Engineering

The constraint

Most of what a new service needs was already built three times, by three teams that could not reuse each other's version.

What goes wrong

Three pipelines, one company

Each team built its own deploy path at the moment it needed one. None can adopt another's, because the differences that matter are the ones nobody wrote down.

Reuse fails on the last tenth

A shared library covers most cases, and every team eventually meets the case it does not cover. Forking is faster than changing the shared version, so a year later there are four versions and one maintainer.

Rollback is a document

The procedure exists, is a page long, and is exercised once a year under pressure by whoever is awake. A recovery path that is not routine is a recovery path you are testing during the incident.

What changes

• What stops

  A new service starts by copying the last service's pipeline and editing the names.

  What replaces it

  A new service starts at the first commit of the product.

• What stops

  Rollback is a runbook somebody follows at two in the morning.

  What replaces it

  The previous version stays addressable, so a rollback is a pointer change.

• What stops

  Capacity is sized ahead of a launch and left at that size afterwards.

  What replaces it

  Instances follow traffic, down to none between requests, and idle capacity is not billed.

• What stops

  Base images, runtimes and certificates are somebody's recurring ticket.

  What replaces it

  They are maintained underneath the application, with no host to log into.

Questions we get asked

We already have a platform team. What is left for you?

If their backlog is full of product requests, nothing here is urgent and you should ignore this page. The case this fits is the team that maintains a platform as a side effect of maintaining an application, where base images and certificate rotation are eating the quarter nobody planned to spend.

How locked in are we?

Every platform holds something, and being told otherwise is a reason to be careful. Here it is the deploy configuration rather than the application, which stays ordinary code. The larger cost of leaving is rebuilding what you stopped maintaining, so price that at the start rather than discovering it at the end.

What about workloads that do not fit this shape?

A process that must hold state in memory between requests, or hold a machine for hours at a time, is a poor fit and we will say so rather than talk you into it. That work usually wants dedicated capacity, and it is a smaller share of most estates than the teams running it expect.

What fits