SNMP Fits Network AI Well

Reactions to SNMP in Network AI

When discussing Network AI, we sometimes get some interesting reactions to SNMP. Specifically when we reveal the amount of effort we have put into a powerful and scalable SNMP “engine”, and our claim that it actually works! For example, when we presented at Network Field Day 28 and the discussion of SNMP arose, one of the NFD28 delegates tweeted this:

The 80's wants its SNMP back!! Reaction to the statement the SNMP fits Network AI well.

Source: Keith Townsend Twitter Account, https://twitter.com/CTOAdvisor/status/1522597235065712642?s=20&t=PA5CAWefHo9dHzIGJ556dQ

The 80’s wants its SNMP back 🤣🤣🤣 That did bring a smile to my face. I do not know the first time I heard the acronym SNMP, but I do know it was a loooooong time ago!

SNMP is Ubiquitous

So why would a cutting-edge technology company even bother with SNMP, and not just focus on OpenConfig, gRPC, gNMI et al?

Recently I was listening to an interview with the founder of an observability platform, and I was reminded of why. This founder explained many of the reasons why network operations products must leverage SNMP so much:

It is in every network
Supported by deployed equipment
No two vendor data models are the same
It provides the information needed to do the job, while other approaches are still maturing

SNMP is a Good Fit for Data Science Pipelines

There is another reason why it is of specific value to our Network AI solution – it is well structured and has evolved to represent the constructs within the equipment. In short, it is a good fit for a data science pipeline.

AI/ML thrives on good data. Take a Syslog message that simply says “Failure”. It takes a good amount of work to understand what to do with a message like that, because in this instance, it is not even saying “what” has failed. With SNMP you always know WHAT has failed, and you also know something about relationships to other entities because of SNMP’s tree-like hierarchy.

Fields are clearly defined, when an event happens you know what entity the event happened to, and you also can understand something about relationships. This makes it much easier for an engineering team to make use of the information in a data science pipeline.

Is gRPC / gNMI / OpenConfig a Good Alternative?

Anyone who has spent any time in a product management position or a strategy position, knows the drill. You always want to be intercepting the next wave. You don’t want to spend a whole bunch of time and money on something that is going to be replaced. The propensity is always to focus on that next wave, and hope that wave comes along fast enough for you to catch it.

Well, as the old saying goes. In the short term, change is much slower than you expect, and in the long term much faster.

In the short term, at least, multiple network operations tool vendors are realizing the reality of equipment in networks. What is there is there, and it is going to be there for a while. Once installed, equipment is sticky, and if you want to create great customer experiences for more than greenfield deployments, and if you want to create customer experiences period, then SNMP is still the way to go today.

Augtera of course supports OpenConfig, gRPC, gNMI, and streaming telemetry will add value in the long run, but for now, a great SNMP implementation is what gets the job done.

Conclusion

It is possible for an engineering team to get some value out of almost any data, and there are approaches to unstructured data, like Augtera’s use of Natural Language Processing on text-based logs, that can accelerate that value. However, there is a tax to the entire industry when there is a significant amount of case by case, and customer by customer data learning and engineering required. The opportunity cost in terms of other innovations can be significant. At Augtera Networks, we have found SNMP to be a good fit for Network AI.

Note: Adapted from the original blog “SNMP is Dead! Long Live SNMP!” by Mark Seery on Substack