Received message larger than max (4199881 vs. 4194304)

[pselden]

I have a TFX pipeline that runs in Kubeflow on GCP and recently one of my pipelines started failing with the following error in a ResolverNode.latest_model_resolver and ResolverNode.latest_blessed_model_resolver

Traceback (most recent call last):
  File "/usr/local/lib/python3.7/dist-packages/ml_metadata/metadata_store/metadata_store.py", line 165, in _call_method
    response.CopyFrom(grpc_method(request))
  File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 826, in __call__
    return _end_unary_response_blocking(state, call, False, None)
  File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 729, in _end_unary_response_blocking
    raise _InactiveRpcError(state)
grpc._channel._InactiveRpcError: <_InactiveRpcError of RPC that terminated with:
	status = StatusCode.RESOURCE_EXHAUSTED
	details = "Received message larger than max (4199881 vs. 4194304)"
	debug_error_string = "{"created":"@1603760693.874743930","description":"Received message larger than max (4199881 vs. 4194304)","file":"src/core/ext/filters/message_size/message_size_filter.cc","file_line":203,"grpc_status":8}"
>
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
  File "/tfx-src/tfx/orchestration/kubeflow/container_entrypoint.py", line 360, in <module>
    main()
  File "/tfx-src/tfx/orchestration/kubeflow/container_entrypoint.py", line 353, in main
    execution_info = launcher.launch()
  File "/tfx-src/tfx/orchestration/launcher/base_component_launcher.py", line 197, in launch
    self._exec_properties)
  File "/tfx-src/tfx/orchestration/launcher/base_component_launcher.py", line 166, in _run_driver
    component_info=self._component_info)
  File "/tfx-src/tfx/components/common_nodes/resolver_node.py", line 73, in pre_execution
    source_channels=input_dict.copy())
  File "/tfx-src/tfx/dsl/experimental/latest_artifacts_resolver.py", line 56, in resolve
    output_key=c.output_key)
  File "/tfx-src/tfx/orchestration/metadata.py", line 323, in get_qualified_artifacts
    executions = self.store.get_executions_by_context(context.id)
  File "/usr/local/lib/python3.7/dist-packages/ml_metadata/metadata_store/metadata_store.py", line 1080, in get_executions_by_context
    self._call('GetExecutionsByContext', request, response)
  File "/usr/local/lib/python3.7/dist-packages/ml_metadata/metadata_store/metadata_store.py", line 140, in _call
    return self._call_method(method_name, request, response)
  File "/usr/local/lib/python3.7/dist-packages/ml_metadata/metadata_store/metadata_store.py", line 170, in _call_method
    raise _make_exception(e.details(), e.code().value[0])  # pytype: disable=attribute-error
ml_metadata.errors.ResourceExhaustedError: Received message larger than max (4199881 vs. 4194304)

Is there a way to fix this on my side?

[pselden]

I was able to fix this by deleting some items from the Associations table for the given context.

This is obviously just a bandaid. It seems like the root of this problem is that since there's no way to do filtering in ml-metadata, that TFX has to load ALL executions just to be able to choose the latest one.

[hughmiao]

hi @pselden , you can use the grpc options to increase the size.

/cc @ruoyu90 on the resolver logic refactoring.

We also also working on filtering. Please stay tuned.

[hughmiao]

@pselden, related to the discussion here, we will surface the pagination API to the python client.

[ConverJens]

hi @pselden , you can use the grpc options to increase the size.

/cc @ruoyu90 on the resolver logic refactoring.

We also also working on filtering. Please stay tuned.

@hughmiao I encountered the same issue when running the Evaluator component in TFX while slicing on continuous features. I'm also running MLMD in Kubeflow, is there a way to set the grpc max message length as a command line option, like grpc-port?

[hughmiao]

@ConverJens , the config can be passed in the tfx pipeline mlmd config settings.

I'm also running MLMD in Kubeflow, is there a way to set the grpc max message length as a command line option, like grpc-port?

For kubeflow deployment, + @dushyanthsc for KFP settings / command line options.

[redramen]

... we will surface the pagination API to the python client.

@hughmiao do you happen to have an ETA for this? We are building atop MLMD at Twitter and pagination is a must have for us.

[dushyanthsc]

@redramen I should have the pagination support surfaced in the python client by end of this week.

[hughmiao]

@redramen sg, will prioritize this.

thanks, @dushyanthsc ! Let's follow up in the cl.

/cc @ruoyu90 for tfx side changes if needed.

[ConverJens]

@ConverJens , the config can be passed in the tfx pipeline mlmd config settings.

I'm also running MLMD in Kubeflow, is there a way to set the grpc max message length as a command line option, like grpc-port?

For kubeflow deployment, + @dushyanthsc for KFP settings / command line options.

@dushyanthsc Any documentation on how to specify this?

[dushyanthsc]

@ConverJens we plan to use the ListOperationOptions(https://github.com/google/ml-metadata/blob/master/ml_metadata/proto/metadata_store.proto#L638) in gRPC service layer and have the python client get executions by calling the gRPC with page size.

I am working on the CL, and should have a point release out early next week.

[dushyanthsc]

@ConverJens @redramen A solution for the problem is checked in and available at HEAD.

The solution was to use the pagination support and retrieve Executions in page of 100 executions per page and abstract this logic behind get_executions_by_context.

If you can elaborate how you consume MLMD i.e. from source code or released version or through TFX we can decide we need to cut a point release.

[ConverJens]

@dushyanthsc That sounds great for the paging problem. However, my problem was that the proto message from a single Evaluator run was too large, hence I would like to specify this option for MLMD in the KubeFlow installation, not just from a client. Any idea how this can be achieved? Perhaps as a command line option, like grpc-port?

And regarding how we consume MLMD: through Jupyter using the kubeflow-metadata package and through TFX pipelines. The pagination is not a blocker for our use case at the moment.

[dushyanthsc]

@ConverJens What is the underlying MLMD API call that the Evaluator makes? Can you provide the log of the error you are seeing, that way I can confirm if the change made solves your problem.

For providing the config flag to increase the allowed payload size. You can set the command line parameter [1] which for a kubeflow-metadata deployment gets passed from the deployment manifest [2]

[1] - https://github.com/google/ml-metadata/blob/master/ml_metadata/metadata_store/metadata_store_server_main.cc#L142

[2] - https://github.com/kubeflow/manifests/blob/master/metadata/base/metadata-deployment.yaml#L26