Q: Should we put our system on Kubernetes? Everyone in our industry uses it.
A: Look at three numbers first, not at your peers. Daily active users, how many machines, how many deploys per month?
The client checked: around 3,000 DAU, 2 ECS instances, 2 deploys a month, and one ops person.
That Time, the Answer Was No
Q: Why not?
A: Kubernetes mainly solves problems like "many services, many machines, frequent scaling, self-healing." With 2 machines, 2 deploys, and one ops person, those problems don't exist yet. Adopting it adds layers you'd have to learn instead: Ingress, Services, Pod scheduling, persistent volumes, and piles of YAML. One person manages it today; after K8s it might take two.
Q: But what if we need to scale later?
A: Then build the foundation that "scaling later" depends on most. That time we did three things: moved deployment scripts into Jenkins with versioned artifacts and one-click rollback; added MySQL replication and nightly backups; moved config files from personal machines into a repository. All three pay off whether or not you ever adopt K8s, and none of them requires a new platform.
Q: So when should we adopt it?
A: A few signals. Services split into 6+ with independent releases; more than 5 machines with frequent ad-hoc additions; clear traffic peaks and troughs needing auto-scaling; or someone starting to spend half their time just on "deployment." Hit two or more of these, then we'll talk — that's when K8s saves real labor.
Q: What happened six months later?
A: Six months later, the client raised it themselves. They'd picked up another channel, DAU had tripled, machines grew to 6, releases moved to twice a week, and ops got more hands. All three numbers had changed, so we did it together — starting with stateless services only, leaving the database and Redis outside, not cramming everything in at once.
In Plain Terms
The trap in technology choices is "everyone else uses it." Everyone else has their own DAU. Put the three numbers on the table and the answer usually arrives on its own — no argument needed.
That conversation took under 40 minutes; 20 of them were spent looking up the three numbers.
