Hi,
We are investigating an issue with our Python SDK streaming pipelines,
and have few questions, but first context.
Our stack:
- Python SDK 2.54.0 but we tried also 2.55.1
- DataFlow Streaming engine with sdk in container image (we tried also
Prime)
- Currently our pipelines do have low enough traffic, so that single
node handles it most of the time, but occasionally we do scale up.
- Deployment by Terraform `google_dataflow_flex_template_job` resource,
which normally does job update when re-applying Terraform.
- We do use a lot `ReadModifyWriteStateSpec`, other states and watermark
timers, but we do keep a the size of state under control.
- We do use custom coders as Pydantic avro.
The issue:
- Occasionally watermark progression stops. The issue is not
deterministic, and happens like 1-2 per day for few pipelines.
- No user code errors reported- but we do get errors like this:
```INTERNAL: The work item requesting state read is no longer valid on
the backend. The work has already completed or will be retried. This is
expected during autoscaling events.
[type.googleapis.com/util.MessageSetPayload='[dist_proc.dax.internal.TrailProto]
{ trail_point { source_file_loc { filepath:
"dist_proc/windmill/client/streaming_rpc_client.cc" line: 767 } } }']```
```ABORTED: SDK harness sdk-0-0 disconnected. This usually means that
the process running the pipeline code has crashed. Inspect the Worker
Logs and the Diagnostics tab to determine the cause of the crash.
[type.googleapis.com/util.MessageSetPayload='[dist_proc.dax.internal.TrailProto]
{ trail_point { source_file_loc { filepath:
"dist_proc/dax/workflow/worker/fnapi_control_service.cc" line: 217 } } }
[dist_proc.dax.MessageCode] { origin_id: 5391582787251181999
[dist_proc.dax.workflow.workflow_io_message_ext]: SDK_DISCONNECT }']```
```Work item for sharding key 8dd4578b4f280f5d tokens
(1316764909133315359, 17766288489530478880) encountered error during
processing, will be retried (possibly on another worker):
generic::internal: Error encountered with the status channel: SDK
harness sdk-0-0 disconnected. with MessageCode: (93f1db2f7a4a325c): SDK
disconnect.```
```Python (worker sdk-0-0_sibling_1) exited 1 times: signal:
segmentation fault (core dumped) restarting SDK process```
- We did manage to correlate this with either vertical autoscaling event
(when using Prime) or other worker replacements done by Dataflow under
the hood, but this is not deterministic.
- For few hours watermark progress does stop, but other workers do
process messages.
- and after few hours:
```Error message from worker: generic::internal: Error encountered with
the status channel: There are 10 consecutive failures obtaining SDK
worker status info from sdk-0-0. The last success response was received
3h20m2.648304212s ago at 2024-04-23T11:48:35.493682768+00:00. SDK worker
appears to be permanently unresponsive. Aborting the SDK. For more
information, see:
https://cloud.google.com/dataflow/docs/guides/common-errors#worker-lost-contact```
- And the pipeline starts to catch up and watermark progresses again.
- Job update by Terraform apply also fixes the issue.
- We do not see any extensive use of worker memory nor disk. CPU
utilization is also most of the time close to idle. I do not think we do
use C/C++ code with python. Nor use parallelism/threads outside beam
parallelization.
Questions:
1. What could be potential causes of such behavior? How to get more
insights to this problem?
2. I have seen `In Python pipelines, when shutting down inactive bundle
processors, shutdown logic can overaggressively hold the lock, blocking
acceptance of new work` in Beam release docs as known issue. What is the
status of this? Can this potentially be related?
Really appreciate any help, clues or hints how to debug this issue.
Best regards
Wiśniowski Piotr
