|By Marten Terpstra||
|July 14, 2014 01:45 PM EDT||
Last week Greg Ferro (@etherealmind) wrote this article about his experience with scripting as a method for network automation, with the ultimate conclusion that scripting does not scale.
Early in my career I managed a small network that grew to be a IP over X.25 hub of Europe for a few years providing many countries with their first Internet connectivity. Scripts were everywhere, small ones to grab stats and create pretty graphs, others that continuously checked the status of links and would send emails when things went wrong.
While it is hard to argue with Greg’s complaints per se, I believe the key point is missing. And it has nothing to do with scripting. In a reply, Ivan’s last comment touches on the real issue.
We have been scripting our networks against CLIs forever and I will bet you most folks will consider it successful, even though it may be a pain. The article lists the pains, but not the reasons why. As a network industry, we have never ever considered the interaction with our network devices an API. Not in the true software engineering sense of an API.
There are many extremely complex clustered applications that rely entirely on exchanging information through APIs that are well documented, well versioned, well abstracted and properly promoted or deprecated. Creating and maintaining APIs is a real software engineering effort, a skill that requires true architecture, engineering and discipline. And we have not given our users anything close to it.
If we (that collective network industry) had truly considered our CLI an API, we would (and should) have been pushed aside a long time ago. The CLI is and always has been a simple interface for a human to tell a device what to do. It was not designed to be automated. It is not structured enough to be automated. Even large vendors have multiple flavors that are all industry standard, but all slightly different. And nowhere would you find a formal, full and complete dictionary of that CLI with all inputs, outputs, versions and options. The closest the network industry has had to a true API is SNMP, and that is indeed a very sad statement.
I think we have mentioned before that the networking industry is a bit slow to get to modern software engineering methods and practices, but the tide is changing. And whether you want to call it SDN or something else, the sheer volume and complexity of interaction with the network is pushing us to provide truly automated access to our devices and our networks.
And creating and maintaining APIs is far more than the technology used to access them. It does not matter whether its XML, JSON, REST, NETCONF or anything else. Those are definitions of how information is carried to and from the device and network. I can build a wonderful REST API that takes a CLI command as an argument and spits me back the output from that CLI command in some format. I am sure that sounds familiar to some, but this is not an API. Not in a truly meaningful way that would elevate our automation abilities.
Designing and implementing APIs is not trivial. Believe me, as an entirely API driven solution, we spend a tremendous amount of time discussing our APIs and abstractions to make sure they find that find balance between granularity, functionality, abstraction, scaling and a few other relevant qualifiers. But the key is that they are part of any feature design from day one, they are part of the overarching architecture, not bolted on at the end. Our APIs are not perfect, there is no such thing, but they are modeled after the workflow of you the user doing the work required to keep the network running and thriving.
So when you need to configure MLAG on a set of Plexxi switches, we do not have a series of API calls to bundle ports together on a switch, give them a unique ID, then tie the switches together as an MLAG pair that shares that unique ID. Oh, and create an MLAG control channel between them, and make each of the switch local LAGs have the same set of VLANs on them. Our API will simply take a list of port objects from any amount of switches in a Plexxi network and turn them into an MLAG. An then you can simply take that entire entity and stick a VLAN on top, we will make sure the participating switches get the pieces they need. That is abstraction, that is workflow encapsulation, that is what APIs are supposed to give you. That is how simple LAG is supposed to be.
We have a long way to go as an industry to get to full APIs the way real software folks think about APIs. The CLI is not it. Scripting against a CLI (or a CLI hidden behind a layer of official sounding API terms) is a useful tool, but one that should be mostly retired to get to true programmable networks that are controlled by real controller (in the broadest definition of the word) using real APIs. Automation is not scripting.
[Today's fun fact: to make sure you do not think I am anti scripting, I once wrote a large chunk of a 10,000 line Perl4 system. It functioned very nicely for years as the RIPE database for IP address allocations back in the mid 90s. Thankfully it has since been tackled by real software engineers.]
The post Scripting is automation, but automation is not scripting appeared first on Plexxi.