Operations

How we plan maintenance windows without surprises

A maintenance window is a promise with a clock attached. We are going to interrupt something on purpose, at a time we chose, and we are asking customers to trust that the interruption will be short. Here is how we try to deserve that.

The rule

Customer-impacting work happens Tuesday to Thursday, between 01:00 and 04:00 UTC. We publish the window at least five working days ahead on the status page and by email to affected customers. Fridays, weekends and the last two weeks of December are off limits unless something is actively on fire.

If a change cannot fit inside a single window, it gets split. A three-hour migration is not a three-hour migration; it is a series of smaller changes that each have a way back.

How a change gets approved

  1. A written plan: what changes, why, who is affected, and exactly how to roll back.
  2. Review by an engineer who is not the author. “Looks fine” is not a review.
  3. A dry run in staging or a single non-critical site wherever possible.
  4. A configuration snapshot and a copy of the previous firmware image.
  5. Two engineers in the window: one changing, one watching graphs and the clock.
  6. A short review within two working days, even when nothing went wrong.

The numbers

Last quarter we ran 47 planned changes. Two were rolled back and none caused a customer-visible incident. The first rollback was a switch firmware update that broke LLDP discovery; we reverted in 11 minutes. The second was a storage tunable that made write latency worse instead of better. Both were caught inside the window because the engineer was watching the latency graph, not waiting for the clock to run out.

Those numbers are not a victory lap. They are the reason the process exists: two wrong changes became two short rollbacks instead of two incidents.

When we break the rule

Security patches do not wait for a Tuesday. An emergency change still gets a named owner, a rollback plan and a status entry. If it needs to happen at an unsociable hour, we wake people up and pay them for it. That is what the rota is for.

We have broken our own Friday rule once in 2026: a critical hypervisor patch that could not wait until Monday. The status entry went up before the change, not after, and the postmortem noted that the timing was correct even though the rule was inconvenient.

What you can do

  • Check the status page or subscribe by emailing status@atlantika-company.uk.
  • Tell support if your own change freeze overlaps ours; we can move a non-urgent window.
  • If you need us on the bridge during a customer-side migration, ask. We would rather be bored together than surprised separately.
← All notes Written by Daniel Okafor, Head of Customer Operations