Deploying ML models with CPU based TFServing, Docker, and Kubernetes
This project shows how to serve a TensorFlow image classification model as RESTful and gRPC based service with TFServing, Docker, and Kubernetes. The idea is to first create a custom TFServing docker image with a TensorFlow model, and then deploy it on a k8s cluster running on Google Kubernetes Engine (GKE). Also we are using GitHub Actions to automate all the procedures when a new TensorFlow model is released.
- Even though this project uses an image classification its structure and techniques can be used to serve other models as well.
- There is a counter part project using FastAPI instead of TFServing. If you wonder from how to convert TensorFlow model to ONNX optimized model to deploy it on k8s cluster, check out the this repo.
Deploying the model as a service with k8s
- Prerequisites: Doing anything beforehand, you have to create GKE cluster and service accounts with appropriate roles. Also, you need to grasp GCP credentials to access any GCP resources in GitHub Action. Please check out the more detailed information here
flowchart LR A[First: Environmental Setup]-->B; B[Second: Build TFServing Image]-->C[Third: Deploy on GKE];
- To deploy a custom TFServing docker image, we define
deployment.ymlworkflow file which is is only triggered when there is a new release for the current repository. It is subdivided into three parts to do the following tasks:
- First subtask handles the environmental setup.
- GCP Authentication (GCP credential has to be provided in GitHub Secret)
- Install gcloud CLI toolkit
- Authenticate Docker to push images to GCR(Google Cloud Registry)
- Connect to the designated GKE cluster
- Second subtask handles building a custom TFServing image.
- Download and extract the latest released model from the current repository
- Run the CPU optimized TFServing image which is compiled from the source code (FYI. image tag is
gcr.io/gcp-ml-172005/tfs-resnet-cpu-opt, and it is publicly available)
- Copy the extracted model into the running container
- Commit the changes of the running container and give it a new image name
- Push the commited image
- Third subtask handles deploying the custom TFServing image to GKE cluster.
- Pick a one of the scenarios from a various experiments
- Download Kustomize toolkit to handle overlay configurations.
- Update image tag with the currently built one with Kustomize
- By provisioning
ConfigMap, the custom TFServing image gets deployed.
ConfigMapis only used for batching enabled scenarios to inject batching configurations dynamically into the
- In order to use this repo for your own purpose, please read this document to know what environment variables have to be set.
- First subtask handles the environmental setup.
If the entire workflow goes without any errors, you will see something silimar to the text below. As you see, two external interfaces(8500 for RESTful, 8501 for gRPC) are exposed. You can check out the complete logs in the past runs.
NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE tfs-server LoadBalancer xxxxxxxxxx xxxxxxxxxx 8500:30869/TCP,8501:31469/TCP 23m kubernetes ClusterIP xxxxxxxxxx <none> 443/TCP 160m
We used Locust to conduct load tests for both TFServing and FastAPI. Below is the results for TFServing(gRPC) on a various setups, and you can find out the result for FastAPI(RESTful) in a separate repo. For specific instructions about how to install Locust and run a load test, follow this separate document.
- This is a follow-up project after ONNX optimized FastAPI deployment, so we wanted to know how CPU optimized TensorFlow runtime could be compared to ONNX based one.
- TFServing's objective is to maximize throughput while keeping tail-latency below certain bounds. We wanted to see if this is true, how reliably it provides a good throughput performance and how much throughput is sacrified to keep the reliability.
- According to the TFServing's official document, TFServing can achieve the best performance when it is deployed on fewer, larger(in terms of CPU, RAM) machines. We wanted to estimate how large of machine and how many nodes are enough. For this, we have prepared a set of different setups in combination of (# of nodes + # of CPU cores + RAM capacity).
- TFServing has a number of configurable options to tune the performance. Especially, we wanted to find out how different values of
--enable_batchingoptions gives different results.
From the results above,
- TFServing focuses more on reliability than performance(in terms of throughput). In any cases, no failures are observed, and the the response time is consistent.
- Req/s is lower than ONNX optimized FastAPI deployment, so it sacrifies some performance to achieve reliability. However, you need to notice that TFServing comes with lots of built-in features which are required in most of ML serving scenarios such as multi model serving, dynamic batching, model versioning, and so on. Those features possibly make TFServing heavier than simple FastAPI server.
- NOTE: We spawned requests every seconds to clearly see how TFServing behaves with the increasing number of clients. So you can assume that the Req/s doesn't reflect the real world situation where clients try to send requests in any time.
- 8vCPU + 16GB RAM seems like large enough machine. At least bigger size of RAM doesn't help much. We might achieve better performance if we increase the number of CPU core than 8, but beyond 8 cores is somewhat costly.
- In any cases, the optimal value of
--tensorflow_inter_op_parallelismseems like 4. The value of
--tensorflow_intra_op_parallelismis fixed to the number of CPU cores since it specifies the number of threads to use to parallelize the execution of an individual op.
--enable_batchingcould give you better performance. However, since TFServing doesn't immediately response to each requests, there is a trade-off.
- By considering cost trade-off, our recommendation from the experiment is to choose
2n-8c-16r-interop4configuration unless you care about dynamic batching capabilities. Or you can write a similar setup by referencing
2n-8c-16r-interop2-batchbut for smaller machines as well.
- Locust doesnt' have a built-in support to write a gRPC based client, so we have written one for ourselves. If you are curious about the implementation, check this locustfile.py out.
- For the legend in the plot,
nmeans the number of nodes(pods),
cmeans the number of CPU cores,
rmeans the RAM capacity,
interopmeans the number of
batchmeans the batching configuration is enabled with this config.
Advancing this repo by providing a semi-automatic model deployment. To be more specific, when new codes implementing new ML model is pull requested, maintainers could trigger model performance evaluable on GCP's Vertex Training via
comments. The experiment results could be exposed through TensorBoard.dev or W&B. If it is approved, the code will be merged, the trained model will be released, and it is going to be deployed on GKE.
ML-GDE program for providing GCP credit support.