Democratization of data in Uber

Hello!

Under Halloween I attended a conference in Budapest ( Data Crunch ) and listened to a number of interesting reports there. One of them was from Uber, who talked about the approaches on which they organized their data management platform. This report was not so much technical as managerial and grocery.

Uber is widely used data that collects as a result of interaction with passengers and drivers. They calculate the cost of the trip, evaluate the flow of people, change the price algorithms, give drivers recommendations on how to make more money and all this based on the data collected. In such a company, all work with data cannot be concentrated in the hands of a group of analysts and DS, since otherwise, you will have to hire them too much, and besides, they are not always immersed in the business context.

From the very beginning, the company took the path of building a data management platform that would allow the use of sufficiently advanced analytical tools to a wide range of users. They identified 4 main groups:

1. Regular users - they know basic SQL, basically they just need tables with data, dashboards)
2. Regional managers - know a little more than SQL, look at the data in different sections, a great need for slice & dice
3. Data analysts - advanced SQL, build dashboards, do research, search for insights in data
4. Data Science - the maximum level of understanding of working with data, build models, conduct experiments, A / B tests, etc.
   On the sidelines, they also learned that in fact there is level 5 - top managers, who mostly use top-level reports and dashboards.

Interestingly, in Uber, people who work with data somehow must know SQL at least at the minimum level.

As examples of the product they created on the basis of their platform, they led the automation of A / B tests. The company conducts a huge number of A / B and allocate for each Data Scientist, that he organized an experiment and then gave an assessment of the tests - again, not a permissible luxury. Therefore, they would like to give ordinary users the ability to interpret and use A / B correctly and without errors, while not loading Data Scientist.

Their construction of this product began with deep work with Data Scientist, because if these guys are not sure that everything is considered to be true, then no Data product will come out. In fact, they began to automate the launch and evaluation of A / B tests, giving Data Scientist a tool that makes their lives easier. After that, they increased the interface to this tool, which would simply show the test results (what they did, what difference, whether the difference is significant). At the same time, they hid under the hood the maximum number of nuances inherent in the A / B tests, so that the user does not need to dive deep into mathematics and statistics.

Interestingly, most people with whom I spoke during coffee breaks said that they didn’t have any A / B tests in their practice, that they use largely qualitative research and intuition when making decisions. So as elsewhere, there is no time to think, you need to saw!