Scaling carrier-grade voice from 10K to 10M minutes a month

Voice scales differently from HTTP. The traffic is bursty, the SLAs are unforgiving, and a bad deployment doesn't show up as a 500 — it shows up as silence.

Pattern 1: Route by geography, not by load balancer

DNS round-robin and L4 balancing are great until you're crossing region boundaries on every leg of a call. Pin sessions to the closest POP, fail over only when latency exceeds your tolerance.

Pattern 2: Separate signaling from media

SIP traffic is tiny. RTP is not. The teams that scale cleanly run them on separate fleets so a media spike doesn't crater registration.

Pattern 3: Multiple termination partners, one routing brain

Single-source termination is fine until the day it isn't. A routing layer that can shift traffic between providers in seconds turns a partner outage from an incident into a footnote.

The thing nobody mentions

Scaling voice is mostly about scaling your ops cadence: paging, post-mortems, capacity reviews. The boxes are easy. The discipline is the hard part.