Salman IqbalSalman Iqbal

Scaling Jupyter notebooks with Kubernetes and Tensorflow

Published in January 2019

Big data

Gathering facts and data to understand better the world we live in has become the new norm.

From self-driving cars to smart personal assistants, data and data science is everywhere.

Even the phones that we carry in our pockets now feature dedicated units for machine learning.

In fact, there has never been more need for a more performant and efficient system to ingest and extract meanings out of large volumes of numbers.

The challenges for data scientists and engineers is how to design pipelines and processes that can operate at scale and in real-time.

Traditionally, the setup of the services and tools required for effective development and deployment of deep learning model has been time-consuming and error-prone.

Build once, run everywhere on my machine

Being able to run your models locally and reproduce the same results in the data farm is what you should expect.

But it's not always like this.

And the issues are not new to any kind of software development either.

You write an algorithm on your computer and test it by training it locally to produce a model.

It works well, and you are proud of your work.

But when you send it to the data farm or your colleague tries to run it, it fails.

It's only after carrying out extensive checks that you realise that there is a conflict of runtime version — your colleague is still stuck at Python 2.7!

It would be much easier if you could send your models alongside your environment and settings.

You could genuinely have reproducible results independently from when you run your tests.

Particularly when you need a larger computer to process even more data.

When your computer is not enough

There comes the point while developing models that the computer you are using will not be enough to train your model.

Sure, you could buy a powerful gaming PC and stuff it with video cards.

However, you will have to wait for it to be delivered and set it all up with the required pre-requisites.

And if you are working in a team and they all require the same setup, buying a powerful machine for each of the members will not be financially viable.

You won't have a good return on investment when you go on holiday.

The kit stays under your desk and collects dust.

A better strategy would be to rent the kit only when you need it.

So you can pay only for what you use.

Perhaps you could use a workstation in the cloud so that even your colleagues can provision large clusters when and if they needed.

The whole setup is too complex

As a data scientist, your expertise and duties span several disciplines, such as provisioning clusters, researching and training models, designing the infrastructure to ingest and extract meaningful information and so on.

On the other hand data scientists need to focus on a handful of things such as creating, testing and evaluating models to deliver the results that the business expect from them.

null

The reality, however, is very sobering.

You have to deal with a myriad of things.

From setting up the infrastructure that can serve the required models, to configuring monitoring tools, tuning resource management, supporting batched jobs and the list goes on!

To become a data scientist, one has to quickly grasp and master other aspects of software development, operations and infrastructure management.

Even if you don't want to, unfortunately.

Scaling machine learning in the cloud

But what if you could outsource all of the non-data science to someone else while still retaining control?

What if you could leverage an existing platform for scaling your models without having to reinvent the wheel?

What if you could train and evaluate your model in the cloud as you do locally?

You could finally focus on what you do best: data science!

Few competing solutions provide an easy entry to a data farm.

You might have heard of:

They are all valid solutions, but they tend to run in public clouds. Additionally, you can't quickly move to another once you committed to one.

Kubeflow — an open source machine learning platform

An excellent alternative for training and evaluating your models in public and private clouds is to use Kubeflow — an open-source toolkit for distributed machine learning.

Kubeflow is designed to make your machine learning experiments portable and scalable.

You start by creating Jupyter notebooks in the cloud.

Since you can rent kit by the hour, you can run your experiment on large compute resources with dedicated hardware such as GPUs and TPUs.

When you sense that you have hit a breakthrough, you can scale your model to run on thousands of machines.

Kubeflow optimises your model and breaks it down into smaller tasks that can be processed in parallel.

Then, it distributes the tasks to several computers and waits until the results are ready.

You can easily rent servers to run thousands of jobs allowing you to train your model orders of magnitude quicker than you're used to!

When the results are collected, you can use Kubeflow built-in web server to offer an API to query the model.

No need to create your own web server and plug in the algorithm. Kubeflow offers that out of the box.

Things get even better.

Kubeflow can run parametrised jobs.

That's great news if you need to tune your hyperparameters.

Instead of starting a single job and having it broken down to smaller tasks, you can submit several jobs and let Kubeflow optimise the processing for you.

And while you wait, you can observe the progress of each model on the built-in dashboard.

Let me recap what Kubeflow does for you:

  • deploys a JupyterHub where you can start notebooks with beefy machines
  • breaks down your models into smaller tasks that can be parallelised
  • serves your models as an API to be queried by other services
  • can run parametrised jobs in parallel — ideal for hyperparameter sweeping

Exciting!

All those annoying tasks that take you away from data science are removed from the equation.

You can finally focus on what you know best: the data science!

So how do you get started with Kubeflow?

Two components make scaling deep learning possible in Kubeflow: Kubernetes and Tensorflow.

What's Kubernetes?

Kubernetes is a smart scheduler for your infrastructure.

It can distribute tasks across multiple nodes and monitor them for completion.

It can also do other things such as automatically adding more nodes as you run out of resources.

It's good practice to start your cluster with a single compute resource and let Kubernetes add more as jobs are submitted.

But what's so special about Kubernetes?

Kubernetes is cloud provider independent.

You can install Kubernetes on Amazon Web Services, Azure or your private cloud.

If you decide to move to another cloud, you can install Kubernetes and continue using the same resources and code.

But Kubernetes doesn't know how to do machine learning.

It can only speak infrastructure.

So how can you train model with it?

Training models with Tensorflow

Kubeflow uses Tensorflow to break down your models into smaller tasks that can be parallelised using Kubernetes.

Tensorflow is one of the frameworks you could use to scale your models. The team behind Kubeflow has the plan to add more frameworks such a PyTorch, MXNet and Chainer.

Please note that you should write your model in Tensorflow (or Keras) to begin with.

Kubeflow cannot parallelise your work if you use vanilla Python of scikit.

You should also pay attention to the Tensorflow API that you use.

TensorFlow exposes three different kinds of APIs: low, mid and high-level APIs.

High-Level TensorFlow APIs Estimators Mid-Level TensorFlow APIs Metrics Layers Datasets Low-Level TensorFlow APIs Python C++ C++ C++ Java C++ Go TensorFlow Kernel Tensorflow Distributed Execution Engine

When you use the high-level APIs such as the Estimators API you benefit from:

  • a more accessible and human-friendly API
  • Estimators take care of abstracting the underlying resources from you

The latter benefit is precisely what you need to test and train your model locally with a subset of the data and then move to the cloud with a much larger dataset and a more powerful cluster.

If you use the estimator API, you don't need to rewrite the code to run in the cloud.

The second detail you should pay attention to is the method you use to train your model.

The only method that works locally and in distributed TensorFlow is tf.estimator.train_and_evaluate from the Estimators API.

Tensorflow offers the same method as two separate commands: train and evaluate.

But they only work locally and not when you deploy in the cloud.

If you pay attention to the previous steps, you have a model that works in a Jupyter notebook, locally and the cloud.

And in combination with Kubernetes, you scale it to thousands of computer.

Once the model is trained at scale, what do you do with it?

Creating APIs from your models with Tensorflow serving

You can expose your model as an API without writing any code.

As long as you use Tensorflow to write your model and export the results, you can leverage Tensorflow serving to read the data and start serving predictions.

Tensorflow serving can be deployed as a web server in Kubernetes thanks to Kubeflow.

Which means that you can use the same Kubernetes primitives to scale your API server to handle more calls.

You don't have to look after the infrastructure yourself anymore.

The setup is great if you need to train and serve one model.

But what if you wish to tune the parameters and train several models at the same time and offer a versioned API for each of them?

Rinse and repeat: hyperparameter sweeping

The jobs that are submitted to Kubernetes can be parametrised.

So what if you could parametrise the hyperparameters and produce multiple results for the same model?

You could explore more variations of your model.

And you'll have a lot of data which you could use to serve an API.

But how do you make sense of it?

You could visualise the detail of your distributed TensorFlow jobs with Tensorboard — a dashboard that makes it easier to understand, debug and optimise TensorFlow programs.

Tensorboard is particularly useful when you wish to monitor and optimise your models for parameter sweeping.

null

With an eye glance, you can tell which parameters are performing better and where you should focus your attention.

Summary

Kubeflow, Kubernetes and Tensorflow are an excellent toolbox to design, train and scale machine learning models.

With Kubernetes you can offload most of the infrastructure-heavy operations such as scaling and provisioning.

Tensorflow has an efficient collection of tools that are neatly designed to work in concert.

And finally, Kubeflow is the glue that makes it possible to leverage the best of both worlds.

If you are interested in a step-by-step tutorial on how to configure Kubeflow in your cluster, you should stay tuned!

In the next episode, you will dive into the details of configuring the cluster and running deep learning jobs at scale.

If you want to be notified when the article is ready, you should subscribe to the mailing list!

Thanks to Leonardo De Marchi (lead instructor @ideai.io) for his feedback!

That's all folks!

If you enjoyed this article, you might find the following articles interesting:

Learn how to tame Kubernetes and deploy applications at scale with an hands-on, self-paced course designed for visual learners.

The Learnk8s Academy is an excellent resource if you wish to learn how to:

  • Tick

    Master Kubernetes networking and handle the busiest traffic websites without breaking a sweat

  • Tick

    Scale your jobs to thousands of servers and reduce the waiting time from days to minutes

  • Tick

    Design reliable clusters and enjoy peace of mind knowing that your apps are highly available in a multi-cloud setup

  • Tick

    Optimise server density with containers and Kubernetes and save a ton of cash on your cloud bill by using only the resources you need

  • Tick

    Supercharge your delivery pipeline and deploy application around the clock

P.S. Don't miss the next experiment, insight, or discount: subscribe to the mailing list!