Every year someone declares SNMP dead, and every year it keeps quietly reporting the interface counters, CPU load, and chassis temperatures of approximately the entire installed base of network equipment on Earth.

Streaming telemetry — gNMI, model-driven push — is genuinely better where it exists. It is also, a decade after introduction, still unevenly supported outside recent-generation gear from a handful of vendors. Meanwhile the branch router, the campus PoE switch, the UPS, and the CRAC unit in your wiring closet all speak fluent SNMP today.

What SNMP is still uniquely good at

Universality. One poller, one protocol, every vendor and every device class — including the environmental gear (temperature, humidity, power draw) that never gets streaming support and that, in our experience, explains a remarkable share of branch-office mysteries. A wiring closet running hot is invisible to every other protocol in your stack.

Cheap history. Polling interface octets, errors, and discards every 60 seconds across a thousand interfaces is a rounding error for both the devices and the collector, and it builds exactly the per-interface baseline that anomaly detection needs.

Truth at the counter level. Flow data describes conversations; SNMP counters describe the physical interface — CRC errors, drops, and discards that no flow record will ever show you.

Polling it well

The difference between "SNMP is noisy garbage" and "SNMP quietly catches failures early" is configuration:

  • Use v3 where you can, v2c where you must — and if you're stuck with v2c, at least scope the community string read-only and restrict it by source address.
  • Poll counters at 60s, health at 30–60s, inventory daily. Interface octets need frequency; serial numbers do not.
  • Collect the boring OIDs. Errors, discards, CPU, memory, temperature, fan status, PSU status. Nobody regrets having temperature history; everybody regrets not having it the day the closet AC failed.
  • Baseline, don't threshold. A branch access switch and a core router have wildly different "normal." Per-device baselines — learned, not configured — are what make SNMP-derived alerts trustworthy.

SNMP isn't the future of telemetry. It's the present of telemetry, and it will remain the present for as long as network hardware outlives hype cycles — which is to say, a very long time. Poll it well.