One of the most daunting tasks of getting going on a personal data science project can be just getting started with the environment. What I generally want is instant access to the latest technology stack (hopefully at low cost), allowing me to get to work quickly while also exploring the most recent tools and sharpening my skills. While there are some quick-start environments around, most seem to suffer from significant vendor lock-in, limiting the actual usefulness of the environments.

Overview

I recently parted ways with my MacBook Pro laptop and faced a decision - buy another, or try something new? It would be convenient to just go with what I’m comfortable with, but I’ve been experimenting with lots of interesting technology lately and want to see if I can do better.

If I were to purchase a new personal computer for data science projects I’d be most tempted by a 15” MacBook Pro. My main motivations for Mac are how close it is to Linux, while still being friendly for applications like Adobe Lightroom/Photoshop and Finale/Sibelius music editing software (hobbiest photographer with composer wife). I used Linux on my desktop/laptop for ~10 years, but finally just got sick of compiling video drivers or plugging my laptop into a projector and having to edit X11 config files to make it work. While I’ve been extremely happy with the Macs, the downsides for modern data science are:

  • Locked into the hardware - Once I’ve paid $4k for a computer I can be hesitant to use anything else. While that $4k computer may be tremendous overkill for many day-to-day tasks, it can be underpowered for anything of scale.

  • Unreproducable - The traditional laptop model is essentially a that of a pet - I carefully curate what software gets installed, and over time I end up with a snowflake - different from every other computer, completly unreproducable, and if I lose it I’m screwed. I can try to use conda/virtual environments to let others reproduce my code, but that can often only go so far.

  • Lack of community - Ideally I’d love to be able to build and share with others, as people create great new things, being able to quickly deploy and access those resources, and share back to the community. This is extremely difficult when everyone only has access to completely different environments. With all the great tools now available (TensorFlow, PyTorch, SparkNLP, etc) this is increasingly improtant to me.

  • Lack of continuity to scale - Getting a pet project working on a laptop is one thing, getting it running at scale and being served with high availability is often a completely separate step, but it doesn’t have to be! I’m convinced that it’s now possible to have a personal environment that’s easy and cheap to build and that can allow both exploratory model development and potential large-scale training and high-availability deployment as needed.

In this series of posts I’ll describe my decisions in building a better personal data science environment - one that is cost-similar to that new laptop, but that allows large-scale model building and facilites reliable algorithm deployment and pipelining.

Design Goals

Some of the user stories that I have in mind for this project are:

  • As a data scientist I need to be able to access the latest software quickly. There can be a wide range of such software, including standard relational databases (Postgres, MariaDB), next-generation databases (Mongo, Elastic Search, OrientDB, Cassandra), scalable compute infrastructure (Spark, Dask), and more.

  • As a data scientist I need to be able to spend the minimal time necessary to prepare and manage my environemnt, so that I can spend the majority of my time performing data science rather than systems administration tasks.

  • As a user I need to be able to reliably and simply access the systems - it should not require more than 1-2 steps to be up and running after the infrastructure is created.

  • As a researcher engaging in personal projects (as opposed to commercial) the infrastructure needs to be inexpensive - defined here as cost-similar to the Mac laptop ($4000) over a 2 year time frame.

  • As a community member, everything I build and use should be open source and openly licensed so that I can share my learnings and benefit from other’s comments.

  • As a data scientist I should be able to have quick access to high-end hardware as needed (e.g. GPU), but I shouldn’t pay for resources that aren’t in use.

  • As a data scientist I should be able to cost-effectively scale up as needed - even for personal projects data sets can easily get into terabyte range, and it must be possible to scale storage and compute resources to accomodate.

  • As a user I want to be free from vendor tie-in - able to move between companies as prices or capabilities change or my needs vary.

  • As a community member it should be possible to invite others into my environment - easy to setup new accounts and collaborate with fellow researchers.


Comments

comments powered by Disqus