How to make your data team efficient for times of crisis
Times have changed and caught most of us unprepared. It is always a part of Bolt’s culture to move quickly and adapt — and the crisis situation that is unfolding due to a pandemic definitely requires significant adaptation. This is a look from inside Bolt’s data team — data analysts, data engineers, data scientists — as we share our experience and advice for times of crisis with all the similar teams out there.
In times of turmoil, the market situation for a lot of businesses changes overnight. Most of the resources are thrown into surviving and, for some, even on seizing new opportunities.
Data teams definitely have a role to play in this. So, what to focus on and what to change?
A crisis develops slowly and then suddenly a single event can erode confidence and cause a real crash. You probably had a roadmap for this quarter, and year. A slow and evolutionary one, making things a bit better than before and taking a bit of risk for larger wins. Now the roadmap becomes less relevant; it’s surely an option to continue as nothing happened, but more value could be found in re-prioritizing and figuring out what makes the most sense now.
Is there any technical debt, like a not-so-optimal algorithm, that wastes precious resources? Does any other team need your help immediately? Do you need to make your company’s business decisions smarter and more informed? Are you set for rapid recovery?
Example: Last summer we launched Bolt Food in Tallinn, Estonia. Today, the service is available in 11 countries and we’ve launched in 10 new cities within the last month alone. We have also launched grocery delivery on Bolt Food in a matter of days. These are among the fastest-growing products we have now, as people are avoiding unnecessary contact and ordering things to be delivered to their home. Launches and initial scaling are usually very operational with optimization not a top priority. Even though the data team had well-defined plans for Q2, we decided to look at whether we can accelerate Bolt Food’s growth and improve efficiency — two things that matter in a crisis more than ever.
Also, a data team is usually the most equipped team in the company to provide robust situation awareness for everyone. If some insights were missing or ignored, the data team can dig them up — possibly incorporating outside data sources, such as how the crisis is progressing and impacting the outside world.
This one is more related to internal functioning of a data team, but applies to all the company’s teams as well.
Review everything you have been doing so far. You definitely have technical debt in the form of sub-optimal SQL queries, ETLs, unused cloud instances and endpoints, which unnecessarily add to costs. In good times this can be sacrificed for the speed of movement, but in bad times the speed of movement is measured in how quickly you can reduce your burn.
Example: In our team, the crisis has accelerated the push to migrate immutable, rarely queried data from our main data warehouse in Amazon Redshift to a cheaper storage service in S3. After two weeks of working on a common goal, the team managed to reduce monthly Data Platform costs by 50%.
Another possible situation: one sub-team consumes resources, another allocates. The first team doesn’t see the costs and doesn’t see a problem in consuming ever more of them. The allocating team sees steadily increasing costs, but has no visibility of the nature of the work of the consuming team — and assumes it’s normal. Get behind a single table and find such points of inefficiency. It is critical in crisis, but such reviews, when done regularly, will also pay off in good times.
Every crisis is an anomaly when people behave in unexpected ways. You probably had all the relationships between your important business metrics figured out. But what happens if your demand drops massively? If you are flooded with a wave of refund requests? If your software starts to be used in an unusual way?
Take a look around and understand what kind of monitoring you are missing. There are probably things you have taken for granted and didn’t develop monitoring for. Now these things may have broken down and may drive even more churn that you definitely don’t need.
This crisis too shall pass. When things start to get better, it will be another anomaly, on the same scale as when things took a turn for the worse. And that time it will be even more decisive to move quickly and accurately, because this would be the kind of situation you can control, and on which your success down the road will depend. Start preparing for resurgence, even though there might be no good news at the current moment. Resilience matters.
Example: We have dynamic pricing, which balances supply and demand by applying a price multiplier. During the current crisis, the demand and supply situation changed dramatically, so the models that were previously trained may have lost touch with reality. Thanks to extensive monitoring and alerting, we can spot serious issues as they arise.
Do simple solutions that just work — even though the tech behind them may be boring.
More established companies likely already have this point figured out, but it still needs to be stressed. Also, if you are an aspiring or beginning data wizard, you should seriously think about it.
If you want to develop some model and think deep learning, go simpler. A linear regression or a simple heuristic may produce most of the results in a fraction of the time needed to do more complex research, requiring much less compute and data. The bottom line is what matters — and you will be rewarded for that.
You may wonder now what to learn or what to experiment with. Some obscure technological niche may seem attractive. Unfortunately, for most niches the demand is low and time-to-market is high. In times of crisis, even more so. Better to put your resource into understanding how to do basic market and customer research. It will also provide you with a great technical and analytical foundation. And it is in the most demand — both in a crisis and in prosperous times.
Example: Let’s sum up the above: most data on customer behavior and the state of business is stored in tabular form in a relational database. At Bolt, we use machine learning to provide the best possible service to our riders and drivers. What kind of models do we use for training on such data? Decision tree models, of course — more specifically, gradient-boosted trees. It is much simpler and faster than deep learning approaches and provides the same, or better, accuracy on tabular data.
And now for the softer part.
At Bolt, we have an explicit policy to overshare rather than to be afraid to look awkward and undershare. Now that we are all in our home offices, it has become absolutely crucial and is working to our benefit.
Unusual times require unusual creativity. The best solutions are born out of the whole team’s effort and rarely from individual heroism. Be nice to each other and brainstorm a lot. We’re in this together and we will come out of this together as well.