What they're testing
Whether you think about ROLLING BACK, and whether you notice the database is the real constraint.
The short answer~30 seconds
Rolling replaces instances gradually — cheapest, but both versions serve traffic during the roll, and rolling back takes as long as rolling out. Blue-green stands up a whole second environment and flips traffic at once — near-instant rollback, at the cost of double the infrastructure while it lasts. Canary sends 1–5% of users to the new version, watches the metrics, then widens — the safest, and it requires observability good enough to make the call.
6 pods running v1. The goal: reach v2 without anyone seeing a 502.
The long answer
The deciding constraint is nearly always the database, not the deployment mechanism. If your migration isn't backward compatible, blue-green loses its meaning: you can't roll back because the old environment can't read the new schema. That's why teams deploying several times a day separate migrations from deploys and follow expand–contract — without that discipline every deployment strategy is slideware.
A canary is only worth having if you know what would stop it. Error rate and latency are the minimum, but they catch technical failures rather than business ones — a change that drops conversion 20% can have a flawless error rate. So a good canary is tied to one or two business metrics, with automatic abort thresholds rather than someone watching a dashboard.
One technique often conflated with these: feature flags. They separate DEPLOYING from RELEASING — the code ships to production switched off, then is enabled per cohort. Rolling back is flipping a flag in seconds, no rebuild. For most teams, a plain rolling deploy plus feature flags gives the benefit of a canary without the traffic-routing machinery.
What they'll ask next
?What's an acceptable rollback time?
Shorter than the outage you can tolerate. If the SLO allows 43 minutes a month and a rollback takes 30, one bad deploy spends nearly the whole budget. Measure rollback time as its own metric rather than assuming it's fast.
These lose points
- No rollback plan for the migration. That's where every deployment strategy breaks.
- A canary with no abort criteria. Then it's just a slower rolling deploy.