Chef arista devops days a'dam 2015

Editor's Notes

1. So let’s start here, let’s say we have a simple leaf/spine design. In order to provision my network I’lI go and connect my console to each and every switch and do all the work manualy I believe it’s clear to everybody why this is not a good idea, some of us still do this for one reason or the other As you scale your infrastructure the cost will grow exponentially. You need more people. Your downtime will be longer because it’s harder to troubleshoot So what easy way we have to do better than this
2. One very easy way to do a lot better is to move the configuration from the devices them selves to a provisioning server So we can use open standards and protocols to achive this. You can creat a DHCP server to create a bi-derectional connection between the switches and the server So whenever the switch boot-up it will do a DHCP recquest and get an IP an the location for the server from the DHCP server. So once we have this bi-directional communication between the server and the switch we are able to do a couple of things We can send the configuration from the server to the switch, so we don’t need any manual intervention any more We can now also push the configuration from the switch to the server, if we’ve made any configuration changes So this way we can backup all configuration and find out who made any configuration changes and when Another advance of building this very simple provisioning server, is that it allow us to do ZTR (zero toucn replacement) What if one of my nodes breaks, how do I replace it In this setup it’s very easy,we aready have the latest configuration of the switch on the server. This is great, and it’s better than doing everything manualy. However it still has some drawback: Single configuration file per switch, and in this case we have a system ID per config. I cannot upgrade or downgrade any of my newly installed switches I cannot install eny extention or application So lets see how we can solve that:
3. So the next step is pretty simple as well. First of all let’s take the configuration and break this into little template. And I believe a lot of people already do this right So for example my LLDP configuration would be the same for every switch, so what’s the point in copy and pasting this 10x Instead I define it once and I test it once and then re-use it in my network Before I move on I need to introduct a couple of new acronyms So we can break the configuration from one file into multiple actions and each action can install a certain template to the devices itself Once we have this concept we can take multiple templates or actions and map them together in what we call a definition file And now we can map the SYS-ID to a defination file, this gives u a lot more flexibility and have less manual intervention JSON can be considered a subset of YAML And on top of this we have the ability to add more actions So far I’ve talked about one action which add a template to the switch But we can have another action with add more complicated logic like upgrading the software of the switch or add an extention to the switch Or before we finish the configuation let’s first check if the field engineers connected the cables to the right ports and act on this. We can do this by adding some custom actions which can enforse all of the cabling checks. Or if you’ve done a ZTR let’s see if we can ping the default gateway and if it doesn’t work what should I do So this is a bit more complex than previous examples so I think its worth talking about how we can monitor the proccess right. How we can make sure that evertthing went fine and what is going to happen if one of the actios would faill And again for this we can leverage opensource tools and protocols. . And the way we do it at Arista is we use XMPP and remote syslog. What this means is that if our customers deploys something like that they use a ZTP server and then the filed engineers have an XMPP app on there phoe and can join a particular XMPP group. Now they get real time notifications if they connect a switch to the network and they possible get tips and tricks if something went wrong. i.e. cable in a wrong port. So what’s the problem here and how can we improve You still need to make the mapping between the system-ID to a definition. In a modern Datacenter there are a lot of devices which has the same functionality based on there location We can have multiple pods which have the same role The conly difference is the hostname and the MGTM IP for example. So let’s see if we can take this one step further?
4. And I think we can do that by defining classes of switches based on there LLDP location in the network. So define a class of switches and map this to a definition file This means that whenever we are adding a new node we don’t have to make any new configuration changes on the server and we don’t need to touch the switch Just because all the switches in that particalur clas follew the same role. Okay so you might think what to do with the host names, one way to do it is to automatic generate them. We can automatic generate them based on the connection to the management network I,e, How about the MGMT IP or the IP subnets between the spine and leaf. Well in this case you don’t need a specific IP for a particular switch. But you need an IP address from subnet A or B. So instead of defining a template for each induvidual switch we can now define a generic template and we configure what we call a pool of resources. Let create a subnet resource pooland than we say to the server. You know what if this switch comes up allocate any IP address from that pool and mark it as allocated So we started in our first example with doing everything manualy and now we are in a stage where we have to define what we are building but once we’ve done this everything is automatic with very little human intervention. So let’s see how ZTR works. Let’s go to the previous example where there was a mapping between sysID and a definition, Now if you replace a switch the switch will send its LLDP neighbor informatio to the server and ask the server if he sees somebody who looks either excactly like me or very close