13

I'm building a Kafka ingest module in EMR 5.11.1, Spark 2.2.1. My intention is to use Structured Streaming to consume from a Kafka topic, do some processing, and store to EMRFS/S3 in parquet format.

Console sink works as expected, file sink does not work.

In spark-shell:

val event = spark.readStream.format("kafka")
.option("kafka.bootstrap.servers", <server list>)
.option("subscribe", <topic>)
.load()

val eventdf = event.select($"value" cast "string" as "json")
.select(from_json($"json", readSchema) as "data")
.select("data.*")

val outputdf = <some processing on eventdf>

This works:

val console_query = outputdf.writeStream.format("console")
.trigger(Trigger.ProcessingTime(10.seconds))
.outputMode(OutputMode.Append)
.start

This doesn't:

val filesink_query = outputdf.writeStream
.partitionBy(<some column>)
.format("parquet")
.option("path", <some path in EMRFS>)
.option("checkpointLocation", "/tmp/ingestcheckpoint")
.trigger(Trigger.ProcessingTime(10.seconds))
.outputMode(OutputMode.Append)
.start //fails

Things I tried that didn't work:

  1. sc.hadoopConfiguration.set("parquet.enable.summary-metadata", "false")
  2. Change format to CSV instead of parquet
  3. Change output mode to Complete (only append is supported)
  4. Different Trigger intervals
  5. .option("failOnDataLoss", false) on readStream

Some digging into the source code took me here: https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/CompactibleFileStreamLog.scala where it says that absence of the .compact file should trigger defaults.

Hence tried: spark.conf.set("spark.sql.streaming.fileSink.log.cleanupDelay", 60000) to make sure that old batch's metadata isn't being removed before the new batch creates a combined metadata file

What makes this error annoying is that it isn't always reproducible. Without changing a single character in the code, writing to parquet will sometimes work, or won't. I have tried cleaning checkpoint locations, Spark/HDFS logs etc in case the "state" of Spark internals was causing this problem.

Here's the error stacktrace:

query: org.apache.spark.sql.streaming.StreamingQuery = org.apache.spark.sql.execution.streaming.StreamingQueryWrapper@56122c1

18/04/09 20:20:04 ERROR FileFormatWriter: Aborting job null.
java.lang.IllegalStateException: history/1523305060336/_spark_metadata/9.compact doesn't exist when compacting batch 19 (compactInterval: 10)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4$$anonfun$apply$1.apply(CompactibleFileStreamLog.scala:174)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4$$anonfun$apply$1.apply(CompactibleFileStreamLog.scala:174)
        at scala.Option.getOrElse(Option.scala:121)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4.apply(CompactibleFileStreamLog.scala:173)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4.apply(CompactibleFileStreamLog.scala:172)
        at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
        at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
        at scala.collection.immutable.NumericRange.foreach(NumericRange.scala:73)
        at scala.collection.TraversableLike$class.map(TraversableLike.scala:234)
        at scala.collection.AbstractTraversable.map(Traversable.scala:104)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog.compact(CompactibleFileStreamLog.scala:172)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog.add(CompactibleFileStreamLog.scala:156)
        at org.apache.spark.sql.execution.streaming.ManifestFileCommitProtocol.commitJob(ManifestFileCommitProtocol.scala:64)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$$anonfun$write$1.apply$mcV$sp(FileFormatWriter.scala:207)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$$anonfun$write$1.apply(FileFormatWriter.scala:166)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$$anonfun$write$1.apply(FileFormatWriter.scala:166)
        at org.apache.spark.sql.execution.SQLExecution$.withNewExecutionId(SQLExecution.scala:65)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$.write(FileFormatWriter.scala:166)
        at org.apache.spark.sql.execution.streaming.FileStreamSink.addBatch(FileStreamSink.scala:123)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch$1.apply$mcV$sp(StreamExecution.scala:666)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch$1.apply(StreamExecution.scala:666)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch$1.apply(StreamExecution.scala:666)
        at org.apache.spark.sql.execution.streaming.ProgressReporter$class.reportTimeTaken(ProgressReporter.scala:279)
        at org.apache.spark.sql.execution.streaming.StreamExecution.reportTimeTaken(StreamExecution.scala:58)
        at org.apache.spark.sql.execution.streaming.StreamExecution.org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch(StreamExecution.scala:665)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1$$anonfun$apply$mcZ$sp$1.apply$mcV$sp(StreamExecution.scala:306)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1$$anonfun$apply$mcZ$sp$1.apply(StreamExecution.scala:294)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1$$anonfun$apply$mcZ$sp$1.apply(StreamExecution.scala:294)
        at org.apache.spark.sql.execution.streaming.ProgressReporter$class.reportTimeTaken(ProgressReporter.scala:279)
        at org.apache.spark.sql.execution.streaming.StreamExecution.reportTimeTaken(StreamExecution.scala:58)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1.apply$mcZ$sp(StreamExecution.scala:294)
        at org.apache.spark.sql.execution.streaming.ProcessingTimeExecutor.execute(TriggerExecutor.scala:56)
        at org.apache.spark.sql.execution.streaming.StreamExecution.org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches(StreamExecution.scala:290)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anon$1.run(StreamExecution.scala:206)
18/04/09 20:20:04 ERROR StreamExecution: Query [id = 5251fe93-2b6b-4dff-bec3-7801dc7e6417, runId = 083547c1-69b7-40e7-8bf9-3c3af11d4c31] terminated with error
org.apache.spark.SparkException: Job aborted.
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$$anonfun$write$1.apply$mcV$sp(FileFormatWriter.scala:213)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$$anonfun$write$1.apply(FileFormatWriter.scala:166)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$$anonfun$write$1.apply(FileFormatWriter.scala:166)
        at org.apache.spark.sql.execution.SQLExecution$.withNewExecutionId(SQLExecution.scala:65)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$.write(FileFormatWriter.scala:166)
        at org.apache.spark.sql.execution.streaming.FileStreamSink.addBatch(FileStreamSink.scala:123)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch$1.apply$mcV$sp(StreamExecution.scala:666)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch$1.apply(StreamExecution.scala:666)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch$1.apply(StreamExecution.scala:666)
        at org.apache.spark.sql.execution.streaming.ProgressReporter$class.reportTimeTaken(ProgressReporter.scala:279)
        at org.apache.spark.sql.execution.streaming.StreamExecution.reportTimeTaken(StreamExecution.scala:58)
        at org.apache.spark.sql.execution.streaming.StreamExecution.org$apache$spark$sql$execution$streaming$StreamExecution$$runBatch(StreamExecution.scala:665)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1$$anonfun$apply$mcZ$sp$1.apply$mcV$sp(StreamExecution.scala:306)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1$$anonfun$apply$mcZ$sp$1.apply(StreamExecution.scala:294)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1$$anonfun$apply$mcZ$sp$1.apply(StreamExecution.scala:294)
        at org.apache.spark.sql.execution.streaming.ProgressReporter$class.reportTimeTaken(ProgressReporter.scala:279)
        at org.apache.spark.sql.execution.streaming.StreamExecution.reportTimeTaken(StreamExecution.scala:58)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anonfun$org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches$1.apply$mcZ$sp(StreamExecution.scala:294)
        at org.apache.spark.sql.execution.streaming.ProcessingTimeExecutor.execute(TriggerExecutor.scala:56)
        at org.apache.spark.sql.execution.streaming.StreamExecution.org$apache$spark$sql$execution$streaming$StreamExecution$$runBatches(StreamExecution.scala:290)
        at org.apache.spark.sql.execution.streaming.StreamExecution$$anon$1.run(StreamExecution.scala:206)
Caused by: java.lang.IllegalStateException: history/1523305060336/_spark_metadata/9.compact doesn't exist when compacting batch 19 (compactInterval: 10)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4$$anonfun$apply$1.apply(CompactibleFileStreamLog.scala:174)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4$$anonfun$apply$1.apply(CompactibleFileStreamLog.scala:174)
        at scala.Option.getOrElse(Option.scala:121)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4.apply(CompactibleFileStreamLog.scala:173)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog$$anonfun$4.apply(CompactibleFileStreamLog.scala:172)
        at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
        at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
        at scala.collection.immutable.NumericRange.foreach(NumericRange.scala:73)
        at scala.collection.TraversableLike$class.map(TraversableLike.scala:234)
        at scala.collection.AbstractTraversable.map(Traversable.scala:104)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog.compact(CompactibleFileStreamLog.scala:172)
        at org.apache.spark.sql.execution.streaming.CompactibleFileStreamLog.add(CompactibleFileStreamLog.scala:156)
        at org.apache.spark.sql.execution.streaming.ManifestFileCommitProtocol.commitJob(ManifestFileCommitProtocol.scala:64)
        at org.apache.spark.sql.execution.datasources.FileFormatWriter$$anonfun$write$1.apply$mcV$sp(FileFormatWriter.scala:207)
        ... 20 more
Bartosz Konieczny
  • 1,985
  • 12
  • 27
maverik
  • 774
  • 6
  • 17
  • 2
    Clearing checkpoint location at every re-run seemed to work but we abandoned EMR soon thereafter. Worth a try – maverik May 04 '18 at 16:21
  • @maverik What your final take on this. I'm facing the problem right now. Was it just EMR or else ? – MaatDeamon Sep 16 '18 at 18:19
  • @MaatDeamon seemed to work out for us when we stopped writing to s3. downside: you need a driver node with large (and fast) EBS storage. don't checkpoint in s3, since its not a real filesystem which can support ordering of files etc. finally take it with a grain of salt since we didn't use EMR a long time and migrated to Databricks, if you have enterprise support talk to AWS staff engineers. This is not a Spark problem, more of an AWS glitch. – maverik Sep 18 '18 at 18:08
  • You said you quit EMR, but you can use HDFS with it, isn’t it ? – MaatDeamon Sep 18 '18 at 18:23
  • ^ EMRFS is just a layer on HDFS – maverik Sep 24 '18 at 18:53

3 Answers3

3

It turns out that S3 does not support the read-after-write semantics needed by Spark checkpointing.

This article suggests using AWS EFS for checkpointing.

S3 remains a good place to ingest data from, or egest data to.

Jacek Laskowski
  • 72,696
  • 27
  • 242
  • 420
Benedetto
  • 711
  • 6
  • 17
1

I solved this question by clearing my checkpoint path:

  1. remove your checkpoint path:

    sudo -u hdfs hdfs dfs -rmr ${your_checkpoint_path}

  2. resubmit your spark job.

Community
  • 1
  • 1
user2894829
  • 775
  • 1
  • 6
  • 26
  • For me also it worked after changing the checkpoint location. Why is it happening like that? – Soumendra Apr 30 '19 at 04:28
  • 1
    @Soumendra Looks like the underlying filesystem that EMR uses is very "unstable" and unreliable so when a streaming query checkpoints its state it simply needs a better filesystem (not whatever EMR uses). – Jacek Laskowski Jul 18 '19 at 09:41
  • 1
    I'd not remove checkpoint directory at every start of a streaming query since the purpose of the directory is to keep state so you won't re-process events and have incorrect results. You should keep a checkpoint directory healthy at all cost and remove it only as your last resort (read: never). What you propose is a workaround not a solution. – Jacek Laskowski Jul 18 '19 at 09:43
  • @Jacek Laskowski Master how to keep a checkpoint directory healthy ? what kind of things should be taken care of for the same ? – BdEngineer Jun 11 '20 at 09:44
0

This error message can show up when the file sink path has changed between re-starts of the structured streaming job or the file system of the output path is unreliable and data can be lost between re-starts.

In case you changed file sink output paths you could copy the missing data from

/old/file/sink/path/_spark_metadata/9.compact

to

/new/file/sink/path/_spark_metadata/9.compact

I have given more details on the background in another answer

Michael Heil
  • 16,250
  • 3
  • 42
  • 77