text stringlengths 0 2.67k |
|---|
**Justin Garrison:** The things we're talking about here - this is 2014. We'll get into it -- you know, we're deploying Puppet at Walmart to a bunch of stores, from no configuration management; just servers set up with actual checklists. Literally word docs. |
**Autumn Nash:** Mainframe influenced configuration -- but it's just wild. For one, who would have known that Walmart was such an innovator in that space? And then it wasn't long enough -- like, we were already out of high school. We were already adults when this was happening. You know what I mean? This is not somethi... |
**Justin Garrison:** And 2012 I think is when I went to Puppet Conf. Puppet Conf in 2012, 2013. And so I was definitely -- I was Puppet certified, I did all the training... I was going to the conferences and deep into it. So I'm pretty sure that I was influenced by what Walmart was doing at the time, because I was look... |
**Autumn Nash:** You know how we were talking about the Netflix blog, and people kind of sharing knowledge? And it's a little wild that they were doing such innovation... And you don't want to end up on the news for it going wrong, but it's also wild that they weren't almost talking about their accomplishments more. No... |
**Justin Garrison:** They did in certain circles. Because that was in the Puppet world. I knew about Walmart moving to Puppet. |
**Autumn Nash:** But did you know the scale, and the -- you know what I mean? |
**Justin Garrison:** In some of the talks. Yeah, for sure. But I didn't want to -- I don't want to spoil the interview here, but the fact that that migration is still going on 10 years later... |
**Autumn Nash:** It's wild. |
**Justin Garrison:** It's still happening. Because the accomplishments of 10 years ago no longer look like accomplishments today. And so now if you're not going to AI, and doing the cloud and doing Kubernetes, you're like "Oh, forget -- get out of here." There's so much legacy. |
**Autumn Nash:** I mean, I don't know... But look at how you turned around, and a lot of people are going from the cloud to on prem. So is it, though? You know what I mean? And yeah, everyone's doing AI... But until when? Until they rip it all out of things all over again. So I feel like as long as they're on this -- d... |
**Justin Garrison:** And the foundations that were laid 10 plus years ago with things like configuration management and automation and just getting the organization ready to adopt that sort of massive changes, and automating and trusting systems - that was foundational to everything we're doing today. There would be no... |
**Autumn Nash:** And all of this leads into how automated infrastructure has gotten, how we got so good at mass rollouts, and deployments, and just kind of the basis of -- and that stuff's not going to change. |
**Justin Garrison:** So let's not spoil the whole the whole interview. Let's jump right into it with Martin, and we'll come back afterwards and talk about CrowdStrike. |
**Break**: \[09:48\] |
**Justin Garrison:** Thank you so much, Martin Jackson, for joining us on the show today. Martin currently works at Red Hat, doing a lot of different automation and container stuff, but also what we we kind of were talking about in the pre-show was some of your work you did at Walmart with configuration management... A... |
**Martin Jackson:** Yeah, sure. So at Walmart I got into configuration management formally in -- it was about 10 years ago. I had been a network engineer where I was responsible for maintaining Walmart's IPAM systems, IP Address Management, DNS, DHCP. |
And one of the things that I'd done in the IPAM space was to help write and then be responsible for the rewrite and redesign of our in-store IPAM system, which maintained multiple DNS zones, DHCP configurations, with both pools and static configurations, and an introductory user interface that allowed field support and... |
After many years of doing this IP address management stuff, we were trying to do IPv6, and IPv6 still hasn't happened.. But I got a bit frustrated with that, and then I got the opportunity to move to infrastructure as a service team, which was just forming at the time. And so we evaluated a bunch of the leading solutio... |
So we did a competitive evaluation of a bunch of these configuration management products, and much to our surprise, open source Puppet won. So our initial configuration management deployment was open source Puppet using the form as the node classifier, and all of that stuff. |
**Justin Garrison:** And for anyone not familiar, Foreman is a Red Hat product that does mostly provisioning, but it has that configuration component. It gives you all your data about machines, and you can combine that metadata with a Puppet manifest, and then deploy a complete system. |
**Martin Jackson:** Yes, yes. Foreman is a pretty amazing thing in and of itself. It was a lot earlier in its lifecycle. It's grown the ability to do an awful lot of interesting things. It's the basis of the Red Hat satellite product, in fact. So there's a lifecycle piece, there's content, there's a lot of that stuff. ... |
Okay, so we were in the process of doing configuration management. We decided on Puppet as being the solution that we were going to use... And so we got heavily into the prospect of writing content for it. We had a couple of really experienced systems engineers that wrote the bulk of our content for the initial deploym... |
\[17:46\] The really exciting thing -- there are projects like this that go on at companies like Walmart all the time. One of the real learning moments for me was when we were talking about how excited we were about Puppet, and showing it to applications teams, and like "Hey, wow, we can make changes across the entire ... |
We learned through brutal experience that along with the ability to do infrastructure as code comes a responsibility of doing those deployments in a conservative and safe manner, because if you send a bad change to a lot of places very quickly, it can have a lot of consequences. |
**Justin Garrison:** Hello, CrowdStrike. |
**Martin Jackson:** Right. Right. We are recording this after the CrowdStrike incident, which now even completely non-technical people kind of understand is a bellwether as far as that goes. But for a lot of other reasons, it was very important that we improve on a lot of our situations. The way the internal Walmart ar... |
So you've got somewhere between four and a dozen server images per store. That racks up the node count really quickly. And the ability to assert policy, and to know for sure what configurations are across that entire state was really powerful. It was a capability that our infrastructure leadership really wanted to be a... |
**Justin Garrison:** Now, with that many stores and servers spread across that broad of an area, were you running as much as you could locally in the store, so internet disconnects are fine? You're just like "We have to deploy the whole stack every time in every area", which is probably - what percentage of overhead pe... |
**Martin Jackson:** We only deployed what absolutely had to be deployed in the store, at the store. Officially, the application-addressable capacity was about half a rack per store. And then with all the necessary backup and other infrastructure, it winds up being about a rack and a half, two racks, give or take. Netwo... |
**Justin Garrison:** Yeah. |
**Martin Jackson:** Deploying Foreman inside a store would have been nuts. We technically would have had the capacity to do it, but one of the things that we were trying to do was to provide a centralized reporting capability of what the current state of everything was. One of the biggest things that we were able to de... |
**Justin Garrison:** All of the kernels, yeah. |
**Martin Jackson:** ...from those 5000 stores... Yeah. Every single kernel that had been released -- |
**Justin Garrison:** It was out there. |
**Martin Jackson:** Several PTF kernels, which were what they call precision-targeted fixed things for very specific problems that had occurred at just that store. Everyone's like "Wow, that's really a problem. I would have thought it would have been a lot more consistent than that." Well, you would, but this isn't you... |
But that was one of the things that we used to help show that there was value in configuration management at that scale. |
**Justin Garrison:** Right. That amount of drift is just impossible to debug anything. Once you say "Oh, store 873 has this weird issue with this one thing that's not responding", it's like "Okay, well, how is it different than store 900?" It's like "Well, who knows?" |
**Martin Jackson:** Yes, that is exactly the problem that you run into. And we were training people to get used to the idea of being able to rely on certain configurations being a certain way. And when you're in the thick of it, your first thing is "Okay, check this, check this, check this, check this." If you're in th... |
**Justin Garrison:** \[22:32\] What did the application layer look here? So you have racks in stores, and you're deploying and configuring Linux on top of it... This is pre-containers, this is pre-Kubernetes stuff, so you're probably just deploying -- an application developer creates an RPM, throws it in a repo, and th... |
**Martin Jackson:** Oh, RPM would have been so much better than some of the actual things that we had to do... |
**Justin Garrison:** Git push. |
**Martin Jackson:** The store workflow was actually developed during the proprietary Unix era. The way that all of this happened was that there were a smorgasbord of potential commercial Unix'es that were available for store deployments. And did you know that NCR made a Unix once upon a time? Oh, buddy... Yes, they did... |
**Justin Garrison:** Everyone made Unix at the time. I'm sure every company out there had a Unix flavor. |
**Martin Jackson:** Yeah. Well, we had MP-RAS, HP-UX, and we had AIX. And all of the basic store deployment systems were based on this. And so all of the tool chains were based on the idea that your application could be running on any of these three Unix'es. Why three, you may ask? Because Walmart is a retailer, and Wa... |
So Walmart had for years operated the strategic vendor environment where vendors were expected to compete with each other on the basis of price. And this worked in technology, too. So for the different Unix variants, they basically threw it out for quarter era bids. "We're gonna retrofit 200 stores this quarter. What's... |
Now, years went by, and at this point of Walmart's history, it was solidly into virtualization. So it was all Zen and VMware at this point. So we didn't have the proprietary Unix component of all of this, but we did have a lot of the tooling that had existed from that era, because all the migrations and everything. So ... |
**Autumn Nash:** Oh, no... |
**Martin Jackson:** Fun fact. We tried several times during my career at Walmart to enumerate the number of applications that were on an individual store server, and no one could agree on the definition of applications. So that answer ranged somewhere between 800 and 5,000. These were individual software artifacts that... |
**Autumn Nash:** What was the purpose of having servers per store, instead of having it all in a centralized area? Why servers in the actual locations? And then running that many applications on those servers. Because it just seems it would be very hard to onboard new people, and to troubleshoot, and then to keep all t... |
**Martin Jackson:** \[26:17\] Oh, yeah, it was a full-time job for a lot of us keeping all that stuff up to date, for sure. But the reason it was done that way was that the basic origins of this architecture go back to the old mainframe days of the '80s and satellite networking. When they had it all centralized on a si... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.