[BEAM-5500] Portable python sdk worker leaks memory in streaming mode

• Type: Bug
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Fix Version/s: 2.8.0
• Component/s: sdk-py-harness
• Labels: portability-flink
• Assignee: Robert Bradshaw
• Reporter: Micah Wylde
• Created: 25/Sep/18 18:15
• Updated: 16/May/20 13:59
• Resolved: 03/Oct/18 18:40

Description

When using the portable python sdk with flink in streaming mode, we see that the python worker processes steadily increase memory usage until they are OOM killed. This behavior is consistent across various kinds of streaming pipelines, including those with fixed windows and global windows.

A simple wordcount-like pipeline demonstrates the issue for us (note this is run on the Lyft beam fork, which provides access to kinesis as a portable streaming source):

counts = (p
| 'Kinesis' >> FlinkKinesisInput().with_stream('test-stream')
| 'decode' >> beam.FlatMap(decode) # parses from json into python objs
| 'pair_with_one' >> beam.Map(lambda x: (x["event_name"], 1))
| 'window' >> beam.WindowInto(window.GlobalWindows(),
                              trigger=AfterProcessingTime(15 * 1000),
                              accumulation_mode=AccumulationMode.DISCARDING)
| 'group' >> beam.GroupByKey()
| 'count' >> beam.Map(count_ones)
| beam.Map(lambda x: logging.warn("count: %s", str(x)) or x))

When run, we see a steady increase in memory usage in the sdk_worker process. Using heapy I've analyzed the memory usage over time and found that it's largely dicts and strings (see attached chart).

 

 

 

Issue Links

• links to: GitHub Pull Request #6517