The deployment must produce no downtime. It does not matter whether it is performed on a testing or a production cluster. Interrupting consumers is disruptive, and that leads to loss of money and confidence in a product.

Gone are the days when users did not care if an application sometimes did not work. There are so many competitors out there that a single bad experience might lead users to another solution. With today’s scale, 0.1% of failed requests is considered disastrous.

While we might never be able to reach 100% availability, we should certainly not cause downtime ourselves and must minimize other factors that could cause downtime.