How to Serve Data Science Models with OpenFaaS

Lucas shows you how you can build and deploy a name classification function using OpenFaaS and PyTorch

Last month I was at Kubecon Barcelona with some of the OpenFaaS community, we were asked about deploying data science models to Kubernetes and, of course, can OpenFaaS help deploy models? In this post we will introduce a new function template aimed at Python data scientists and walk through a concrete example of deploying a PyTorch model.

The pydatascience template

The pydatascience template showcases a couple of ideas that anyone can leverage in their own functions and templates:

  1. using the Conda package manager
  2. setting up a non-root user Python environment
  3. multi-module Python function
  4. using HTTP mode in of-watchdog to load an asset (in this case a model) into memory during startup

The name classifier function

In this post, we share an example name classifier. It is a relatively simple function that accepts a name and then attempts to guess the nationality of that name, returning the top three guesses.

echo "Lucas" | faas-cli invoke classify
[
  ["(-1.4300)","Irish"],
  ["(-1.6112)","Italian"],
  ["(-2.0314)","Dutch"]
]

You can try this in your own OpenFaaS cluster using

faas-cli deploy --image=theaxer/classify:latest --name=classify

Caution, this model was built on a relatively small data set and is intended to be a simple demonstration. And remember, with all things data science, “with great power comes great responsibility”.

Understanding the template

What is Conda?

I have recently started to using the Conda package manager for many of my own experiments with Python, especially when they involve Pandas and Numpy, because the install is very fast and environment management is fairly easy. For function development and build, the speed is really nice. For this project it was also the easiest way to install PyTorch. Finally, Conda also supports non-Python packages, e.g. curl, yarn-js, and zeromq are installable via Conda.

Set up a non-root function

A best practice for deploying Docker images is to make sure that the user running your code is not privileged, this is also known as “non-root”. These images tend to be much safer to deploy. Fortunately, OpenFaaS makes it easy to enforce, even when the original image is not non-root by default. When we build new templates, it is important that we consider non-root deployments as the default. For Python environments, Conda made this very easy.

RUN addgroup app && adduser app --system --ingroup app \
    && mkdir -p /opt/conda && chown -R app /opt/conda

ENV HOME /home/app
ENV PATH=$HOME/conda/bin:$PATH

RUN apt-get update \
    && apt-get -y install curl bzip2 ${ADDITIONAL_PACKAGE} \
    && curl -sSL https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -o /tmp/miniconda.sh \
    && curl -sSL https://github.com/openfaas-incubator/of-watchdog/releases/download/0.5.1/of-watchdog > /usr/bin/fwatchdog \
    && chown app /tmp/miniconda.sh \
    ...

A multi-module template

Earlier this year we highlighted how to use multiple files and modules to organize your Python functions. This new template follows this pattern out of the box. It provides a “core” module for putting utility methods, a handler file, which is where you need to implement your function logic, and a “training” file where you can put the logic for training your model.

function
├── __init__.py
├── core
│   ├── README.md
│   ├── __init__.py
│   └── utils.py
├── handler.py
├── requirements.txt
└── train.py

Like the basic python template, only the handler.py implementation is required, and both the core module and the train.py could be replaced if you needed, but this structure starts you on a path to keeping the code organized.

From a template to actual code

The name classifier function uses the neural network implementation in PyTorch. PyTorch has a great introduction and walk-through for neural network package and model training. In this post we focus on what a concrete and complete implementation looks like. All of the interesting details about the implementation of the model, the utilities used to parse the training data, and the utility for applying the model to a new inputs are stored in the core module. This core module can then be used in both the function handler and the training script. This code is also easier to reuse in a new function because it is cleanly isolated from the function itself.

Train the model

Another point to note is the training data folder is also included here, data/names and a serialized model is also include data/char-rnn-classification.pt.

This template is designed to run the training as part of the build process. You can easily trigger the training using

python train.py

There are two methods for saving PyTorch models:

  1. save the entire model using pickle, or
  2. saving just the model state_dict.

We have opted for the second method because it is more portable, the pickle method is sensitive to the Python environment and folder structure.

torch.save(rnn.state_dict(), "data/char-rnn-classification.pt")

Loading the model into function memory

Using the HTTP mode in the of-watchdog enables us to load this model into memory once and reuse it for multiple requests. We only need to load the model at the start of the handler.py implementation

import json
import os
from pathlib import PurePath
from typing import Any, List

import torch
from torch.autograd import Variable

from .core import const, model, utils

FUNCTION_ROOT = os.environ.get("function_root", "/home/app/function/")

# init model
RNN = model.RNN(const.N_LETTERS, const.N_HIDDEN, const.N_CATEGORIES)
# fill in weights
RNN.load_state_dict(
    torch.load(str(PurePath(FUNCTION_ROOT, "data/char-rnn-classification.pt")))
)

def predict(line: str, n_predictions: int = 3) -> List[Any]:
    """omitted for brevity"""

def handle(req: bytes) -> str:
    """handle a request to the function
    Args:
        req (bytes): request body
    """

if not req:
        return json.dumps({"error": "No input provided", "code": 400})

name = str(req)
    output = predict(name)

return json.dumps(output)

When a function instance is instantiated, e.g. during scaling or deployment, this file is parsed and the model is loaded into memory. This will happen exactly once because the HTTP mode loads this handler as a very small background web server. The original forking mode in the watchdog would instead load this file for every invocation. When loading models, this creates additional latency.

Deployment

This template is designed to bundle the pre-trained model into the final Docker image. This means that the deployment steps look like this

cd classify && python train.py && cd ..
faas-cli build classify

This results in a completely self-contained Docker image that does not need access to a database or S3 during runtime. This provides several benefits:

  1. This model becomes very easy to share because you don’t need to share access to your model storage.
  2. It also means that each new deployment is completely versioned, you know the version of the model from the Docker tag. This makes rolling backward or forward as simple as changing the Docker tag on the deployment.
  3. One final benefit is that startup time can be slightly faster because once a node has the Docker image in its cache, it has everything it needs to start new instances of the function. If you load the model from an external source, then every function instance that starts must curl/copy from that external location, there is no way to cache it for new instances.

On the other hand, this does result in slightly larger images, in this case the model is 107k. This size will, of course, depend on the libraries being used and the type of model being serialized. For many models, the simplicity of a self-contained image outweighs the cost of the additional size. Often Docker layer caching will hide almost all of it.

You can try the latest version of this function in your own cluster using

faas-cli deploy --image=theaxer/classify:latest --name=classify

A note about Python 2 deprecation

Remember that Python 2 will reach end of life on 1/1/2020. For the impact on your project, see https://python3statement.org

Wrapping up

Data science is of course a popular topic these days and OpenFaaS can help simplify, standardize, and speed up the deployment process for your team. This post introduces just one example of what a data science focused template could look like.

Have you struggled to package your machine-learning model for platforms like AWS Lambda? OpenFaaS functions don’t have an arbitrary size limit and may provide an easier alternative for you.