[SPARK-42102] Using checkpoints in Spark Structured Streaming with the foreachBatch sink

• Type: Question
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 3.3.1
• Component/s: PySpark, Structured Streaming
• Assignee: Unassigned
• Reporter: Kai-Michael Roesner
• Created: 17/Jan/23 14:48
• Updated: 17/Jan/23 14:48

Description

I want to build a fault-tolerant, recoverable Spark job (using Structured Streaming in PySpark) that reads a data stream from Kafka and uses the foreachBatch sink to implement a stateful transformation before writing the resulting data to the actual sink.

The basic structure of my Spark job is like this:

counter = 0

def batch_handler(df, batch_id):
  global counter
  counter += 1
  df.withColumn('counter', lit(counter)).show(truncate=30)

spark = (SparkSession.builder
  .appName('test.stateful.checkpoint')
  .config('spark.jars.packages', f'{KAFKA_SQL},{KAFKA_CLNT}')
  .getOrCreate())

source = (spark.readStream
  .format('kafka')
  .options(**KAFKA_OPTIONS)
  .option('subscribe', 'topic-spark-stateful')
  .option('startingOffsets', 'earliest')
  .option('includeHeaders', 'true')
  .load())

(source
  .selectExpr('CAST(value AS STRING) AS data', 'CAST(timestamp AS STRING) AS time')
  .writeStream
  .option('checkpointLocation', './checkpoints/stateful')
  .foreachBatch(batch_handler)
  .start()
  .awaitTermination())

where the simplified batch_handler function is a stand-in for the stateful transformation + writer to the actual data sink. Also for simplicity I am using a local folder as checkpoint location.

This works fine as far as checkpointing of Kafka offsets is concerned. But how can I include the state of my custom batch handler (counter in my simplified example) in the checkpoints such that the job can pick up where it left after a crash?

The Spark Structured Streaming Guide doesn't say anything on the topic. With the foreach sink I can pass a custom row handler object but this seems to support only open, process, and close methods.

Would it make sense to create a "Request" or even "Feature" ticket to enhance this with methods for restoring state from a checkpoint and exporting state to support checkpointing?

PS: I have posted this on SOF, too. If anyone cares to answer or comment I'd be happy to upvote their post.