ปรับขนาดอินสแตนซ์และเรพลิกาให้เหมาะสม

Most ML serving bills come from instances sitting half-idle. Right-sizing means matching hardware and replica count to your real load, not your worst-case fear.

You cannot right-size what you have not measured. Start by watching real CPU and memory utilization over a normal traffic day.

Picking a giant instance just in case feels safe but burns cash every hour. Chronically low utilization is the clearest sign of overprovisioning.

If average CPU sits near 15 percent, your instance is far too big. Aim for a healthy target utilization with headroom for short spikes.

Vertical scaling gives one replica a bigger machine. Horizontal scaling adds more small replicas instead, which usually handles bursty traffic better.

On Kubernetes, your resource request tells the scheduler how much each replica truly needs. Set it from observed usage, not a round guess.

Too few replicas means queued requests and slow responses. Too many means idle pods you still pay for. Size replicas to your peak concurrent load.

A Horizontal Pod Autoscaler adds replicas when load rises and removes them when it falls, so you stop paying for off-peak capacity.

A minimum replica count keeps a few pods warm so traffic never hits a cold, empty service. This is your trade-off between cost and availability.

GPU instances cost many times more than CPU. Only request a GPU when your model truly needs it, and pack work tightly so it never sits idle.

Traffic patterns drift over weeks and months. Revisit instance type and replica counts on a schedule so your fleet stays a good fit.

Let us see what low utilization is telling you.

You measured utilization, chose between vertical and horizontal scaling, set requests, and added autoscaling. Your fleet now matches real demand. 💸