Five-year check-in for the Startup Founder’s Guide to Analytics

[dbeatty10]

Tristan wrote the The Startup Founder’s Guide to Analytics five years ago, and the modern data stack (MDS) has continued to take shape since then.

I'm wondering if any of the recommendations for each stage has changed since then?

[krevitt]

Such a great topic. My hunch is that earlier-stage startups have become much more sophisticated about how they do this stuff, because a single person doing data work just has a lot higher leverage / better tooling than they did when this was written.

Thinking primarily about stuff like data activation / reverse ETL, and how data becoming operationalized also forces you to do stuff like adopt data testing / version control earlier than you would when all of your earlier data analysis was done in spreadsheets.

[nicholasyager]

My hunch is that earlier-stage startups have become much more sophisticated about how they do this stuff, because a single person doing data work just has a lot higher leverage / better tooling than they did when this was written.

I strongly agree with this sentence, @krevitt! When I was starting up an analytics discipline at Torch in 2019, the bar to entry for setting up a warehouse, event tracking, testing, etc. was low enough that a single data person could set up large portions of the MDS quickly and easily.

[nicholasyager]

As someone who used this as a resource not too long ago to go through all of the listed stages, there are some sections I'd love to tweak:

• Implement data testing and documentation as soon as possible. This is important for maintaining trust between the data team and data consumers even in the very early stages. This will slow you down a little, but it's faster than rebuilding trust.
• Once there are two or more people working on the data platform (e.g. Early Stage or Mid-Stage), use pull requests and code review. This helped our team leverage CI/CD during development, which cut our defect rate significantly.
• The "Implement a solid process for SQL-based modeling" is a good candidate for Early Stage. Having that been said, the primary reason that I'd move it earlier is so that practitioners have a chance to iterate before the company gets too big.
• Mid-Stage is a great time for data activation. It's usually around this size that more opportunities for leveraging data open up (e.g. enriching a CRM or marketing tool with product data). As long as you have only one or two targets in mind, the pricing can still be reasonable.
• I'd add in some "What not to do" warnings around setting up tools that you don't need yet. For instance: metadata catalogs, augmented analytics tools (Sisu/Thoughtspot), and most entitlement tools (Immuta) will not be useful until very late growth stages, if not even later.

[nicholasyager]

Oh, and I would add using dbt Cloud to the list for Early Stage. It's a little self-serving now, but at the time it did remove the need to setup any dataflow automation tools, which save a lot of time.