Latest / Microsoft Cloud IT Pro Podcast / Episode 378 – Azure Compute Fleet
Transcript
- 0:03Welcome to episode 378
- 0:05of the Microsoft cloud It pro podcast recorded
- 0:09live. On 05/31/2024.
- 0:12This is a show about Microsoft 3 65
- 0:14in Azure from the perspective of It pros
- 0:16and end users where we discuss a topic
- 0:19or news and how it relates to you.
- 0:21Today, we're diving into a new preview feature
- 0:24that was recently announced at Microsoft build,
- 0:27Microsoft compute.
- 0:28Scott and Ben dive into what exactly the
- 0:31services is and some of the intricacies of
- 0:33what it entails. They also discuss what a
- 0:36deployment involves and what some of the possible
- 0:40use cases they see of the service based
- 0:42on the features implemented within the preview.
- 0:46You wanna get in a smash or adventures
- 0:48today? Yes. Like fleets, Azure fleets, boat fleets,
- 0:53helicopter fleets.
- 0:54Compute fleets of what? Azure compute fleets, fleets
- 0:58of
- 0:59computers. Yeah. III figured it's been a hot
- 1:01minute since we talked about Smash stuff and
- 1:04it's only fair that we get back to
- 1:06it. And this is a new preview feature
- 1:08that came out or was announced in the
- 1:11build time frame. What was that? Week ago,
- 1:142 weeks ago. Something like that. I don't
- 1:15know. Time has no meeting to me anymore.
- 1:17Could've have been, like, a month ago.
- 1:20A But it recently at Microsoft built. Yeah.
- 1:23May 20 24
- 1:25is when this 1 publicly came out. And
- 1:28landed in public preview. So I figured between
- 1:33virtual machines,
- 1:34virtual machine scale sets or Vms s,
- 1:38virtual machines scale sets,
- 1:40flex or Vms s Flex, which we've talked
- 1:43about in the past. And now compute fleet.
- 1:45We've got yet another perm mutation for at
- 1:49scale deployment of
- 1:51virtual machines. And I think this 1 is
- 1:53kind of interesting when it comes to the
- 1:57quote unquote scale aspect of it because it
- 2:00goes a little bit above and beyond what
- 2:02things like Vms do today,
- 2:04just with the units of compute that can
- 2:07be deployed, like, if you're thinking, like individual
- 2:09Vm counts,
- 2:11things like that. So real quick just to
- 2:13kinda ground everybody. So compute, like I said
- 2:17is this new thing. It's effectively a new
- 2:20infrastructure service
- 2:22within Azure around compute. And it's meant to
- 2:26kinda just
- 2:27streamline end to end that whole provisioning and
- 2:30management aspect
- 2:32of provisioning compute... Compute capacity
- 2:35across a whole bunch of different Vm types,
- 2:37potentially at the same time,
- 2:39availability zones, kinda of mix and match to
- 2:41your pricing model.
- 2:43Also that you can get to Vm deployments
- 2:46at scale that are performance because
- 2:49let's be honest, it's a little bit hard
- 2:51to do like multi vm deployments today, especially
- 2:54if you're doing, like, a scripted deployment or
- 2:56something, say with, like, powershell and just en,
- 2:59like, Hey. I'm gonna spin up a Vm,
- 3:00and then I'm gonna wait, and I'm gonna
- 3:02spin up another Vm, and I'm gonna spin
- 3:03up another Vm. It's a lot of Api
- 3:05calls. It's a lot of potential polling operations,
- 3:08things like that. So the cool thing about
- 3:10this is we need to kinda talk about
- 3:12it as, like, a
- 3:14infrastructure service or an infrastructure component.
- 3:17It's a single Api call to
- 3:21do an Azure compute fleet deployment, which is
- 3:26really kind of a cool thing, especially when
- 3:29you think about the scale of it. So
- 3:31you can deploy up to 10000 vms,
- 3:34in a single call through
- 3:36compute fleet.
- 3:38You've got a bunch of different settings in
- 3:39there where you can prioritize things like deployment
- 3:43speed,
- 3:43you could prioritize
- 3:45operational costs, say, like spot compute versus
- 3:48regular, pay go, compute,
- 3:50you can kinda have a mix and match
- 3:52model, potentially balance both those dimensions together
- 3:56because you're deploying different units of compute and
- 3:59those different units of compute the Skews underlying
- 4:01Skus and compute. They all have different costs
- 4:04associated with them. You've also got to, kinda
- 4:06cost management aspects.
- 4:08And being able to mix and match pricing
- 4:10models in a wide swath single Api call
- 4:14deployment.
- 4:16What else Oh?
- 4:17You can also... Because you're deploying potentially so
- 4:20many virtual machines at a time. Like, but
- 4:23deploying 10000 Vms at once isn't necessarily a
- 4:26small ask. When it comes to things like
- 4:28quotas, availability of that compute in a region,
- 4:32anything along that
- 4:33dimension as well. So it's kinda nice from
- 4:36the fire in forget aspect to say, hey,
- 4:38fire off an Api call and kinda load
- 4:40me up based on what my quota is
- 4:43for my subscription. A given region things like
- 4:46that along the way. So it looks like
- 4:49a kind of fun nifty new feature. It
- 4:52is definitely not an f for everybody feature,
- 4:55I think, given the scale aspect of it.
- 4:58I don't know that many people are looking
- 5:00to deploy deploy 10000
- 5:02vms at once, at least not broadly.
- 5:05But there's definitely customers out there that would
- 5:09love to deploy 10000
- 5:11and frankly, more than 10000 virtual machines
- 5:15at a given and go to get through
- 5:17some of the stuff. So you think about,
- 5:18like, customers with a large Ai training workload
- 5:22kind of thing. That's a ton of compute,
- 5:25ton jeep can be fairly ep.
- 5:28So, you know, you might not wanna run
- 5:29the thing the whole time. But, like, yeah,
- 5:30Definitely, while we're training a model, we need
- 5:32to get that. Those units have compute up
- 5:34and running and those Vms configured all those
- 5:37kinds of things along the way. So looks
- 5:40like a fun 1. Yeah. This was interesting
- 5:42and you first brought it up me We're
- 5:44like, hey, ben. Go check this 1 out
- 5:45because
- 5:47can't remember. It was a couple months ago.
- 5:49We were talking about Av. And I was
- 5:51creating some maybe the environments where we were
- 5:54actually, like, destroying and
- 5:57rebuilding or tearing down and recreating Vms on
- 6:00a nightly basis,
- 6:01and we were only doing...
- 6:03I mean, we're only doing like, 30 or
- 6:0440 Vms.
- 6:05But we were running into issues where when
- 6:07we did all 30 of them at once.
- 6:09Again, it was a bunch of single Api
- 6:10calls. Not all of them are coming back
- 6:12up. We split it up into 5 or
- 6:1305:10 vms at once. And it worked better.
- 6:16And you're were like, hey Ben, now you
- 6:17can do complete. Compete and do, like,
- 6:19do compute fleet and do, like, a whole
- 6:22bunch from that once. And it led to
- 6:24a little bit of like that use case
- 6:26where
- 6:27this is very much, like, when you stand
- 6:29up a compute fleet, go to play a
- 6:31bunch of them, but then you have to
- 6:34continue
- 6:35managing it through the fleet. And that was
- 6:371 thing I was curious about when you
- 6:39told me to go check it out is
- 6:40it's like, can I use a fleet to
- 6:42go deploy 30 vms? And then like, in
- 6:45the case of Av, if I wanna scale
- 6:47up or scale down,
- 6:49how does that look within a compute fleet?
- 6:51Like, can I go? Turn Vms off or
- 6:54delete them or
- 6:55how do I manage these museums once it's
- 6:58the fleet. And
- 6:59this is not really for that use case.
- 7:01Like you said, this is I'm gonna go
- 7:03deploy a bunch of Vms, but then you
- 7:05continue to manage
- 7:07that capacity
- 7:08of Vms
- 7:10within the fleet, and it's not something that
- 7:13you could easily...
- 7:14I would say it's not something you can
- 7:15easily scale. It's like what you said. We
- 7:17wanna go spin up 10000 vms. We're gonna
- 7:19go spin 5000 vms for a workload and
- 7:22we were even talking probably more of like
- 7:24an a a ep type workload where it's...
- 7:27We're gonna go train something for a week
- 7:29or train something for a month or...
- 7:32I mean, certain industries have certain times a
- 7:34year where it's busy and they're spinning up
- 7:36a huge number of Vms. Think of shopping
- 7:39sites over the month of November and December
- 7:41where... Mh. They need a massively scale up.
- 7:44For a short period of time, and then
- 7:46they scale back down or they just frankly,
- 7:48like blow them away. It's like, we hit
- 7:50the middle of January,
- 7:52nobody's shopping anymore because everybody spends their Christmas
- 7:55money and has done shopping. Let's just delete
- 7:57everything. Like, this is very much
- 8:00spin up a bunch of Vms for a
- 8:02short period of time and then get rid
- 8:04of them all. Is kinda how I interpret
- 8:07this. It's the Nuance potentially of that statement
- 8:10of
- 8:12compute fleet is kind of like an infrastructure
- 8:14service
- 8:15versus
- 8:16a holistic
- 8:18compute management service.
- 8:20Right? So I think about things maybe, like,
- 8:22the
- 8:23Vms s or Vms Flex as being a
- 8:26little bit more feature rich there. Like, 1
- 8:28of the things with compute fleet, at least
- 8:31I haven't seen a way to do it.
- 8:33It's not documented
- 8:34nor, like, that I see, like, an easy
- 8:37way to do it through the Api surface
- 8:38or the portal or anything like that. But,
- 8:40like, 1 of the big differences is is
- 8:42when you deploy a compute,
- 8:44you're
- 8:45deploying compute, you're not necessarily configuring that compute.
- 8:49So when I think about, like, a traditional
- 8:51single Vm deployment or, like, a Vms fast
- 8:54deployment things like that. It's not just about
- 8:57deploying the compute. It's also about configuring and
- 9:00managing that compute throughout its life cycle. So
- 9:03I don't know. Maybe I'm standing up, like
- 9:05that, you know, traditional multi tier web application
- 9:09thing where, you know, I've got a front
- 9:10end, some middle wear and a back end.
- 9:13I might do that in S flex, and
- 9:15then within my deployment
- 9:17configuration, I would say, hey for this Ubuntu
- 9:20based front end deploy Ng x on it.
- 9:22And load these Ssl cert and do this
- 9:26configuration for Ng x.
- 9:28My middle aware, do this kinda
- 9:31configuration for my Api hosting on my data
- 9:35side, like, with my
- 9:37sql server. Like, Sam doing, like, my sequel
- 9:39or something like that. Well, go ahead and
- 9:41actually spin up my sequel install it on
- 9:43the box for me or even deploy that
- 9:45as a bad service like whatever it happens
- 9:47to be. And you don't have that same
- 9:50flexibility within compute.
- 9:52Compute is literally, like, fire and forget I
- 9:55want to create just a ton of virtual
- 9:58machines at the same time. I'm gonna have,
- 10:00like, min and Max targets
- 10:02for
- 10:03how much I want to deploy. And and
- 10:06of what type I want to deploy. But
- 10:08then once it's deployed, it's up to you
- 10:11to go and stand up your applications,
- 10:14do all that kind of management
- 10:17on top of it. So it's kinda nice
- 10:20in that it's got a single Api, fire
- 10:23and forget for the at scale deployment piece.
- 10:26It'd be interesting to see where this goes
- 10:27in the future, like, if they bring in,
- 10:29like, at scale management as well. Like, I
- 10:32would love to have a way. And if
- 10:34it's possible today, maybe the docs just need
- 10:35to be updated. I would love to have
- 10:37a weighted to say, like, go in and
- 10:39create an arm template to do a compute
- 10:41fleet deployment. And then within that fleet, as
- 10:44I'm defining my Vm skus, and I'm saying,
- 10:46okay, you know, I want 10 of this
- 10:48spot, up to a hundred of this spot.
- 10:50I want 10 of these, you know, ds
- 10:52s 96,
- 10:54up to, you know, 50 of them, whatever
- 10:56happens to be. Like, in that same arm
- 10:59template, having, like, further child properties that I
- 11:02can call out to so that I can
- 11:04do things, like, execute boot automations, so maybe,
- 11:07like, cloud in it scripts, things like that
- 11:09on a Linux box
- 11:10just to get them up and running. But
- 11:13that stuff's kind of not there today.
- 11:17The other thing is it's a little bit
- 11:19different and I do think of as maybe
- 11:22a little bit more of an ep thing.
- 11:24So when you fire off a request to
- 11:28instant initiate a compute fleet. So you go
- 11:30in fire off this Api surface. And again
- 11:32you do that through arm template. You do
- 11:34it through the portal. You do it through
- 11:35the Arm Apis. Like, however you do it?
- 11:37You fire off that request.
- 11:39That request can take... It's effectively asynchronous compute
- 11:42deployment at that point, like, it all happens
- 11:44in the background it just churn and churn
- 11:46and churn.
- 11:47These are, like, super long running jobs.
- 11:51So, like, an initial compute fleet request can
- 11:53be active for up to a year.
- 11:55365
- 11:57days.
- 11:58If you need to clear a request, like,
- 12:01hey I I did the fire and forget
- 12:02saying of that Api, and now there's a
- 12:04bunch of async stuff happening in the background.
- 12:06Your only option is, like, you can't really,
- 12:08like, pause it. You can just delete the
- 12:11compute fleet request. And if you do that,
- 12:13all the Vms inside that compute fleet come
- 12:16down at the same time.
- 12:18So, yeah. I... It's a little bit weird.
- 12:20In that it doesn't have parity with things
- 12:22like,
- 12:24Flex. It's its own new
- 12:27weird
- 12:28thing. Like, whether you wanna think about it
- 12:30as
- 12:30potentially, like, a new Api surface, like, as
- 12:33a new infrastructure service,
- 12:35built on top of existing constructs like virtual
- 12:37machines, things like that.
- 12:39The public docs refer to it as a
- 12:43building block. So it's just kinda like this
- 12:45foundational little block or, like Lego brick. That's
- 12:48going to
- 12:50accelerate your
- 12:51access to compute capacity in a region. So
- 12:54if you think about it that way, like,
- 12:55hey, I just need to
- 12:57secure a bunch of compute. Really good for
- 12:59that. I need to secure a bunch of
- 13:02compute, and I need to configure it, and
- 13:04I need all this post
- 13:05configuration. I need down level management and all
- 13:07this other stuff. I don't know. It might
- 13:09not be the best thing for that, but
- 13:11it's not, like Vms or those other things
- 13:13are going away. You might have to kinda
- 13:16mix and match to or it just very
- 13:19well could not be the right thing for
- 13:21you along the way. I'm curious, and I
- 13:23don't know that you have an answer because
- 13:24we talked about this a little bit, and
- 13:26this is in the Faq. The 365
- 13:29day thing you brought up. Is if you
- 13:30go in and look at the Faqs for
- 13:33compute fleet. It says how long does my
- 13:35compute fleet request active, compute fleet requests are
- 13:38active for 365
- 13:39days. If you delete it, it deletes all
- 13:42the Vms.
- 13:43If I don't delete it, it continues the
- 13:45Vms continue to run and charge you. Does
- 13:47this mean that if I request
- 13:50or maybe I already know what you're gonna
- 13:52answer. I'm curious how this plays out is
- 13:54could it actually take 365
- 13:57days to provision all 10000 vms? Or does
- 14:01this mean that, like, you put in a
- 14:03request
- 14:04And within that request, you're setting, I need
- 14:07certain amounts of Vms,
- 14:08and you can go in and update your
- 14:11fleet, I believe to change that number of
- 14:13Vms,
- 14:14Like, do you only have a year to
- 14:16change it and after a year, a complete
- 14:18fleet or a compute fleet is like.
- 14:20Stock or locked, I'm I can't imagine anybody
- 14:23would wanna spin up 10000 vms and not
- 14:25know if it's gonna complete in a week
- 14:28in a year.
- 14:29Like, the whole
- 14:31what is our active request
- 14:34mean?
- 14:36Don't know. Right? Like, I I... So a
- 14:38couple of things. So it's in preview.
- 14:41I I think you've gotta kinda look at
- 14:42it from that lens of, hey, it's all
- 14:45not gonna be there, and I think as
- 14:47questions or answer or, like, as questions or
- 14:50asked
- 14:50those will get added and and build some
- 14:52of that out over time. You know, do
- 14:55you wanna wait a year to deploy your
- 14:5810000 Vms? No, Probably not, but that's why
- 15:01you're gonna have men's and max. Like, so
- 15:03you had the portal up earlier.
- 15:05I was gonna say we should walk through
- 15:07that and just talked through what a deployment
- 15:09looks like. Yeah. I think that gives you
- 15:10a little bit of kind of a a
- 15:12view into
- 15:14that. World and what some of the dimensions
- 15:18are that you're looking at. So, like, any
- 15:20other arm deployment. Right?
- 15:22Subscription,
- 15:23resource groups, that's pretty easy. You're gonna have
- 15:25a resource name,
- 15:27your fleet name, what region are you deploying
- 15:30into which this is limited right now in
- 15:33preview. It's East. East Us 2 west 2
- 15:35west 2 and West Us. So really East
- 15:37and West, the original and 2 are the
- 15:40only 4 options at this point in time.
- 15:42Effectively some Us hero regions in Azure. Yeah.
- 15:46So you've got that
- 15:47availability zones is
- 15:49your
- 15:50next 1. So
- 15:52do you wanna use Az z's for your
- 15:54Vm as to deploy and then just like
- 15:56a regular Vm, you know, zone 1, zone
- 15:592,
- 16:00zone 3 kind of thing.
- 16:02What is
- 16:03your
- 16:04security type for those virtual machines? So I
- 16:07think we discussed this a couple episodes back.
- 16:10Trusted launch Vms or the default everywhere right
- 16:13now, but you can revert to standard,
- 16:17or you can do confidential computer on the
- 16:19way.
- 16:20Know, trusted launch are the ones that give
- 16:21you, like, you've got the screen up now.
- 16:23Those are the ones that give you secure
- 16:24boot. They give you virtual T pms.
- 16:27And then some of the
- 16:29integrity monitoring around, like, memory things like that
- 16:33coming in from the hyper visor.
- 16:36Single image that you're gonna choose. So it...
- 16:39It's kind of interesting. Like, again, this isn't
- 16:42like Has flex. Where you would say, hey.
- 16:45Maybe I have this Vm. That's on a
- 16:46Ubuntu. I have this 1 that's on windows,
- 16:48things like that. That doesn't occur over here.
- 16:51So you just kinda pick a single image,
- 16:54and then you're lining up the units of
- 16:56compute that are all gonna use that image
- 16:59underneath it. Yep. And you can do with
- 17:01that image, you can do your custom images
- 17:03to you. So you still get that option
- 17:05like Go see marketplace
- 17:06images, see other images, where you can go
- 17:09select my images, shared images.
- 17:11You can do ubuntu Red hat, oracle windows,
- 17:15like you have all of your normal
- 17:18images that you would expect. So this isn't
- 17:20limiting in that aspect. It's just limiting and
- 17:22that you could only pick 1 for your
- 17:24entire fleet. Correct. Yeah.
- 17:26Yep.
- 17:30Do you feel overwhelmed by trying to manage
- 17:32your Office 3 65 environment are you facing
- 17:35unexpected issues that disrupt your company's productivity?
- 17:38Intelligent is here to help much like you
- 17:40take your car to the mechanic that has
- 17:42specialized knowledge on how to best keep your
- 17:44car running, intelligent helps you with your Microsoft
- 17:47cloud environment because that's their expertise. Intelligent keeps
- 17:51up with the latest updates in the Microsoft
- 17:53cloud to help keep your business running smoothly
- 17:55and ahead of the curve. Whether you are
- 17:57a small organization with just a few users
- 17:59up to an as of several thousand employees.
- 18:02They want to partner with you to implement
- 18:04and administer your Microsoft cloud technology.
- 18:08Visit them at intelligent dot com slash podcast
- 18:12that's INTELLIGINK
- 18:16dot com slash podcast
- 18:19for more information or to schedule a 30
- 18:21minute call to get started with them today.
- 18:23Remember intelligent focuses on the Microsoft cloud, so
- 18:27you can focus on your business.
- 18:31Next up for you is the types virtual
- 18:34machines. So I think this is an interesting
- 18:361 Like I I'm Mentioned in kinda of
- 18:37the opening that you can do mix and
- 18:39match.
- 18:40For cost optimizations. So being able to mix
- 18:42and match regular Vms and spot Vms.
- 18:46And then you can select the sizes or
- 18:49the Vm series that you wanna use.
- 18:51Along the way for those specific components. So
- 18:54I think it's a minimum today of 3
- 18:57Vm series that you have to select and
- 18:59you can go up to 15. I'm gonna
- 19:01spin up some big ones, Scott. I'm gonna
- 19:03go do 10000
- 19:05d 96
- 19:06b fives. Yeah. Go for it. That's a
- 19:08beef you bill. So, yeah, shooting can do
- 19:10us select a bunch of those. Like you
- 19:12said, spot regular.
- 19:13So... And then after that, you start picking
- 19:15capacity. This is the 1 where it gets
- 19:18interesting for me. Like, after you select all
- 19:20your sizes,
- 19:21now you can go choose these
- 19:23capacity preferences.
- 19:25So you can
- 19:26maintain a capacity, and this is once the
- 19:29target is matte,
- 19:31replace evicted vms to maintain target capacity. So
- 19:34it sounds like once you spin this up,
- 19:36you can set a capacity. So target capacity
- 19:38is a number between 1 and 10000,
- 19:41and it sounds like once you spin these
- 19:43up, you must be able to kick Vms
- 19:45out. Like you could essentially go in and
- 19:46say, I'm missed guessing eviction is essentially
- 19:50deleting that Vm saying this Vm isn't part
- 19:52of this this fleet anymore. It's a spot
- 19:56compute thing. 1 of the things that you
- 19:58run into with spot compute is it's kinda
- 20:01based on availability.
- 20:03And it's based on demand shaping within the
- 20:05region. So if you're doing spot,
- 20:08spot can be evicted, at any time. So
- 20:11it's really good for kinda, like, state workloads.
- 20:14If you have state full things, you know,
- 20:16you kinda make sure like, hey, that you're
- 20:18responding to things like eviction notices, so you're
- 20:20potentially persisting state and, like, object storage or
- 20:24off to disc someplace else.
- 20:26Things like that. So basically, what it's saying
- 20:28is if you wanna maintain your capacity, like,
- 20:31let's say you did a thousand spot vms,
- 20:34and you tell it to maintain. Well, over
- 20:36time, once you've turned those spot vms off,
- 20:38they're gonna get evicted just naturally through the
- 20:40life cycle as other customers come in and
- 20:42out of the region and the elasticity there.
- 20:44So what you're telling it to do is
- 20:46saying, like, hey, I need a thousand vms
- 20:49all the time. So if any get evicted,
- 20:52so it avail a hundred Vms and you
- 20:54go down to 900,
- 20:55the async engine in the background that's deploying
- 20:58that compute for you, will come in and
- 21:00request another hundreds spot Vms to get you
- 21:02back up to a thousand. And if you
- 21:04can't get you to a thousand, it'll kinda
- 21:06keep churning away in the back end. And
- 21:08keep trying to get you to where you
- 21:10need to be, or you can say don't
- 21:12maintain that capacity. Like, hey, I wanna start
- 21:14at, like, a thousand, and then stuff's just
- 21:16naturally gonna get it over time. And once
- 21:19it gets evicted,
- 21:20I don't really need it back or maybe
- 21:22I spun up another compute fleet or something
- 21:24else along the way. So you can kinda
- 21:27let it go
- 21:28naturally kinda downhill that way if you want
- 21:30to. This would get interesting then with the
- 21:32365
- 21:33days because as it technically then like, no
- 21:36longer active after 365
- 21:38days and if something gets evicted, it isn't
- 21:40active to maintain capacity. There is my question.
- 21:44I guess so. I don't know. Preview service,
- 21:47there's no Sla, anything like that. Figure it
- 21:49out. I imagine some of this like just
- 21:51to be brutally honest is kinda like a
- 21:53shot in the dark from folks going like,
- 21:55hey, how long do we need to do
- 21:57this?
- 21:58Is, like, a year sounds like a long
- 22:01time for a long running async job just
- 22:04to kinda keep sitting in the background. It's
- 22:06like a job scheduler.
- 22:08Especially when you consider that
- 22:10at least as of today for compute fleet,
- 22:12it doesn't have any cost. So it's kinda
- 22:14like Ak and that, you know, you're getting
- 22:16the management piece for free,
- 22:18but you're paying for the underlying compute along
- 22:21the way and and all the constructs that
- 22:23that come with that. So you're You're paying
- 22:25for the virtual machines. You're not paying for
- 22:27the ability to deploy those virtual machines and
- 22:30manage them
- 22:31via compute fleet. I don't know if part
- 22:34of it is maybe, like, a little bit
- 22:36of, like, just a cost thing, even like,
- 22:39internally. Like, hey, how many these jobs do
- 22:41we need to run? How long do they
- 22:42need to pull for? What does that look
- 22:44like and how does it manifest? I imagine
- 22:46a lot of that stuff gets cleaned up
- 22:48be before it go was to a
- 22:52production state. I would assume so because there's
- 22:54a lot of I have a lot of
- 22:56questions.
- 22:57Like, how some of that works that aren't
- 22:58answered. I think some of the the questions
- 23:00you have. You know, if folks are kinda
- 23:02sitting in the background and they're asking these
- 23:04same questions.
- 23:05A, lot of it is... This might not
- 23:07be the service for you. Like, that's true.
- 23:10There are other things out there. Like, this
- 23:12definitely,
- 23:13like, has a
- 23:16use case and a place. And if you're
- 23:19looking at it, and you're going, like, yeah,
- 23:20I don't really see the use case. Right
- 23:21I don't see place, like, that's okay. That's
- 23:23why things like Vm
- 23:25and everything else exists out there. Because like
- 23:28I said, this isn't Vms, So even though
- 23:31you can do the capacity deployment through a
- 23:33single Api call, you've still got the potential,
- 23:36like, configuration concerns and all the other stuff
- 23:39that comes along the way with it. So
- 23:40you have this little trade off there that
- 23:43you're kinda balancing
- 23:44between
- 23:45do I want to deploy a lot of
- 23:47compute or do I want to deploy a
- 23:50little less compute and then have all the
- 23:52manage ability aspects that come with it? Continuing
- 23:55down? Setting
- 23:56maintain capacity don't maintain, what is your target
- 23:59capacity? So what are you gonna set this
- 24:01at? And then you can set the eviction
- 24:02policy. So...
- 24:04That's all spot stuff, spot specific. Delete,
- 24:08allocation strategies, max hourly, price per spot is
- 24:10all the spot stuff, and then you get
- 24:12down to
- 24:14the Vm capacity so non spot Vms,
- 24:17and you lose some of those extra settings
- 24:19because now it's just regular compute where it's
- 24:21your capacity, and this one's...
- 24:24I don't know if this is spot. Yeah.
- 24:26Maybe you can help me out with this.
- 24:27Like, you can set a capacity for these
- 24:29normal Vms of
- 24:30say a thousand. And then you set your
- 24:32minimum starting capacity
- 24:34of something like 10, and this minimum starting
- 24:39capacity is essentially saying that Vm in the
- 24:42fleet our provision gradually. It starts with your
- 24:45desired in capacity if you specify it, so
- 24:48it could start with 10, and then it's
- 24:50gonna slowly grow
- 24:51up to 100. It doesn't say how slow,
- 24:54but this is not something you have with
- 24:56spot Vms. Like spot vms you just set
- 24:58the capacity, and I'm guessing it's does it
- 25:02all out once
- 25:04versus regular Vms where if you want to
- 25:07use set a minimum and slowly ramp up
- 25:09to your capacity. I think it's implied that
- 25:12spot is also going to ramp because by
- 25:16nature of spot via. By nature of spot.
- 25:18Like, you're basically saying with spot compute.
- 25:21Give me your extra compute in the region,
- 25:23and I'm gonna pay a little bit less
- 25:25for that compute, but the reason that I'm
- 25:27paying less is because you, Microsoft as the
- 25:30as the host and hyper scaler can rip
- 25:33that compute away from me. I can be
- 25:35evicted. And when you've ripped that compute away
- 25:37from me, you're gonna go give it to
- 25:39another customer. So let's say in this case,
- 25:42like, I'm the 1 who's consuming a lot
- 25:43of spot, and you come in and you
- 25:45say, hey. I'm here as a as a
- 25:47customer and I don't need spot compute. I
- 25:49need real non ep compute, and it's in
- 25:53the same series that I have in spot.
- 25:55It's in Microsoft best interest to e me
- 25:58out of there and just let it go.
- 26:00So I I think a lot of it
- 26:01is just kinda, like, sitting and and trying
- 26:03to balance
- 26:05that nature of spot versus
- 26:08everything else that's they're and available for you.
- 26:11The other interesting thing about this too is
- 26:14like, you said earlier, you can select up
- 26:16to 15 different sizes. So you could go
- 26:18in and select 15 different
- 26:21d series
- 26:22Vms, z e series Vm spot Vms, but
- 26:25you can't set capacity on a per
- 26:28sku level. I can go in here, and
- 26:31I can say, I want my target capacity
- 26:33to be 3000,
- 26:35my minimum is a hundred, but I've selected
- 26:385 different sizes.
- 26:39It's gonna go... Well, it's technically gonna break
- 26:43the way I have it now. If I'm
- 26:44spinning up 5 different sizes with 3000
- 26:47capacity.
- 26:48I believe it's spinning up 3000
- 26:52of each or is it taking? Now I'm
- 26:56questioning myself. Total capacity. So this is total
- 26:59capacity. It's taking 3000 dividing it by 5
- 27:02probably and spinning up
- 27:05600
- 27:06of each
- 27:07Skew that I've selected, doing that math in
- 27:09the background versus 3000 of each sku that
- 27:12I've selected. Again, very unclear from the documentation,
- 27:16even from the sense of, like, hey, self
- 27:18documenting code. If you go look at, like,
- 27:20the arm templates for these or, like, the
- 27:23actual arm
- 27:24endpoint that sits out there. I don't have,
- 27:28you know, 10000 course
- 27:30that I can go play with in my
- 27:32in my Pe go subscription. So, like, I
- 27:34wasn't even able to to 1 of these
- 27:36at scale, like, even, you know, even getting
- 27:38yourself lifted beyond, like, the default core limits
- 27:41and, like, a visual studio sub is
- 27:44pretty rough these days
- 27:46for how that stuff comes together. But, yeah,
- 27:48you're basically doing
- 27:50your allocations that way, and it's gonna split
- 27:53amongst them. It's unclear how it divides things
- 27:56up. So the interesting thing is it's not
- 27:59exposed in the portal really or at least
- 28:01it is a math back clearly between with
- 28:03the options in the portal. But in the
- 28:05arm template, if you go look at the
- 28:06arm template. So Mh. The spot profile and
- 28:10the regular priority profile.
- 28:13They have this property that's called
- 28:16allocation strategy.
- 28:17And
- 28:18allocation strategy
- 28:20can be capacity optimized,
- 28:22or it can be lowest priced.
- 28:25So that's kind of interesting because
- 28:28then what it's doing, if I'm reading it
- 28:30the right way based on the Api, is
- 28:32it using the logic of, hey, you selected
- 28:345 d series Vms.
- 28:37What are you phonetic clicking on over there?
- 28:39So so you've
- 28:41nuts. Yes.
- 28:43You've selected 5 d series Vms, you can
- 28:46go in and you can make your Alex
- 28:47vacation strategy to be lowest price. And then
- 28:50what it'll do based on that is,
- 28:53I... It would literally choose the lowest price
- 28:55Vm series. And it would start with deploying
- 28:58those first and then come behind it, like,
- 29:00it would just keep kinda incrementally stepping up
- 29:02the ladder based on the price. Or like
- 29:05I said, the other dimension is capacity optimized.
- 29:07So what I was clicking and I was
- 29:08trying to see what a... I can't find
- 29:10small Vms. That's another weird thing with this.
- 29:13Like the smallest Vm I was able to
- 29:15find to select was 8 cores. I was
- 29:17curious if I just selected 2 small ones
- 29:19and set it to 4. If I get
- 29:218 Vms or if I get 2 of
- 29:22each. Okay. Not to service for you if
- 29:24you're looking for small tabs. If I'm looking
- 29:26for small Vms, I know. I was trying
- 29:28to test it out to answer my questions.
- 29:30This very much does seem geared towards, like
- 29:33Ai training scenarios,
- 29:36Hp scenarios, where I'm gonna spin up a
- 29:39bunch of compute to
- 29:41run some kind of
- 29:43some kind of job on top of things.
- 29:46It's just going above and beyond, like, the
- 29:48click stops of some of the
- 29:50scale out capabilities of things like,
- 29:52you know, the current Hp offerings, current Vms
- 29:55offerings, but, yeah. It's a little bit of
- 29:58weird nuance there. So then once you set
- 30:00your capacities, it's really, after that, it's just
- 30:02what's my admin username password for all these
- 30:05vms, again, set it once, get it for
- 30:07everything.
- 30:09If you wanna do hybrid benefit, if you're
- 30:10doing windows,
- 30:12and then networking is just virtual network subnet,
- 30:15and then you can go in and tweak
- 30:17network interfaces.
- 30:19Really, if you wanna do Ns msg or
- 30:21accelerated networking. They do have an option 2
- 30:24to do a new load balance in front
- 30:26of it with networking. So relatively basic, you're
- 30:29essentially just saying dump all these vms in
- 30:31a network.
- 30:32Use an ns msg if you want to...
- 30:34And if you're gonna do any type of
- 30:35load balancing. Yeah. I believe you have to
- 30:37create a load balance. I don't think it's
- 30:40optional Do you? I'm gonna try it? Oh,
- 30:42basics. What did I fail on? Oh, field
- 30:44name? Benz test fleet. Who it validated without
- 30:47a load balance or Scott? Create gonna initialize
- 30:50the deployment. So I tried to spin it
- 30:52up through an arm template, and I just
- 30:54pulled their default arm template. So the default
- 30:56arm template at least as it's documented today.
- 30:59Includes a V handle load balance in it.
- 31:02So, yes. I just said V subnet, no
- 31:04load balance. So it doesn't look like it's
- 31:07required. We'll see what this actually creates when
- 31:10it goes and deploy everything.
- 31:12Well, mean if I have to come back
- 31:13and check in a year.
- 31:15Yeah. It's What else? There's that Faq out
- 31:17here. I'm trying to think if there was
- 31:19anything else in here
- 31:21that we wanted to cover, start and stop
- 31:23time. Yeah. No start and stop time. You
- 31:26spin it up? It's running. It will be
- 31:28interesting like you said, do they change some
- 31:29of the stuff going forward,
- 31:31but also like you said it's not for
- 31:33me. This is not a service I would
- 31:35use, the only use I kinda thought of
- 31:37it, but even as we've looked at it
- 31:39is if you wanted to do it for
- 31:40a lab. If you're spinning up, like a
- 31:42large lab for a learning environment. But even
- 31:45that, given you'd have to do an image
- 31:47and that there's no start and stop time,
- 31:49you're probably better off going other routes even
- 31:52for labs. Like you said it does tend
- 31:54to be... It appears to be more Hp
- 31:57Ai training based stuff that you were mentioning.
- 31:59I'd go back to you
- 32:01there are services to do those kinds of
- 32:04things. Right? Like, there there's literally lab services.
- 32:07Lab services. We've talked about that 1 before,
- 32:09which does sit out there. There's Dev. So
- 32:12there's kinda all these different
- 32:14constructs. So
- 32:15you do have to kinda go through and
- 32:17pick the right
- 32:18functionality,
- 32:19like, underlying deployment model, all all those things
- 32:23that
- 32:24that are gonna work for you within there.
- 32:26The only thing I'd call out is, like,
- 32:27if folks are watching this, and they're like,
- 32:29trying to get hands on with it and
- 32:30if you are more of, like, an arm
- 32:31template deployment person,
- 32:33versus
- 32:34a go and click in the portal to
- 32:36do your first deployment. You do have to
- 32:38register the resource provider. For this 1. So
- 32:42there's a little bit of a flag that
- 32:43needs to be cleared, register the resource provider
- 32:46and then you can go ahead and do
- 32:47your deployments.
- 32:48You
- 32:49didn't have to register the resource provider because
- 32:51you initiated through the portal and then that
- 32:54registration happens automatically. And then it broke because
- 32:57My subscription is not registered to use the
- 32:59name base Microsoft dot compute. You've never deployed
- 33:02a vm in there. Apparently not. That's a
- 33:05weird 1.
- 33:06I thought I had. Now I'm gonna go
- 33:08look. Oh, this is gonna be annoying. Yeah.
- 33:11We'll go look later. That wraps that'll be
- 33:13for another time.
- 33:15Yes yes. Yeah. So kind of a interesting
- 33:18service to go look at explore play with.
- 33:21If you wanna go spin up a whole
- 33:24lot of vms. Compute fleet. Get it done.
- 33:26Alright. As always. Thank you, Ben. Thank you.
- 33:29Enjoyed it. Enjoy your weekend. Try to stay
- 33:31cool in this. Florida got hot really fast
- 33:34this year. We went from summer to summer,
- 33:37like that yes. In an instant. So enjoy,
- 33:41we will talk to you again soon. Alright.
- 33:42Thanks, Ben.
- 33:45If you enjoyed the podcast,
- 33:47Go leave us a 5 star rating in
- 33:48itunes. It helps to get the word out
- 33:50so more It pros can learn about Office
- 33:523 65 in Azure.
- 33:54If you have questions you want us to
- 33:56address on the show or feedback about the
- 33:58show, feel free to reach out via our
- 34:01website, Twitter, or Facebook.
- 34:03Thanks again for listening and have a great
- 34:05day.