提供。

本页提供结构化流式处理中实时模式查询的工作代码示例，从简单的无状态转换到使用自定义状态管理进行复杂的有状态处理。有关概念和配置，请参阅结构化流式处理中的实时模式，动手教程请查看实时模式入门。

先决条件

若要在此页上运行示例，需要：

配置并运行实时模式群集。有关分步设置说明，请参阅实时模式入门。
Databricks Runtime 16.4 LTS 或更高版本。
访问支持的流式处理源和接收器：
- 对于 Kafka 示例：配置了输入/输出主题的 Kafka 代理
- 对于 Kinesis 示例：AWS 凭据和为增强型 Fan-Out (EFO) 模式配置的 Kinesis 数据流
- 对于自定义接收器示例：已配置的目标数据库或服务（提供的示例中的 PostgreSQL）
基本熟悉结构化流概念。如果你是流式处理新手，请参阅结构化流式处理概念。

注释

这些示例使用占位符值，例如 broker_address， input_topic和 checkpoint_location。在运行代码之前，请将这些值替换为实际配置值。

无状态查询示例

无状态查询独立处理每个记录，而无需维护记录之间的任何状态。这些查询通常更简单，延迟比有状态查询低，因为它们不需要管理状态存储或执行查找。对转换、筛选、与静态数据和路由作联接使用无状态查询。

Kafka 源到 Kafka 接收器

在此示例中，从 Kafka 源读取并写入 Kafka 接收端。

Python

query = (
    spark.readStream
        .format("kafka")
        .option("kafka.bootstrap.servers", broker_address)
        .option("startingOffsets", "earliest")
        .option("subscribe", input_topic)
        .load()
        .writeStream
        .format("kafka")
        .option("kafka.bootstrap.servers", broker_address)
        .option("topic", output_topic)
        .option("checkpointLocation", checkpoint_location)
        .trigger(realTime="5 minutes")
        .outputMode("update")
        .start()
)

Scala

import org.apache.spark.sql.streaming.OutputMode
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic)
      .load()
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply())
      .outputMode(OutputMode.Update())
      .start()

重新分区

在此示例中，从 Kafka 源读取数据，将数据重新分区到 20 个分区，并写入 Kafka 接收器。

由于当前实现限制，必须在使用重新分区之前将 Spark 配置从spark.sql.execution.sortBeforeRepartition设置为false。

Python

# Sorting is not supported in repartition with real-time mode, so you must set this to false to achieve low latency.
spark.conf.set("spark.sql.execution.sortBeforeRepartition", "false")

query = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("subscribe", input_topic)
    .option("startingOffsets", "earliest")
    .load()
    .repartition(20)
    .writeStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("topic", output_topic)
    .option("checkpointLocation", checkpoint_location)
    .trigger(realTime="5 minutes")
    .outputMode("update")
    .start()
)

Scala

import org.apache.spark.sql.streaming.OutputMode
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

// Sorting is not supported in repartition with real-time mode, so you must set this to false to achieve low latency.
spark.conf.set("spark.sql.execution.sortBeforeRepartition", "false")

spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic)
      .load()
      .repartition(20)
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply())
      .outputMode(OutputMode.Update())
      .start()

流 - 快照联接（仅广播）

在此示例中，从 Kafka 读取数据，将数据与静态表联接，并写入 Kafka 接收器。仅支持将静态表广播的流-静态联接，这意味着静态表应能放入内存中。

Python

from pyspark.sql.functions import broadcast, expr

# We assume the static table in the path `static_table_location` has a column 'lookupKey'.

query = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("subscribe", input_topic)
    .option("startingOffsets", "earliest")
    .load()
    .withColumn("joinKey", expr("CAST(value AS STRING)"))
    .join(
        broadcast(spark.read.format("parquet").load(static_table_location)),
        expr("joinKey = lookupKey")
    )
    .selectExpr("value AS key", "value")
    .writeStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("topic", output_topic)
    .option("checkpointLocation", checkpoint_location)
    .trigger(realTime="5 minutes")
    .outputMode("update")
    .start()
)

Scala

import org.apache.spark.sql.functions.{broadcast, expr}
import org.apache.spark.sql.streaming.OutputMode
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic)
      .load()
      .join(broadcast(spark.read.format("parquet").load(staticTableLocation)), expr("joinKey = lookupKey"))
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply())
      .outputMode(OutputMode.Update())
      .start()

Kinesis 源到 Kafka 接收器

在此示例中，从 Kinesis 数据源读取并写入 Kafka 数据接收端。

Python

query = (
    spark.readStream
        .format("kinesis")
        .option("region", region_name)
        .option("awsAccessKey", aws_access_key_id)
        .option("awsSecretKey", aws_secret_access_key)
        .option("consumerMode", "efo")
        .option("consumerName", consumer_name)
        .load()
        .selectExpr("partitionKey AS key", "CAST(data AS STRING) AS value")
        .writeStream
        .format("kafka")
        .option("kafka.bootstrap.servers", broker_address)
        .option("topic", output_topic)
        .option("checkpointLocation", checkpoint_location)
        .trigger(realTime="5 minutes")
        .outputMode("update")
        .start()
)

Scala

import org.apache.spark.sql.streaming.OutputMode
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

spark.readStream
      .format("kinesis")
      .option("region", regionName)
      .option("awsAccessKey", awsAccessKeyId)
      .option("awsSecretKey", awsSecretAccessKey)
      .option("consumerMode", "efo")
      .option("consumerName", consumerName)
      .load()
      .select(
        col("partitionKey").alias("key"),
        col("data").cast("string").alias("value")
      )
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply())
      .outputMode(OutputMode.Update())
      .start()

Union

在此示例中，你合并来自两个不同主题的两个 Kafka 数据帧并写入 Kafka 接收器。

Python

df1 = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("startingOffsets", "earliest")
    .option("subscribe", input_topic_1)
    .load()
)

df2 = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("startingOffsets", "earliest")
    .option("subscribe", input_topic_2)
    .load()
)

query = (
    df1.union(df2)
    .writeStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("topic", output_topic)
    .option("checkpointLocation", checkpoint_location)
    .trigger(realTime="5 minutes")
    .outputMode("update")
    .start()
)

Scala

import org.apache.spark.sql.streaming.OutputMode
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

val df1 = spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic1)
      .load()

val df2 = spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic2)
      .load()

df1.union(df2)
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply())
      .outputMode(OutputMode.Update())
      .start()

有状态查询示例

有状态查询在跨记录的过程中维护状态信息，从而实现去重、聚合和窗口化等操作。这些查询对于需要一段时间内或跨多个事件跟踪信息的用例至关重要。实时模式支持具有与微批处理模式相同的语义的有状态作，但会持续处理数据以降低延迟。有状态查询需要比无状态查询更多的内存和计算资源，因为它们必须维护和更新状态。

去重

在此示例中，您将基于 timestamp 列和 value 列去重记录。

Python

query = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("startingOffsets", "earliest")
    .option("subscribe", input_topic)
    .load()
    .dropDuplicates(["timestamp", "value"])
    .writeStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("topic", output_topic)
    .option("checkpointLocation", checkpoint_location)
    .trigger(realTime="5 minutes")
    .outputMode("update")
    .start()
)

Scala

import org.apache.spark.sql.streaming.OutputMode
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic)
      .load()
      .dropDuplicates("timestamp", "value")
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply())
      .outputMode(OutputMode.Update())
      .start()

集合体

在此示例中，按timestamp和value将记录分组，然后计数其出现次数。

Python

from pyspark.sql.functions import col

query = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("startingOffsets", "earliest")
    .option("subscribe", input_topic)
    .load()
    .groupBy(col("timestamp"), col("value"))
    .count()
    .selectExpr("CAST(value AS STRING) AS key", "CAST(count AS STRING) AS value")
    .writeStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("topic", output_topic)
    .option("checkpointLocation", checkpoint_location)
    .trigger(realTime="5 minutes")
    .outputMode("update")
    .start()
)

Scala

import org.apache.spark.sql.functions.col
import org.apache.spark.sql.streaming.OutputMode
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic)
      .load()
      .groupBy(col("timestamp"), col("value"))
      .count()
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply("5 minutes"))
      .outputMode(OutputMode.Update())
      .start()

与聚合的联合

在此示例中，首先将两个不同主题中的两个 Kafka DataFrame 合并，然后执行聚合。最后，将数据写入 Kafka sink。

Python

from pyspark.sql.functions import col

df1 = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("startingOffsets", "earliest")
    .option("subscribe", input_topic_1)
    .load()
)

df2 = (
    spark.readStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("startingOffsets", "earliest")
    .option("subscribe", input_topic_2)
    .load()
)

query = (
    df1.union(df2)
    .groupBy(col("timestamp"), col("value"))
    .count()
    .selectExpr("CAST(value AS STRING) AS key", "CAST(count AS STRING) AS value")
    .writeStream
    .format("kafka")
    .option("kafka.bootstrap.servers", broker_address)
    .option("topic", output_topic)
    .option("checkpointLocation", checkpoint_location)
    .trigger(realTime="5 minutes")
    .outputMode("update")
    .start()
)

Scala

import org.apache.spark.sql.functions.col
import org.apache.spark.sql.execution.streaming.RealTimeTrigger

val df1 = spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic1)
      .load()

val df2 = spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic2)
      .load()

df1.union(df2)
      .groupBy(col("timestamp"), col("value"))
      .count()
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply())
      .outputMode(OutputMode.Update())
      .start()

transformWithState

在此示例中，你将使用 transformWithState TTL（生存时间）来维护自定义状态。处理器对每个键看到的记录数进行计数。

Python

from typing import Iterator, Tuple

from pyspark.sql import Row
from pyspark.sql.streaming import StatefulProcessor, StatefulProcessorHandle
from pyspark.sql.types import LongType, StringType, TimestampType, StructField, StructType

class RTMStatefulProcessor(StatefulProcessor):
  """
  This processor counts the number of records it has seen for each key using state variables
  with TTLs. It redundantly maintains this count with a value, list, and map state to put load
  on the state variable cleanup mechanism. (In practice, only one value state is needed to maintain
  the count for a given grouping key.)

  The input schema it expects is (String, Long) which represents a (key, source-timestamp) tuple.
  The source-timestamp is passed through so that we can calculate end-to-end latency. The output
  schema is (String, Long, Long), which represents a (key, count, source-timestamp) 3-tuple.
  """

  def init(self, handle: StatefulProcessorHandle) -> None:
    state_schema = StructType([StructField("value", LongType(), True)])
    self.value_state = handle.getValueState("value", state_schema, 30000)
    map_key_schema = StructType([StructField("key", LongType(), True)])
    map_value_schema = StructType([StructField("value", StringType(), True)])
    self.map_state = handle.getMapState("map", map_key_schema, map_value_schema, 30000)
    list_schema = StructType([StructField("value", StringType(), True)])
    self.list_state = handle.getListState("list", list_schema, 30000)

  def handleInputRows(self, key, rows, timerValues) -> Iterator[Row]:
    for row in rows:
      # row is a tuple (key, source_timestamp)
      key_str = row[0]
      source_timestamp = row[1]
      old_value = value.get()
      if old_value is None:
        old_value = 0
      self.value_state.update((old_value + 1,))
      self.map_state.update((old_value,), (key_str,))
      self.list_state.appendValue((key_str,))
      yield Row(key=key_str, value=old_value + 1, timestamp=source_timestamp)

  def close(self) -> None:
    pass

output_schema = StructType(
  [
    StructField("key", StringType(), True),
    StructField("value", LongType(), True),
    StructField("timestamp", TimestampType(), True),
  ]
)

query = (
  spark.readStream
  .format("kafka")
  .option("kafka.bootstrap.servers", broker_address)
  .option("subscribe", input_topic)
  .load()
  .selectExpr("CAST(key AS STRING)", "CAST(value AS STRING)", "timestamp")
  .groupBy("key")
  .transformWithState(
    statefulProcessor=RTMStatefulProcessor(),
    outputStructType=output_schema,
    outputMode="Update",
    timeMode="processingTime",
  )
  .writeStream
  .format("kafka")
  .option("kafka.bootstrap.servers", broker_address)
  .option("topic", output_topic)
  .option("checkpointLocation", checkpoint_location)
  .trigger(realTime="5 minutes")
  .outputMode("Update")
  .start()
)

Scala

import org.apache.spark.sql.Encoders
import org.apache.spark.sql.execution.streaming.RealTimeTrigger
import org.apache.spark.sql.streaming.{ListState, MapState, StatefulProcessor, OutputMode, TTLConfig, TimeMode, TimerValues, ValueState}

/**
 * This processor counts the number of records it has seen for each key using state variables
 * with TTLs. It redundantly maintains this count with a value, list, and map state to put load
 * on the state variable cleanup mechanism. (In practice, only one value state is needed to maintain
 * the count for a given grouping key.)
 *
 * The input schema it expects is (String, Long) which represents a (key, source-timestamp) tuple.
 * The source-timestamp is passed through so that we can calculate end-to-end latency. The output
 * schema is (String, Long, Long), which represents a (key, count, source-timestamp) 3-tuple.
 *
 */

class RTMStatefulProcessor(ttlConfig: TTLConfig)
  extends StatefulProcessor[String, (String, Long), (String, Long, Long)] {
  @transient private var _value: ValueState[Long] = _
  @transient private var _map: MapState[Long, String] = _
  @transient private var _list: ListState[String] = _

  override def init(outputMode: OutputMode, timeMode: TimeMode): Unit = {
    // Counts the number of records this key has seen
    _value = getHandle.getValueState("value", Encoders.scalaLong, ttlConfig)
    _map = getHandle.getMapState("map", Encoders.scalaLong, Encoders.STRING, ttlConfig)
    _list = getHandle.getListState("list", Encoders.STRING, ttlConfig)
  }

  override def handleInputRows(
      key: String,
      inputRows: Iterator[(String, Long)],
      timerValues: TimerValues): Iterator[(String, Long, Long)] = {
    inputRows.map { row =>
      val key = row._1
      val sourceTimestamp = row._2

      val oldValue = _value.get()
      _value.update(oldValue + 1)
      _map.updateValue(oldValue, key)
      _list.appendValue(key)

      (key, oldValue + 1, sourceTimestamp)
    }
  }
}

spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("subscribe", inputTopic)
      .load()
      .select(col("key").cast("STRING"), col("value").cast("STRING"), col("timestamp"))
      .as[(String, String, Timestamp)]
      .groupByKey(row => row._1)
      .transformWithState(new RTMStatefulProcessor(TTLConfig(Duration.ofSeconds(30))), TimeMode.ProcessingTime, OutputMode.Update)
      .as[(String, Long, Long)]
      .select(
            col("_1").as("key"),
            col("_2").as("value")
      )
      .writeStream
      .format("kafka")
      .option("kafka.bootstrap.servers", brokerAddress)
      .option("topic", outputTopic)
      .option("checkpointLocation", checkpointLocation)
      .trigger(RealTimeTrigger.apply("5 minutes"))
      .outputMode(OutputMode.Update())
      .start()

注释

结构化流式处理中，实时模式和其他执行模式的运行方式 StatefulProcessor 在 transformWithState 之间存在差异。请参阅在实时模式下使用 transformWithState。

开发和测试

使用显示进行交互式开发

可以使用函数 display 直接在笔记本中可视化实时流数据。这对于交互式开发、测试和调试实时模式查询非常有用，无需设置外部接收器或生产基础结构。

display具有realTime触发器的函数在 Databricks Runtime 17.1 及更高版本中可用。在开发期间，使用 display 来验证您的查询逻辑和数据转换，然后再通过 Kafka 或自定义接收器部署到生产环境。有关使用速率源 display的完整示例，请参阅实时模式入门。

显示速率源

在此示例中，您从速率源读取数据，并在笔记本中显示流式数据帧。

Python

inputDF = (
  spark
  .readStream
  .format("rate")
  .option("numPartitions", 2)
  .option("rowsPerSecond", 1)
  .load()
)
display(inputDF, realTime="5 minutes", outputMode="update")

Scala

import org.apache.spark.sql.streaming.Trigger
import org.apache.spark.sql.streaming.OutputMode

val inputDF = spark
  .readStream
  .format("rate")
  .option("numPartitions", 2)
  .option("rowsPerSecond", 1)
  .load()
display(inputDF, trigger=Trigger.RealTime(), outputMode=OutputMode.Update())

自定义接收器示例

当需要将流式处理数据写入没有内置结构化流式处理支持的目标时，请使用 foreachSink 来实现自定义写入逻辑。使用自定义数据接收器可以让您完全掌控数据的写入方式，从而能够与任何数据库、API或存储系统集成。以下示例演示如何使用 JDBC 写入 PostgreSQL 数据库。

使用 foreachSink 写入 PostgreSQL

import java.sql.{Connection, DriverManager, PreparedStatement}

import org.apache.spark.sql.{ForeachWriter, Row}

/**
 * Groups connection properties for
 * the JDBC writers.
 *
 * @param url JDBC url of the form jdbc:subprotocol:subname to connect to
 * @param dbtable database table that should be written into
 * @param username username for authentication
 * @param password password for authentication
 */
class JdbcWriterConfig(
    val url: String,
    val dbtable: String,
    val username: String,
    val password: String,
) extends Serializable

/**
 * Handles streaming data writes to a database sink via JDBC, by:
 * - connecting to the database
 * - buffering incoming data rows in batches to reduce write overhead
 *
 * @param config connection parameters and configuration knobs for the writer
 */
class JdbcStreamingDataWriter(config: JdbcWriterConfig)
  extends ForeachWriter[Row] with Serializable {
  // The writer currently only supports this hard-coded schema
  private val UPSERT_STATEMENT_SQL =
    s"""MERGE INTO "${config.dbtable}"
       |USING (
       |  SELECT
       |    CAST(? AS INTEGER) AS "id",
       |    CAST(? AS CHARACTER VARYING) AS "data"
       |) AS "source"
       |ON "test"."id" = "source"."id"
       |WHEN MATCHED THEN
       |  UPDATE SET "data" = "source"."data"
       |WHEN NOT MATCHED THEN
       |  INSERT ("id", "data") VALUES ("source"."id", "source"."data")
       |""".stripMargin

  private val MAX_BUFFER_SIZE = 3
  private val buffer = new Array[Row](MAX_BUFFER_SIZE)
  private var bufferSize = 0

  private var connection: Connection = _

  /**
   * Flushes the [[buffer]] by writing all rows in the buffer to the database.
   */
  private def flushBuffer(): Unit = {
    require(connection != null)

    if (bufferSize == 0) {
      return
    }

    var upsertStatement: PreparedStatement = null

    try {
      upsertStatement = connection.prepareStatement(UPSERT_STATEMENT_SQL)

      for (i <- 0 until bufferSize) {
        val row = buffer(i)
        upsertStatement.setInt(1, row.getAs[String]("key"))
        upsertStatement.setString(2, row.getAs[String]("value"))
        upsertStatement.addBatch()
      }

      upsertStatement.executeBatch()
      connection.commit()

      bufferSize = 0
    } catch { case e: Exception =>
      if (connection != null) {
        connection.rollback()
      }
      throw e
    } finally {
      if (upsertStatement != null) {
        upsertStatement.close()
      }
    }
  }

  override def open(partitionId: Long, epochId: Long): Boolean = {
    connection = DriverManager.getConnection(config.url, config.username, config.password)
    true
  }

  override def process(row: Row): Unit = {
    buffer(bufferSize) = row
    bufferSize += 1
    if (bufferSize >= MAX_BUFFER_SIZE) {
      flushBuffer()
    }
  }

  override def close(errorOrNull: Throwable): Unit = {
    flushBuffer()
    if (connection != null) {
      connection.close()
      connection = null
    }
  }
}

spark.readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", testUtils.brokerAddress)
      .option("subscribe", inputTopic)
      .load()
      .writeStream
      .outputMode(OutputMode.Update())
      .trigger(defaultTrigger)
      .foreach(new JdbcStreamingDataWriter(new JdbcWriterConfig(jdbcUrl, tableName, jdbcUsername, jdbcPassword)))
      .start()

后续步骤

了解这些实时模式示例后，以下是用于深化知识和构建可用于生产的流式处理应用程序的资源：

实时模式参考 - 了解实时模式概念、配置选项、支持的功能和限制
实时模式入门 - 按照分步说明配置群集并运行第一个实时查询
结构化流概念 - 了解 Databricks 上的结构化流式处理的基础概念
有状态应用程序 - 深入了解 transformWithState 和自定义状态管理模式

Last updated on 2026-03-11

通过

实时模式示例

先决条件

无状态查询示例

Kafka 源到 Kafka 接收器

Python

Scala

重新分区

Python

Scala

流 - 快照联接（仅广播）

Python

Scala

Kinesis 源到 Kafka 接收器

Python

Scala

Union

Python

Scala

有状态查询示例

去重

Python

Scala

集合体

Python

Scala

与聚合的联合

Python

Scala

transformWithState

Python

Scala

开发和测试

使用显示进行交互式开发

显示速率源

Python

Scala

自定义接收器示例

使用 foreachSink 写入 PostgreSQL

后续步骤

其他资源