beam: [Bug]: IcebergIO - Write performance issues
What happened?
I am trying to stream data from PubSub (with a throughput of 10-50 RPS) into an Iceberg Table (not partitioned) using the IcebergIO connector and Hive Metastore.
However, after some time, I see warning logs in the console like below:
...
Operation ongoing in bundle for at least 40m00s without completing Current user step name: Managed.ManagedTransform/ManagedSchemaTransformProvider.ManagedSchemaTransform/IcebergWriteSchemaTransformProvider.IcebergWriteSchemaTransform/IcebergIO.WriteRows/Write Rows
Operation ongoing in bundle for at least 50m00s without completing Current user step name: Managed.ManagedTransform/ManagedSchemaTransformProvider.ManagedSchemaTransform/IcebergWriteSchemaTransformProvider.IcebergWriteSchemaTransform/IcebergIO.WriteRows/Write Rows
Operation ongoing in bundle for at least 01h00m00s without completing Current user step name: Managed.ManagedTransform/ManagedSchemaTransformProvider.ManagedSchemaTransform/IcebergWriteSchemaTransformProvider.IcebergWriteSchemaTransform/IcebergIO.WriteRows/Write Rows
Operation ongoing in bundle for at least 01h10m00s without completing Current user step name: Managed.ManagedTransform/ManagedSchemaTransformProvider.ManagedSchemaTransform/IcebergWriteSchemaTransformProvider.IcebergWriteSchemaTransform/IcebergIO.WriteRows/Write Rows
...
From the thread dump (see threadump.txt) it looks like a significant number of DataflowWorkUnits threads (287) are in the TIMED_WAITING state, specifically waiting for a Hive client from the pool.
java.base@11.0.20/java.lang.Object.wait(Native Method)
app//org.apache.iceberg.ClientPoolImpl.get(ClientPoolImpl.java:133)
However, I find it surprising that all these threads are attempting to create a new writer each time, which results in the concurrent reload of the table we saw above. Is this behavior expected? I suspect that the performance issues stem from the RecordWriterManager class not being thread-safe. Specifically, it appears that a race condition is occurring in this code snippet due to the check-then-act logic: https://github.com/apache/beam/blob/4cf99ea31aa2f3b4e8fa0da479939bbe92bd46dd/sdks/java/io/iceberg/src/main/java/org/apache/beam/sdk/io/iceberg/RecordWriterManager.java#L151-L161
Indeed, when I stop the job I am prompted with the error message: the evidence that too many writers got created.
java.lang.IllegalArgumentException: Expected all data writers to be closed, but found 8 data writer(s) still open
at org.apache.beam.vendor.guava.v32_1_2_jre.com.google.common.base.Preconditions.checkArgument(Preconditions.java:191)
at org.apache.beam.sdk.io.iceberg.RecordWriterManager.close(RecordWriterManager.java:279)
at org.apache.beam.sdk.io.iceberg.WriteUngroupedRowsToFiles$WriteUngroupedRowsToFilesDoFn.finishBundle(WriteUngroupedRowsToFiles.java:241)
Issue Priority
Priority: 2 (default / most bugs should be filed as P2)
Issue Components
- Component: Python SDK
- Component: Java SDK
- Component: Go SDK
- Component: Typescript SDK
- Component: IO connector
- Component: Beam YAML
- Component: Beam examples
- Component: Beam playground
- Component: Beam katas
- Component: Website
- Component: Infrastructure
- Component: Spark Runner
- Component: Flink Runner
- Component: Samza Runner
- Component: Twister2 Runner
- Component: Hazelcast Jet Runner
- Component: Google Cloud Dataflow Runner
About this issue
- Original URL
- State: closed
- Created 2 years ago
- Comments: 25 (15 by maintainers)
Yep the triggering frequency was one of the recently added features. Streaming writes before 2.60.0 may work in some cases but its rather unpredictable.
This is also one of the improvements – we drastically reduced the number of manifest files we write. From my experience, the more manifest files we have, the longer it takes to update the next one
Let me know how the new pipeline goes!