Datasets:
willtheorangeguy
/
2021-Ship-It-Transcripts

Dataset card Data Studio Files
xet
 Community
text
stringlengths
0
1.8k
[2858.50 --> 2863.00] across our entire fleet, just because of natural inefficiencies between teams, right?
[2863.18 --> 2866.82] You need a new app out, you throw a couple of VMs out there, you call it a day.
[2867.26 --> 2871.88] And the CIOs job, part of it, is to reduce infrastructure costs, right?
[2872.18 --> 2874.72] And so the CIOs looked around, they said, oh, this is great.
[2874.78 --> 2876.58] We can bim pack the f*** out of this, right?
[2876.60 --> 2881.24] We can take all that stuff and just shove it into one massive cluster, save so much money.
[2881.24 --> 2884.08] And I think that drove a lot of initial Kubernetes adoption.
[2884.22 --> 2887.48] I mean, obviously, there was a lot of grassroots adoption of Kubernetes, but there was also
[2887.48 --> 2891.78] a lot of, there was a lot of adoption coming out of the IT organizations in larger companies
[2891.78 --> 2894.20] because of that driving factor.
[2894.60 --> 2900.70] Now, when the operators started using Kubernetes, they saw what I think of as the real benefits.
[2900.82 --> 2904.20] I don't think the benefit of Kubernetes is about orchestrating containers.
[2904.42 --> 2909.36] I think it's about that beautiful, idempotent, declarative, and ubiquitous API.
[2909.36 --> 2915.54] And especially when you start extending that into external services, external resources
[2915.54 --> 2922.62] that you're managing, like using, for example, Crossplane to provision AWS resources through
[2922.62 --> 2923.18] KubeCuttle.
[2923.30 --> 2924.74] It's a fantastic experience, right?
[2924.92 --> 2925.54] Yes.
[2925.66 --> 2929.24] And the operators looked at it and said, this whole Kubernetes thing is pretty cool.
[2929.50 --> 2932.42] However, Blast Radius is a thing, right?
[2932.42 --> 2937.14] And so if you've got everything in one big cluster, and especially those poor operators
[2937.14 --> 2946.90] who went through the 1.8 through 111 upgrade path got burned so many times on trying to upgrade
[2946.90 --> 2947.90] these clusters in place.
[2947.90 --> 2951.54] And they started developing these complicated blue-green cluster upgrade strategies where
[2951.54 --> 2953.22] they deploy an entirely new cluster.
[2953.42 --> 2955.50] And that's necessary and great.
[2955.76 --> 2960.52] But now we've figured out that, well, you should just be running many small clusters.
[2960.52 --> 2961.70] And there's two different ways you could do it.
[2961.70 --> 2966.06] You run a cluster per kind of bounded context for your microservices.
[2966.22 --> 2970.66] In other words, you could have a cluster just for your shopping cart stuff and a cluster
[2970.66 --> 2975.56] just for your front-end stuff and a cluster for your back-end and all that.
[2975.90 --> 2980.92] But a better way of doing it is to run all these clusters as homogenous workloads, where they
[2980.92 --> 2982.68] are all running identical workloads.
[2983.24 --> 2987.44] In fact, one of our clients is doing that, and they're referring to it as fleets internally.
[2987.44 --> 2990.36] So what they do is actually really smart.
[2990.62 --> 2995.06] They run a cluster in AWS per availability zone.
[2995.36 --> 2996.46] And that does a couple of things.
[2996.74 --> 2999.26] It's a natural dividing point for the different clusters.
[2999.78 --> 3004.88] And it means that they also keep all of their traffic inside each AD because all the services
[3004.88 --> 3007.86] in cluster A are always talking to other services in cluster A.
[3007.92 --> 3009.72] They don't try and do cross-cluster traffic.
[3010.18 --> 3013.58] And that saves them a good amount of money because they have a lot of networking that's happening
[3013.58 --> 3014.10] in AWS.
[3014.60 --> 3019.60] But also, it means that when they're upgrading these clusters, they can just upgrade one.
[3019.86 --> 3021.20] And if it goes sideways, who cares?
[3021.40 --> 3023.68] Burn it down, rebuild it, and you're fine.
[3023.94 --> 3027.00] You've only lost, what, 20%, 25% of your capacity?
[3027.18 --> 3028.46] And you just keep moving.
[3028.98 --> 3031.72] Now, of course, the big elephant here is state.
[3032.10 --> 3034.00] You can't do that with databases.
[3034.38 --> 3039.00] And so the best solution that we always propose to our customers is, look, if you're going to
[3039.00 --> 3043.06] run stateful workloads in Kubernetes, which, by the way, that's a lot of innovation points.
[3043.48 --> 3046.14] You really need a team to manage that if you're going to do that.
[3046.20 --> 3048.66] That's a dangerous thing to do as a small company.
[3048.98 --> 3053.52] But if you're going to run stateful workloads in Kubernetes, at least shove them into a smaller
[3053.52 --> 3055.48] cluster that you know you have to treat as a pet.
[3055.94 --> 3058.90] You've taken all of your other clusters, your stateless ones, and you've made them into
[3058.90 --> 3060.26] cattle, which is beautiful.
[3060.80 --> 3062.86] Then you constrain all your stateful workloads into one.
[3062.96 --> 3064.52] Or just use RDS.
[3065.06 --> 3067.74] Just externalize your databases entirely.
[3067.74 --> 3068.42] Right?
[3068.66 --> 3069.46] It's a tough problem.
[3069.64 --> 3073.46] And yeah, unless you've been solving that problem for some years, it's really difficult
[3073.46 --> 3074.18] to appreciate.
[3074.64 --> 3078.16] And even the operators, I'm glad that you mentioned it earlier for PostgreSQL.
[3078.46 --> 3079.78] Do you know how we run PostgreSQL?
[3079.96 --> 3080.32] How do you?
[3080.80 --> 3082.78] We run it as a stateful set.
[3083.04 --> 3085.34] No help, no operator, nothing like that.
[3085.62 --> 3088.40] And since we did that, it's been more stable.
[3088.58 --> 3092.12] It has not failed since we went to a stateful set.
[3092.24 --> 3095.68] Simple stateful set, PostgreSQL container, sorry, PostgreSQL image.
[3095.68 --> 3097.36] And what were you doing before that?
[3097.36 --> 3099.36] Were you doing RDS or were you doing?
[3099.66 --> 3106.96] We tried running the Crunchy data, PostgreSQL operator, and it failed because of replication.
[3107.48 --> 3110.36] Actually, we even covered this in like an episode at length.
[3110.46 --> 3114.74] But the point was the primary stopped replicating to the replica.
[3114.96 --> 3115.10] Yeah.
[3115.10 --> 3117.08] So the write-ahead log filled up on the primary.
[3117.50 --> 3119.10] The second it crashed.
[3119.24 --> 3121.54] The secondary could not be promoted.
[3121.62 --> 3125.84] The replica could not be promoted to primary because it was too far behind.
[3125.98 --> 3127.68] And then we didn't have a database.
[3129.62 --> 3129.98] Ouch.
[3129.98 --> 3133.46] We couldn't reboot the main one because the PVC filled up.
[3133.72 --> 3135.18] We couldn't resize the PVC either.
[3135.52 --> 3137.66] And we thought, nah, let's just crunch data.
[3137.86 --> 3142.20] We actually went to Zalanda one, the other PostgreSQL operator, and the same thing happened.
[3142.86 --> 3146.28] So obviously the networking, there was an issue at that point with networking.
[3146.78 --> 3153.02] And that broke replication, PostgreSQL replication, which resulted in a less stable database.
[3153.02 --> 3155.56] Yeah, but I mean, come on, that's not because of those operators.
[3156.16 --> 3158.70] You would have the same problem running a stateful set.
[3158.80 --> 3162.54] I think you probably changed other things at the same time as moving to a stateful set,
[3162.58 --> 3164.72] or maybe changed the way you use it or something like that.
[3164.74 --> 3165.56] We don't replicate.
[3165.92 --> 3166.66] Like it's single instance.
[3166.88 --> 3167.06] Oh, okay.
[3167.16 --> 3167.62] Well, there you go.
[3167.82 --> 3168.62] We back everything up.