8

我已经设置了 Spark Structured Streaming (Spark 2.3.2) 以从 Kafka (2.0.0) 读取。如果消息在 Spark 流作业启动之前进入主题，我将无法从主题的开头消费。这是 Spark 流的预期行为，它忽略了在 Spark Stream 作业初始运行之前产生的 Kafka 消息（即使使用 .option("stratingOffsets","earliest")）？

重现步骤

  1. 在开始流式作业之前，创建test主题（单个代理，单个分区）并为主题生成消息（在我的示例中为 3 条消息）。

  2. 使用以下命令启动 spark-shell：spark-shell --packages org.apache.spark:spark-sql-kafka-0-10_2.11:2.3.2.3.1.0.0-78 --repositories http://repo.hortonworks.com/content/repositories/releases/

  3. 执行下面的 spark scala 代码。

// Local
val df = spark.readStream.format("kafka")
  .option("kafka.bootstrap.servers", "localhost:9097")
  .option("failOnDataLoss","false")
  .option("stratingOffsets","earliest")
  .option("subscribe", "test")
  .load()

// Sink Console
val ds = df.writeStream.format("console").queryName("Write to console")
  .trigger(org.apache.spark.sql.streaming.Trigger.ProcessingTime("10 second"))
  .start()

预期与实际产出

我希望流从 offset=1 开始。但是，它从 offset=3 开始读取。可以看到，kafka 客户端实际上是在重置起始偏移量：2019-06-18 21:22:57 INFO Fetcher:583 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] Resetting offset for partition test-0 to offset 3.

我可以看到火花流处理我在开始流式传输作业后生成的消息。

这是 Spark 流的预期行为，它忽略了在 Spark Stream 作业初始运行之前产生的 Kafka 消息（即使使用.option("stratingOffsets","earliest")）？

2019-06-18 21:22:57 INFO  AppInfoParser:109 - Kafka version : 2.0.0.3.1.0.0-78
2019-06-18 21:22:57 INFO  AppInfoParser:110 - Kafka commitId : 0f47b27cde30d177
2019-06-18 21:22:57 INFO  MicroBatchExecution:54 - Starting new streaming query.
2019-06-18 21:22:57 INFO  Metadata:273 - Cluster ID: LqofSZfjTu29BhZm6hsgsg
2019-06-18 21:22:57 INFO  AbstractCoordinator:677 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] Discovered group coordinator localhost:9097 (id: 2147483647 rack: null)
2019-06-18 21:22:57 INFO  ConsumerCoordinator:462 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] Revoking previously assigned partitions []
2019-06-18 21:22:57 INFO  AbstractCoordinator:509 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] (Re-)joining group
2019-06-18 21:22:57 INFO  AbstractCoordinator:473 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] Successfully joined group with generation 1
2019-06-18 21:22:57 INFO  ConsumerCoordinator:280 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] Setting newly assigned partitions [test-0]
2019-06-18 21:22:57 INFO  Fetcher:583 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] Resetting offset for partition test-0 to offset 3.
2019-06-18 21:22:58 INFO  KafkaSource:54 - Initial offsets: {"test":{"0":3}}
2019-06-18 21:22:58 INFO  Fetcher:583 - [Consumer clientId=consumer-2, groupId=spark-kafka-source-e948eee9-3024-4f14-bcb8-75b80d43cbb1--181544888-driver-0] Resetting offset for partition test-0 to offset 3.
2019-06-18 21:22:58 INFO  MicroBatchExecution:54 - Committed offsets for batch 0. Metadata OffsetSeqMetadata(0,1560910978083,Map(spark.sql.shuffle.partitions -> 200, spark.sql.streaming.stateStore.providerClass -> org.apache.spark.sql.execution.streaming.state.HDFSBackedStateStoreProvider))
2019-06-18 21:22:58 INFO  KafkaSource:54 - GetBatch called with start = None, end = {"test":{"0":3}}

火花批处理模式

我能够确认批处理模式从一开始就读取 - 所以 Kafka 保留配置没有问题

val df = spark
  .read
  .format("kafka")
  .option("kafka.bootstrap.servers", "localhost:9097")
  .option("subscribe", "test")
  .load()

df.count // Long = 3
4

2 回答 2

2

哈哈，这是一个简单的错字：“stratingOffsets”应该是“startingOffsets”

于 2019-06-27T16:57:32.590 回答
1

您可以通过两种方式做到这一点。将数据从 kafka 加载到流数据帧或将数据从 kafka 加载到静态数据帧（用于测试）。

我认为您因为 group-id 而看不到数据。kafka 会将消费者组和偏移量提交到内部主题。确保每次读取的组名都是唯一的。

这是两个选项。

选项 1：从 kafka 读取数据到流式数据帧 

// spark streaming with kafka 

import org.apache.spark.sql.streaming.ProcessingTime

val ds1 = spark.readStream.format("kafka")
.option("kafka.bootstrap.servers","app01.app.test.net:9097,app02.app.test.net:9097")
.option("subscribe", "kafka-testing-topic")
.option("kafka.security.protocol", "SASL_PLAINTEXT")
.option("startingOffsets","earliest")
.option("maxOffsetsPerTrigger","6000")
.load()

val ds2 = ds1.select(from_json($"value".cast(StringType), dataSchema).as("data")).select("data.*")
val ds3 = ds2.groupBy("TABLE_NAME").count()
ds3.writeStream
.trigger(ProcessingTime("10 seconds"))
.queryName("query1").format("console")
.outputMode("complete")
.start()
.awaitTermination()

选项2：从kafka读取数据到静态数据帧（为了测试，它将从头开始加载）


// Subscribe to 1 topic defaults to the earliest and latest offsets
val ds1 = spark.read.format("kafka")
.option("kafka.bootstrap.servers","app01.app.test.net:9097,app02.app.test.net:9097")
.option("subscribe", "kafka-testing-topic")
.option("kafka.security.protocol", "SASL_PLAINTEXT")
.option("spark.streaming.kafka.consumer.cache.enabled","false")
.load()

val ds2 = ds1.selectExpr("CAST(key AS STRING)", "CAST(value AS STRING)","topic","partition","offset","timestamp")
val ds3 = ds2.select("value").rdd.map(x => x.toString)
ds3.count()
于 2019-06-19T05:16:01.043 回答