Production Ready Data Science
Becoming a data-driven organization is not an easy task. Off the top of my head, and in no particular order, these are the most frequent challenges a company faces:
- Attracting, retaining, and training the right talents;
- Collecting and making data available cross-silos;
- Modernize their tech stack or increase the complexity of the IT landscape by adding new technologies;
- Fear of the unknown i.e. many people are afraid about losing control, or their job, to data and data science;
- Lack of vision1.
As I was thinking about this list, however, I felt there was something deeper about the troubles some organizations are facing. I know in fact about companies that have made significant progress in all five points, but that are still not reaping the fruits they were expecting. When I looked at these companies more closely, they were all not putting the models they developed into production. The reasons varied, from being content with report-driven decision making (either a one-off report, or periodic reporting) to simply struggling with all the pieces of the puzzle.
Being the curious type, I set out to investigate what was making the puzzle so difficult for them. Productionizing a model involves a series of (moving) pieces:
- The data should be automatically available for the model, i.e. there should be no2 human intervention in the ETL phase;
- The model should also run automatically, follow the DRY principles, should be (battle) tested, and possibly be flexible in its sources/sinks, either by being used as a library, or by exposing accessible APIs (REST being one the most in vogue today);
- Refreshing and/or re-training the model should not impact the front-end accessing it, or, in the easiest cases, should not impact it during business hours3.
The companies struggling with becoming data driven, are failing in on or more of the above points. What they are doing is a mix of the following4:
- They copy the data around manually; this makes it basically impossible to bring a model into production because, if the data does not flow automatically, the owners of the data (wherever they might be) are probably not even aware that their data is used in a model. If the data comes in automatically, instead, they know that systems are depends on them; on top of that you should get almost immediate alerts if the ingestion fails (if not: set it up);
- They don't test the code so software engineers basically refuse to touch the code; this problem is exacerbated when the engineers have to re-write the code because of performance or other reasons; with tests unavailable, modifying the model becomes daunting;
- Related to the previous point, many data scientists are either ex data analysts that thought their job was more secure by changing their business title, or are coming from disciplines like Physics, Math, or Statistics, often with research experience. Coming myself from four years of research in Physics, I can attest that, except some unicorns, we (used to) write unreadable or very complicated code5. When PhDs end up doing software development, they quickly pick up the good practices. But in data science, exploration driven modelling can worsen the situation: you start poking around the data until it suddenly makes sense but you leave around all the steps you took, even the unused ones; coupling this to a lack of documentation, you can easily end up with thousands of lines of code that are basically acting as a scarecrow for your engineers;
- Sometimes your data scientists code in a language that doesn't nicely operate with the outside world; if you're in charge, make this stop as there are no solutions other than complete lock-in;
- Many data scientists approach the problem at hand with a Kaggle-like mentality: delivering the best model in absolute terms, no matter what the practical implications are. In reality it's not the best model that we implement, but the one that combines quality and practicality. Take the Netflix competition for example: the company made 1 million dollars available to the group that would improve its recommendation engine; the winning team found a combination of algorithms improving Netflix one by 8.43%. Netflix however never implemented it, as the method was built to handle 100 million ratings, much less that the 5 billions that Netflix had! Moreover the algorithms were not built to adapt as members added more ratings. I am quoting here but think about it for a second: the winner reported more than 2000 hours of work in order to come up with the final combination of 107 algorithms that gave them this prize. They gave Netflix the source code. And yet they did not think how the algorithms were going to be used, that is daily updated as new users were rating additional movies. 2000 hours of work!6
If you've payed attention to these points, you probably start seeing a pattern: data scientists usually suck at software quality, that is7: reliability, usability, efficiency, portability, and maintainability. Because data-driven models are implemented through software, they suffer from bad software quality just as much as your typical application.
Let me be clear: this is not an easy task! To create a (great) model you need creativity, a scientific attitude, knowledge of various modeling techniques, etc. Getting data scientists able to create these models is one of the biggest challenges for an organization. But focusing on the modeling at the cost of software quality will produce something great and admirable that ends up not being used.
This is the reason we actively hire data scientists that can code, and can do it well.
I imagine you now have the next burning question which is: what if the data scientists working at my organization are not good at it? What if someone left the company, implemented a great new method, but nobody can actually make sense of what she wrote?
This is where I pitch you our services, training and consultancy, because it's not like I write 12 hundreds words for nothing! We can train your data scientists to write code of higher quality and we can review the code they wrote. And we're very good at it and have fun while doing it! Get in touch.
I wrote this post after a lot of brainstorming with the team. A big thank you goes in particular to Gabriele for reading everything and giving me precious feedback, together his experience at some of the largest Dutch enterprises.
A lack of vision is a much broader issue than 1-4 as it can bring even the largest and most flourishing corporations to the ground (a great read about this is Good to Great). I included it nonetheless as it will cut or make budget unavailable or prevent management buy-in of data-driven products. And lack of management buy-in is even worse that lack of budget. One of our first clients installed its first Hadoop cluster on dismissed machines, built a type-ahead and recommendation engine for their web shop, and see profits surge right after they put it into production. There was nothing a budget could do had management not agreed about "letting" the model into production. ↩
Unless something breaks of course. ↩
Whatever that means for you. ↩
This is probably one of the post with the highest density of bullet points I've ever written. Apologies. ↩
I still vividly remember when a professor suggested that using
kkkas a variable name was not a very wise choice, to which I replied that I was using
kkfor something else. ↩
It is not my intention to denigrate their work. I often use the matrix factorization methods implemented in Spark to train my recommendation engines. I am merely stating that they set out to solve a problem without thinking about productionizing their work. ↩
Follow us for more of this
Fairness in Machine Learning with PyTorch
May 22, 2018
Elitist shuffle for recommendation systems
May 13, 2018
GoDataDriven open source contribution: April 2018 edition
May 06, 2018
Event Report: Blockchain Ask Me Anything
May 03, 2018
Towards fairness in ML with adversarial networks
April 27, 2018
GDPR: "May 25th is when it really begins
April 10, 2018