Infrastructure for an event that happens once

01Where it started

Three thousand people arriving in the same minute, as the vote was announced. Not a curve but a jump, and an architecture sized for the average fails exactly there.

02What could not move

• 99.9% availability. Budget was not the constraint.
• The date does not move: the peak happens once.
• If something fails there is no time to restore: it has to carry on alone.
• Everything is switched off when the event ends.

03The decision

Managed cloud: that availability is bought ready-made and there was no time to build it. Containers behind a load balancer, with a firewall filtering SQL injection, cross-site scripting and request bursts. Redis in front of the database, since a vote is almost all reads, and a standby in another zone ready to take over as primary with nobody stepping in. Both zones in Ireland, the lowest latency for Europe. All of it as code, to stand it up, rehearse it and take it down without relying on anyone's memory.

05What was not in the plan

The web server logs went to the same network volume, the only way to get them out of the container, and started failing there. They were moved off it. The rule that came out of it: a shared volume carries only what is shared.

06The result

The peak held. Thirty days later everything was switched off, as planned: no account left with resources running that nobody dares touch. Here managed cloud was the right answer; elsewhere, the same question gives the opposite one.