yangchenghuang-docusign / streaming-model-serving Goto Github PK

This repository contains the implementation and experiments configuration to compare different model serving alternatives over streaming data. Our repository is organized as it follows: the implementation/ diretcory contains the code for all the benchmarked pipelines, alonside the scripts to run the experiments and the corresponding evaluation configurations. The models/ directory contains: (1) the scripts used to train the TensorFlow and PyTorch models (under models/training/), and (2) the pre-trained feed-forward models (under models/feedforward/).

1. Unix-like environment
2. Maven
3. Java 8
4. Docker installation

All the different benchmarks can be executed using their corresponding run_benchmark.sh shell script. The script will first run the experiments (with their corresponding configurations), write the results to disk, and then compute the quality metrics per experiment type. The results will be printed to the console.

The code for the embedded inference approach is contained inside implementation/flink. The run_benchmark.sh shell script is used to experiment with all the embedded serving libraries (i.e., ND4J, ONNX, and TensorFlow SavedBundle). Simply running the run_benchmark.sh shell script will print all the computed quality metrics to the console. Thus, one only needs to execute implementation/flink/run_benchmark.sh once to obtain the ND4J, ONNX, and TensorFlow SavedBundle results.

We test two different model serving systems: Tensorflow Serving, and TorchServe. The code and experiments configurations are located at implementation/tensorflowserve and implementation/torchserve respectively. The run_benchmark.sh shell script will run the evaluation for each approach. Please note that the corresponding shell scripts must be executed manually for each approach: execute implementation/tensorflowserve/run_benchmark.sh to gather the TensorFlow Serving results and implementation/torchserve/run_benchmark.sh for the TorchServe tests.

The experiment configuration files are located at implementation/{flink | tensorflowserve | torchserve}/expconfigs/ and can be adjusted to run the experiments with different configurations.