tobe

Thanks @petro-rudenko . That may help and we can test for the new client in our environment.

@RummySugar You need to install libdisni so in each server or upload with `spark-submit`.

It seems the travis-ci still use `libboost1.55-all-dev` which fail the tests in the PR. Ping @ipkn to fix for boost 1.7.

You can export with the preprocess python function in the SavedModel. This needs more document and you may find it in the source code.

We can not hot-reload models with config file or using non-tensorfllow models now. We may use load balancer or AB testing tools for this scenario.

It seems that you have serialize the `preprocess_function` as byte object but not UTF string. Please paste more log or your testing script so that we can re-produce your problem.

@yhfwww you can shutdown the service by killing the process? Use ctrl-c to send the int signal or `docker stop` the container.

Thanks for reporting. You can pass the model path as the specified HDFS path. We may add more detailed document about this soon.

Could you please post the files in this HDFS directory? Try with `hadoop dfs -ls hdfs://GZGD/171587/1/model/NeuralCF/base-model-output`. We have to make sure that the output path is the SavedModel files with...

So the `1551684206` is the model version, right? And are there `saved_model.pb` and `variables` in these directories of the versions?