ETL pipelines: a practical guide

The version of this that works is simpler than the version most people imagine. This guide covers what ETL pipelines actually involves, where it usually goes wrong, and how to tell whether yours is in reasonable shape.

Numbers get quoted in meetings long after anyone remembers how they were calculated. Nothing below assumes a large team or a large budget — most of it is a decision somebody has to make and then write down.

What is actually at stake

Pipelines fail silently unless you design them not to. In practice this is a scheduling problem more than a technical one. Most teams find the first pass takes an afternoon and the maintenance takes minutes a month.

For most businesses the question is not whether this matters but how much of it is worth doing right now. That depends on what you are trying to achieve in the next few months, not on best practice in the abstract. In practice this is a scheduling problem more than a technical one.

The practical version

Make every run re-runnable without duplicating data. The reasoning matters more than the rule, because the rule has exceptions. Budget a little time for it every quarter and it never becomes a project of its own.

Most data problems are ownership problems that turned into technical ones. The version that works in practice is usually less elaborate than the version described in the guides.

Validate at the boundary, not three steps later. Where this goes wrong is almost never a lack of knowledge. The teams that stay on top of it are the ones who put it on a calendar rather than a wish list.

A working checklist

If you want a quick read on where you stand, work through this. Anything you cannot answer confidently is where to start.

• Pipelines fail silently unless you design them not to
• Make every run re-runnable without duplicating data
• Validate at the boundary, not three steps later
• Someone is named as the owner, not just assumed to be
• There is a date in the calendar to review it again
• The decision and the reasoning behind it are written down somewhere findable
• You could explain the current setup to a new hire in five minutes

Common failure modes

The most common failure is not doing this badly. It is doing it once, during a launch, and never revisiting it. Circumstances move, the setup does not, and the gap widens quietly until something breaks or somebody notices the numbers.

• It was configured during a launch and has not been touched since
• Different people in the business believe different things are true about it
• There is no way to tell whether the last change helped or hurt
• The only person who understands it has left, or is about to

Data outlives the applications built on top of it, which is why the model deserves more thought than the screens. Getting it slightly wrong is survivable. Ignoring it entirely is not.

How we approach it

On our projects this gets handled during the build rather than added afterwards, because retrofitting it costs several times more than including it. We write down what was decided and why, so the next person to touch it is not guessing.

If you are working with someone else, the questions worth asking are simple: who owns this, how will we know it is working, and what happens when it needs to change?

What to do next

Pick the single item from the checklist above that would cause the most trouble if it turned out to be wrong. Fix that one, confirm it worked, then move on. If any of that sounds like a description of your current setup, it is fixable.