Does it have basic functioning other stuff? I am shocked at how our production usage of Fly has gone. Even basic stuff as support not being able to just... look up internal platform issues. Cryptic/non-existent error messages. I'm not impressed. It feels like it's compelling to those scared of or ignorant of Kubernetes. I thought I was over Kubernetes, but Fly makes me miss it.
I was hoping to migrate to Fly.io and during my testing I found that simple deploys would drop connections for a few seconds during a deploy switch over. Try a `watch -n 2 curl <serviceipv4>` during a deploy to see for yourself (try any one of the the strategies documented including blue-green). I wonder how many people know this?
When I tested it I was hoping for at worst early termination of old connections with no dropped new connections and at best I expected them to gracefully wait for old connections to finish. But nope, just a full downtime switch over every time. But then when you think about the network topology described in their blog posts, you realize theres no way it could've been done correctly to begin with.
It's very rare for me to comment negatively on a service but that fact that this was the case paired with the way support acted like we were crazy when we sent video evidence of it definitely irked me for infrastructure company standards. Wouldn't recommend it outside of toy applications now.
> It feels like it's compelling to those scared of or ignorant of Kubernetes
I've written pretty large deployment systems for kubernetes. This isn't it. Theres a real space for heroku-like deploys done properly and no one is really doing it well (or at least without ridiculously thin or expensive compute resources)
Yeah I had a similar experience where I got builds frozen for a couple days, such that I was not able to release any updates. When I emailed their support, I got an auto-response asking me to post in the forum. Pretty much all hosts are expected to offer a ticket system even for their unmanaged services if its a problem on their side. I just moved over all my stuff to Render.com, it's more expensive, but its been reliable so far.
That forum post just says what OP said, that they will ignore all tickets from unnmanaged customers. Which is a pretty shitty thing to do to your customers.
The cheapest plan that gets email support is nothing more than a commitment to spend a minimum of $29/mo on their services. That is, if you spend >=$29/mo, it costs nothing extra. Not what I'd call "managed".
> I've written pretty large deployment systems for kubernetes. This isn't it. Theres a real space for heroku-like deploys done properly and no one is really doing it well (or at least without ridiculously thin or expensive compute resources)
Have you tried Google Cloud Run(based on KNative) I've never used it in production, but on paper seems to fit the bill.
Yeah we're mostly hosted there now. The cpu/virtualization feels slow but I haven't had time to confirm (we had to offload super small ffmepg operations).
It's in a weird place between heroku and lambda. If your container has a bad startup time like one of our python services, autoscaling can't be used as latency becomes a pain. Its also common deploy services on there that need things like health checks (unlike functions which you assume are alive), this assumes at least 1 instance of sustained use as well, assuming you do minute health checks. Their domain mapping service is also really really bad and can take hours to issue a cert for a domain so you have to be very careful about putting a lb in front of it for hostname migrations.
I don't care right now but the fact that we're paying 5x in compute is starting to bother me a bit. A 8core 16gb 'node' is ~$500/month ($100 on DO) assuming you don't scale to zero (which you probably wont). Plus I'm pretty sure the 8 cores reported isn't a meaty 8 cores.
But its been pretty stable and nice to use otherwise!
A 6c / 12t Dedicated Server with 32GB of ram is 65$ a month with OVH
I do get that it is a bare server, but if you deploy even just bare containers to it, you would be saving a good bit of money and get better performance from it.
It depends on what the 6 cores are. Like I have a 8C/8T dedicated server sitting in my closet that costs $65 per the number of times you buy it. (Usually once.) The cores are not as fast as the highest-end Epyc cores, however ;)
At the $65/month level for an OVH dedicated server, you get a 6-core CPU from 2018 and a 500Mbps public network limit. Doesnt even seem like that good a deal.
There is also a $63/month option that is significantly worse.
We also run some small ffmpeg workloads and experimented with Cloud Run consuming Pub/sub via EventArc triggers. Since Cloud Run's opaque scaling is tied to http requests, EventArc uses a push subscription. In pub/sub these don't give you any knobs to turn regarding rate limiting/back pressure, so it basically tries to DoS your service and then backs off. This setup was basically impossible to tune or monitor properly.
Our solution was to migrate the service to Kubernetes using an HPA scaling on the number of un-acked messages in the subscription, and then use a pull subscription to ensure reliable delivery (if the service is down they just sit in the queue rather retrying indefinitely).
I'm convinced Cloud Run/Functions are only useful for trivial HTTP workloads at this point and I rarely consider them.
Triggered deploys to Kubernetes you mean? There's a million ways to solve this problem for better or worse. We use Gitlab CI so we invoke helm in our pipelines (I'm sure there's a way to do this with github actions), but there's also flux cd, argo, etc. etc.
We use Kubernetes (GKE) elsewhere so we already had this machinery in place luckily. I can see the appeal of CloudRun/Functions as a way to avoid taking that plunge
I have yet to gain positive experience with Cloud Run. I have one project with it, and Cloud Run is very unpredictable with autoscaling. Sometimes, it can start spinning up/down containers without any apparent reason, and after hunting Google support for months, they said it is an "expected behavior". Good luck trying to debug this independently because you don't have access to knative logs.
Starting containers on Cloud Run is weirdly slow, and oh boy, how expensive that thing is. I'm getting the impression that pure VMs + Nomad would be a way better option.
> I'm getting the impression that pure VMs + Nomad would be a way better option
As a long time Nomad fan (disclaimer: now I work at HashiCorp), I would certainly agree. You lose some on the maintenance side because there's stuff for you to deal with that Google could abstract for you, but the added flexibility is probably worth it.
> Starting containers on Cloud Run is weirdly slow
What is this about? I assumed a highly throttled cpu or terrible disk performance. A python process that would start in 4 seconds locally could easily take 30 seconds there.
I just use AWS EC2, load balancer, auto scaling groups. The user_data pulls and runs a docker image. To deploy I do an instance refresh which has no downtime. Obvious downside is more configuration than more managed services.
I have been using Google Cloud Run in production for a few years and have had a very good experience. It has the fastest auto scaler I have ever seen, except only for FaaS, which are not a good option for client-facing web services.
Cloud Run is compatible with KNative YAML but actually runs on Borg under the hood, not Kubernetes. At least when not using the "Cloud Run on GKE" option via Anthos.
Can you email the first two letters of my username at fly.io with more details? I'd love to find out what you've been having trouble with so I can help make the situation better any way I can. Thanks!
I've been a paying Fly.io customer for 3 years now, and for the past 18 months, I've had no real issue with any of my apps. In fact, I don't even monitor our Fly.io servers any more than I monitor S3 buckets; the kind of zero devops I expect from it is already a reality.
> it's one person who had issues
Issues specific to an application or one particular account have to be addressed as special cases (like any NewCloud platform, Fly.io has its own idiosyncrasies). The first step anyway is figuring out just what you're dealing with (special v common failure).
> looks like a theatre
I have had the Fly.io CEO do customer service. Some may call it theatre, but this isn't uncommon for smaller upstarts, and indicative of their commitment, if anything.
Yep they have terrible reliability and support. Couldn’t deploy for 2 days once and they actually told me to use another company. Unmanaged dbs masquerading as managed. Random downtime. I could go on but it’s not a production ready service and I moved off of it months ago.
The header at the top of their Getting Started is "This Is Not Managed Postgres
" [1]
and they have a managed offering [2] in private beta now...
> Supabase now offers their excellent managed Postgres service on Fly.io infrastructure. Provisioning Supabase via flyctl ensures secure, low-latency database access from applications hosted on Fly.io.
I was shocked because I didn't realise it wasn't managed. Even Digital Ocean offer managed Postgres.
If you are offering a service like Fly I think the database should be managed personally, the whole point of Fly.io is to provide abstractions to make production simpler.
Do you think the type of user who is using fly.io is interested in or capable of managing their own Postgres database? I'd rather just trust RDS or another provider.
Yep,
I never really trusted managed databases. It just feels like one of those things that's so important to your app that not having full control of it is weird.
Unfortunately this is a pretty common story. Half the people I know who adopted Fly migrated off it.
I was very excited about Fly originally, and built an entire orchestrator on top of Fly machines—until they had a multi-day outage where it took days to even get a response.
Kubernetes can be complex, but at least that complexity is (a) controllable and (b) fairly well-trodden.
Kubernetes on AWS, GCP, and Linode are all controllable and well-trodden.
I definitely understand the comparison between Kubernetes and fly. You have couple apps that are totally unrelated, managed by separate teams, and you want to figure out how you can avoid the two teams duplicating effort. One option is to use something like fly.io, where you get a command line you run to build your project and push the binary to a server. Another option is to self-host infrastructure like Kubernetes, and eventually get that down to one command to build and push (or have your CI system do it).
The end result that organizations are aiming for are similar; developers code the code and then the code runs in production. Frankly, a lot of toil and human effort is spent on this task, and everyone is aiming to get it to take less effort. fly.io is an approach. Kubernetes is an approach. Terraform on AWS is an approach.
That’d be a slightly more valid comparison albeit flyctl is much less ambitious by choice and design. That said, using flyctl to orchestrate your deployments is not the only way to Fly. Example:
I don't see any reason use Fly. There are more mature, feature-richer and cheaper solutions out there. We have the big complex ones like AWS, Azure, GCP, the easier more affordable all rounder like DO, Render, the hosting plattforms like Vercel, Heroku and finally the biggest bang for your money barebones like Hetzner.
Why should I choose Fly? How come they are so prominent on hackernews? Are they backed by VC and get their default 400 upvotes by backers? I get the impression that Fly posts here are kind of sponsored.
It's hard to tell how meaningful the reviews are. I have used AWS, GCP, DigialOcean, and Linode throughout my career. Every single one of these, through no fault of myself or my team, messed up and caused downtime. Like, you can get most SRE types in a room to laugh if you blurt out "us-east-1", because it's known to be so unreliable. And yet, it's where every Fortune 500 puts every service; we laugh about the reliability and it's literally powering the economy just fine.
So yes, a lot of people on HN complain about fly's reliability. fly posts to HN a lot and gives them the opportunity. Is it actually meaningful compared to the alternatives? It's very hard to tell.
First: this is 100% a "live by the sword, die by the sword" situation for us. We're as aware as anybody about our weird HN darling status (this is a post from two months ago, about an announcement from many months ago, that spent like 12 hours plastered to the front page; we have no idea why it hit today, and it actually stepped on another thing we wanted to post today so don't think we secretly orchestrated any of this!). We've allowed ourselves to be ultra-visible here, and threads like this are natural consequence.
Moreover: a lot of this criticism is well warranted! I can cough up a litany of mitigating factors (the guy who stored his database in ephemeral instance storage instead of a volume, for instance), but I mean, come on. The single most highly upvoted and trafficked thing we've ever written was a post a year ago owning up to reliability issues on the platform. People have definitely had issues!
A fun cop-out answer here is to note all the times people compare us to AWS or Cloudflare, as if we were a hyperscaler public cloud. More fun still is to search HN for stories about us-east-1. We certainly do that to self-sooth internally! And: also? If your only consideration for picking a place to host an application is platform reliability? You're hosting on AWS anyways. But it's still a cop-out.
So I guess I'd sum all this up as: we've picked a hard problem to work on. Things are mathematically guaranteed to go wrong even if we're perfect, and we are not that. People should take criticisms of us on these threads seriously. We do. This is a tough crowd (the threads, if not the vote scores on our blog post) and there's value in that. Over the last year, and through this upcoming year, staffing for infra reliability has been the single biggest driver of hiring at Fly.io, I think that's the right call, and I think the fact that we occasionally get mauled on threads is part of what enabled us to make that call.
(Ordinarily I'd shut up about this stuff and let the thread die out itself, but some dearly loved user of ours took a stand and said they'd never had any problems on us, which: you can imagine the "ohhhhh nooooooo" montage that took place in my brain when I read that someone had essentially dared the thread to come up with times when we'd sucked for some user, so I guess all bets are off. Go easy on Xe, though: they really are just an ultra-helpful uncynical person, and kind of walked into a buzzsaw here).
I also don't know why HN is so upset about people willing to help out in the threads. The way I see it is, if you talk about your product on HN, inevitably someone will remember they have a support inquiry while HN is open, and ask it there instead of over email. Since employees are probably reading HN, they are naturally going to want to answer or say they escalated there. I don't think it's some sort of scam, just what any reasonable person would do.
It's become a YC cliche, that the way to get support for any issue is to get a complaint upvoted to the top of a thread. People used to talk about "Collison installs", which are real-use product demos that are so slick your company founder (in this case Stripe's 'pc) can just wander around installing your product for people to evangelize it; there should be another Collison term for decisively resolving customer support issues by having the founder drop into a thread, and I think that's the vibe people are reacting to here.
ok possibly not alone, maybe the issues happened before I started using them extensively. I've had ~no downtime that affects me in 7 months.
I do wish they had some features I need, but their support and responses are top notch. And I've lost much less hair and time than I would going full-blown AWS or another cloud provider.
To be fair most hosting providers come with plenty of public complaints about downtime. The big ones do way better, the best one is AWS, then GC and last Azure. They cost stupid money though.
Digital ocean has been terrible for me, some regions just go down every month and I lose thousands of requests, increasing my churn rate.
Fly.io had tons of weird issues but it got better in the last months. It's still very incomplete in terms of functionality and figuring out how to deploy the first time is a massive pain.
My plan is to add Hetzner and load balance with bunnycdn across DO and H
Actually here is a good example: Cloudflare. Sure people complain a ton about privacy but I haven't seen a single complaint about the reliability of Cloudflare Workers or similar product in the dozens of threads I've seen on HN
this is what I thought, until once I spent two days to publish a new, trivial code change to my Fly.io hosted API — it just wouldn't update! And every time I tried to re-publish it'd give me a slightly different error.
When it works, it's brilliant. The problem is that it hasn't worked too well in the last few months.
Hi, author of the post and Fly.io devrel here in case anyone has any questions. GPUs went GA yesterday, you can experiment with them to your heart's content should the fraud algorithm machine god smile upon you. I'm mostly surprised my signal post about what the "GPUs" are didn't land well here: https://fly.io/blog/what-are-these-gpus-really/
I'd be fascinated to hear your thoughts on Apple hardware for inference in particular. I spend a lot of time tuning up inference to run locally for people with Apple Silicon on-prem or even on-desk, and I estimate a lot of headroom left even with all the work that's gone into e.g. GGUF.
Do you think the process node advantage and SoC/HBM-first will hold up long enough for the software to catch up? High-end Metal gear looks expensive until you compare it to NVIDIA with 64Gb+ of reasonably high memory bandwidth attached to dedicated FP vector units :)
One imagines that being able to move inference workloads on and off device with a platform like `fly.io` would represent a lot of degrees of freedom for edge-heavy applications.
Well, let me put it this way. I have a MacBook with 64 GB of vram so I can experiment with making an old-fashioned x.ai clone (the meeting scheduling one, not the "woke chatgpt" one) amongst other things now. I love how Apple Silicon makes things vroomy on my laptop.
I do know that getting those working in a cloud provider setup is a "pain in the ass" (according to ex-AWS friends) so I don't personally have hope in seeing that happen in production.
However, the premise makes me laugh so much, so who knows? :)
This is right on time. I'm evaluating "severless" GPU services for my upcoming project. I see on the announcement that pricing is per hours. Is scaling to zero priced based on minutes/seconds? For my workflow, medical image segmentation, one file takes about 5 minutes.
Part of it is for people that want to do GPU things on their fly.io networks. One of the big things I do personally is I made Arsène (https://arsene.fly.dev) a while back as an exploration of the "dead internet" theory. Every 12 hours it pokes two GPUs on Fly.io to generate article prose and key art with Mixtral (via Ollama) and an anime-tuned Stable Diffusion XL model named Kohaku-XL.
Frankly, I also see the other part of it as a way to ride the AI hype train to victory. Having powerful GPUs available to everyone makes it easy to experiment, which would open Fly.io as an option for more developers. I think "bring your own weights" is going to be a compelling story as things advance.
Enough that I'd probably need to write a blogpost about it and answer some questions that I have about it. The biggest one I want to do is a sentiment analysis of these horoscopes vs market results to see if they are "correct".
Because I have secret magical powers that you probably don't, it's basically free for me. Here's the breakdown though:
The application server uses Deno and Fresh (https://fresh.deno.dev) and requires a shared-1x CPU at 512 MB of ram. That's $3.19 per month as-is. It also uses 2GB of disk volume, which would cost $0.30 per month.
As far as post generation goes: when I first set it up it used GPT-3.5 Turbo to generate prose. That cost me rounding error per month (maybe like $0.05?). At some point I upgraded it to GPT-4 Turbo for free-because-I-got-OpenAI-credits-on-the-drama-day reasons. The prose level increase wasn't significant.
With the GPU it has now, a cold load of the model and prose generation run takes about 1.5 minutes. If I didn't have reasons to keep that machine pinned to a GPU (involving other ridiculous ventures), it would probably cost about 5 minutes per day (increased the time to make the math easier) of GPU time with a 40 GB volume (I now use Nous Hermes Mixtral at Q5_K_M precision, so about 32 GB of weights), so something like $6 per month for the volume and 2.5 hours of GPU time, or about $6.25 per month on an L40s.
In total it's probably something like $15.75 per month. That's a fair bit on paper, but I have certain arrangements that make it significantly less cheap for me. I could re-architect Arsène to not have to be online 24/7, but it's frankly not worth it when the big cost is the GPU time and weights volume. I don't know of a way to make that better without sacrificing model quality more than I have to.
For a shitpost though, I think it'd totally worth it to pay that much. It's kinda hilarious and I feel like it makes for a decent display of how bad things could get if we go full "AI replaces writers" like some people seem to want for some reason I can't even begin to understand.
I still think it's funny that I have to explicitly tell people to not take financial advice from it, because if I didn't then they will.
This isn't the target user, but the boy's been using it at the soil bacteria lab he works in to do basecalling for a FAST5 data from a nanopore sequencer.
As far as I know, Fly uses Firecracker for their VMs. I've been following Firecracker for a while now (even using it in a project), and they don't support GPUs out of the box (and have no plan to support it [1]).
I'm curious to know how Fly figured their own GPU support with Firecracker. In the past they had some very detailed technical posts on how they achieved certain things, so I'm hoping we'll see one on their GPU support in the future!
There has been weirdly little discussion on HN about Cloud Hypervisor.
I guess because it's such a horribly bland non-descriptive Enterprise Naming name?
It looks pretty sweet. Rust & sharing libraries with Firecracker and ChromeOS's crosvm, with more emphasis on long-running stateful services than in Firecracker.
For anyone else wanting to check on the status of this: it seems they're looking at a combination of seccomp, landlock and a systemd service instance per VM, with systemd doing DynamicUser, namespacing, and initial seccomp. Work seems to be happening right now, but of course it's telling and sad that it wasn't part of the original design.
Way simpler than what I was expecting! Any notes to share about Cloud Hypervisor vs Firecracker operationally? I'm assuming the bulkier Cloud Hypervisor doesn't matter much compared to the latency of most GPU workloads.
They are operationally pretty much identical. In both cases, we drive them through a wrapper API server that's part of our orchestrator. Building the cloud-hypervisor wrapper took me all of about 2 hours.
It’s cool to see that they can handle scaling down to zero. Especially for working on experimental sites that don’t have the users to justify even modest server costs.
I would love an example on how much time a request charges. Obviously it will vary, but is it 2 seconds or “minimum 60 seconds per spin up”?
We charge from the time you boot a machine until it stops. There's no enforced minimum, but in general it's difficult to get much out of a machine in less than 5 seconds. For GPU machines, depending on data size for whatever is going into GPU memory, it could need 30s of runtime to be useful.
It takes about 7s to load a 9GB model on Beam (they claim, and tested as about right), I imagine it is similar with Fly - I've not had any performance issues with Fly.
I see the whisper transcription article. Is there an easy way to limit it to, say $100 worth of transcription a month and then stop till next month? I want to transcribe a bunch of speeches but I want to spread the cost over time
I can probably just run these through whisper locally for you if you want and are able to share. Email is in my bio (ignore the pricing, I'm obv not charging)
So just to confirm, for these workloads, it’d start a machine when the request comes in, and then shut it down immediately after the request is finished (with some 30-60s in between I suppose)? Is there some way to keep it up if additional requests are in the queue?
Unfortunately true. Also jumped the fly.io ship after initial high excitement for their offering. Moved back to DigitalOcean's app platform. A bit more config effort, significantly pricier, but we need stability on production. Can't have my customers call me b/c of service interruption.
+1 - It's the most unreliable hosting service I've ever used in my life with "nice looking" packaging. There were frequently multiple things broken at same time, status page would always be green while my meetings and weekends were ruined. Software can be broken but Fly handles incidents with unprofessional, immature attitude. Basically you pay 10x more money for an unreliable service that just looks "nice". I'm paying 4x less to much better hardware with Hetzner + Kamal; it works reliably, pricing is predictable, I don't pay 25% more for the same usage next month.
Comments like these are just sad to see on HN. It is not constructive. What is these basic features that need fixing you're speaking about and what is the fixes required?
Reliability and support. Having even “the entire node went down” tickets get an auto-response to “please go fuck off into the community forum” is insane. What is the community forum gonna do about your reliability issues? I can get a 4€/mo server at Hetzner and have actual people in the datacenter respond to my technical inquiries within minutes.
Well I use Hetzner myself, I have never used Fly or their support. But this comment is much more helpful for readers so it's good that you elaborated :)
About Fly but not about the GPU announcement, I wish they had a S3 replacement, they suggest a GNU Affero project that is a dealbreaker for any business, needing to leave Fly to store user assets was a dealbreaker for us to use Fly on our next project, sad cause I love the simplicity, the value for money, the built in VPN.
> I wish they had a S3 replacement, they suggest a GNU Affero project that is a dealbreaker for any business
AGPL does not mean you have to share everything you've built atop a service, just everything you've linked to it and any changes you've made to it. If you're accessing an S3-like service using only an HTTPS API, that isn't going to make your code subject to the AGPL.
Yep, our lawyers say to not use and we have to check components and libs we use too. People are really shooting themselves in the foot with that license.
You assume that people want you to use their project. For MinIO, the AGPL seems to be a way to get people into their ecosystem so they can sell exceptions. Others might want you to contribute code back.
I have no problem with contributing back: we do that all the time on MIT / BSD projects even if we don't have to. AGPL just restricts the use-cases and (apparently) there is limited legal precedence in my region to see if we don't have to give away everything that's not even related but uses it, so the lawyers (I am not a lawyer, so I cannot provide more details) say to avoid it completely. Just to be safe. And I am sure it hurts a lot of projects... There are many modern projects that are the same thing, but they don't share code because the code is agpl.
Sounds more like the license is doing its job as intended, and businesses that can afford lawyers but not bespoke licenses are shooting themselves in the foot with that policy
Exactly. If someone doesn't want to use your software because the copyleft license is stricter than they would prefer, that's an opportunity to sell them a license.
> AGPL does not mean you have to share everything you've built atop a service, just everything you've linked to it and any changes you've made to it. If you're accessing an S3-like service using only an HTTPS API, that isn't going to make your code subject to the AGPL.
I am not so sure about that. Otherwise, you could trivially get around the AGPL by using https services to launder your proprietary changes.
There is not enough caselaw to say how a case that used only http services provided by AGPL to run a proprietary service would turn out, and it is not worth betting your business on it.
> > AGPL does not mean you have to share everything you've built atop a service, just everything you've linked to it and any changes you've made to it. If you're accessing an S3-like service using only an HTTPS API, that isn't going to make your code subject to the AGPL.
Correct, this is a known caveat, that's also covered a bit more in the GNU article about the AGPL when discussing Software as a Service Substitutes, ref: https://www.gnu.org/licenses/why-affero-gpl.html.en
Anything “clever” in a legal sense is a red flag for me… Computer people tend to think of the law as a black and white set of rules, but it is and it isn’t. It’s interpreted by people and “one clever trick” doesn’t sound like something I’d put a lot of faith in. Intent can matter a lot.
> Computer people tend to think of the law as a black and white set of rules
I've never seen someone put this into words, but it makes a lot of sense. I mean, idealistically computers are deterministic, whereas the law is not (by design), yet there exists many parallels between the two. For instance, the lawbook has strong parallels to the documentation for software. So it makes sense why programmers might assume the law is also mostly deterministic, even if this is false
I'm an engineer with a passing interest in the law. I've frequently had to explain to otherwise smart and capable people that their one weird trick will just get them a contempt charge.
Even if that wasn't directly targeted at me, I'll elaborate on my concern:
That it's possible to interpret the AGPL both ways (that the prior hack is legal, and that it is not), and that the project author could very well believe either one, suggests to me that the AGPL's terms aren't rigidly binding, but ultimately a kind of "don't do what the author thinks the license says, whatever that is".
Who is the target market for this? Small/unproven apps that need to run some AI model, but won't/can't use hosted offerings by the literally dozens of race-to-zero startups offering OSS models?
We run plenty of our own models and hardware, so I get wanting to have control over the metal. I'm just trying to figure out who this is targeted at.
We have some ideas but there's no clear answer yet. Probably people building hosting platforms. Maybe not obvious hosting platforms, but hosting platforms.
Fly is an edge network - in theory, if your GPUs are next to your servers and your servers are next to your users, your app will be very fast, as highlighted in the article. In practice this might not matter much since inference takes a long time anyway.
We're really a couple things; the edge stuff was where we got started in 2020, but "fast booting VMs" is just as important to us now, and that's something that's useful whether or not you're doing edge stuff.
- having the GPU compute in the same data center or at least from the same cloud provider can be a huge plus
- it's not that rare for various providers we have tried out to run out of available A100 GPUs, even with large providers we had issues like that multiple times (less an issue if you aren't locked to specific regions)
- not all providers provide a usable scale down to zero "on demand" model, idk. how well it works with fly long term but that could be another point
- race-to-zero startups have the tendency to not last, it's kind by design from a 100 of them just a very few survive
- if you are already on fly and write a non-public tech demo which just gets evaluated a few times their GPU offering can act like a default don't think much about it solution (through you using e.g. Huggingface services would be often more likely)
- A lot of companies can't run their own hardware for various reasons, at best they can rent a rack in another Datacenter but for small use use-cases this isn't always worth it. Similar there are use cases which do might A100s but only run them rarely (e.g. on weekly analytics data). Potentially less then 1h/w in which case race-to-zero pricing might not look interesting at all
To sum up I think there are many small reasons why some companies, not just startups, might have interest in fly GPUs, especially if they are already on fly. But there is no single "that's why" argument, especially if you are already deploying to another cloud.
It's not like Fly has GPUs in every PoP...so there goes all the same datacenter stuff (unless you just want to be in the PoP with GPUs in which case...)
But none of this answers my question.
I'm trying to understand the intersection of things like "people who need GPU compute" and "people who need to scale down to zero".
I am not seeing any race-to-zero in the hosted offering space. Most charge multiples of what you would pay on GCP, and the public prices on GCP are already several times what you would pay as an enterprise customer.
I don't know what you think I'm talking about, or who is charging multiples of GCP? But I'm talking about hosted inference, where many startups are offering Mistral models cheaper than Mistral are.
The recipe example or any any LLM use case seems like a very poor way of highlighting “inference at the edge” given the extra few hundred ms round trip won’t matter.
The better use case is obviously voice assistant at the edge. As in voice 2 text 2 search/GPT 2 voice generated response. That is where ms matter but it is also a high abuse angle no one wants to associate with just yet. My guess is they are going to do this in another post, and if so they should make their own perplexity style online-gpt. For now they just wanted to see what else people can think up by making the introduction of it boring.
> $2.24/hour pricing is for a 3-year commitment. On-demand pricing for H100 is $5.95/hour under our special promo price.\$1.15/hour pricing is for a 3-year commitment.
Company I work for had multiple times problems of not being able to allocate any gpus from some larger cloud providers (with the region restrictions we have, which still include all of EU as regions).
(I'm not sure which of them it was, we are currently evaluating multiple providers and I'm not really involved in that process.)
I don't want to deploy an app, I just want to play around with LLMs and don't want to go out and buy an expensive PC with a highend GPU just now. Is Fly.io a good way to go? What about alternatives?
Use https://vast.ai and rent a machine for as long as you need (minutes, hours, days). You pick the OS image, and you get a root shell to play with. An RTX 4090 currently costs $0.50 per hour. It literally took me less than 15 minutes to sign up for the first time a few weeks ago.
For comparison, the first time experience on Amazon EC2 is much worse. I had tried to get a GPU instance on EC2 but couldn't reserve it (cryptic error message). Then I realized as a first-time EC2 user my default quota simply doesn't allow any GPU instances. After contacting support and waiting 4-5 days I eventually got a response my quota was increased, but I still can't launch a GPU instance... apparently my quota is still zero. At this point I gave up and found vast.ai. I don't know if Amazon realizes how FRUSTRATING their useless default quotas are for first-time EC2 users.
Pretty much had the same experience with EC2 GPUs. No permission, had to contact support. Got permission a day later. I wanted to run on A100 ($30/hour, 8GPU minimum) but they were out of them that night. I tried again next day, same thing. So I gave up and used RunPod.io.
You might actually be better off building a gaming rig and using that. The datacenter GPUs are silly expensive, because this is how NVIDIA price discriminates. The consumer, game GPUs work really well and you can buy them for almost as cheap as you can lease datacenter ones.
I can recommend runpod.io after a few months of usage - very easy to spin up different GPU configurations for testing and the pricing is simple and transparent. Using TheBloke docker images you can get most local models up and running in a few minutes.
Paperspace is a great way to go for this. You can start by just using their notebook product (similar to Collab), and you get to pick which type of machine/GPU it runs on. Once you have the code you want to run, you can rent machines on demand:
I used paperspace for a while. Pretty cheap for mid tier gpu access (A6000 for example). There were a few things that annoyed me though. For one, I couldn’t access free GPUs with my team account. So I ended up quitting and buying a 4090 lol.
Great, more competition for the price-gouging platforms like Replicate and Modal is needed. As always with these, I would be curious about the cold-start time -- are you doing anything smart about being able to start (load models into VRAM) quickly? Most platforms that I tested are completely naive in their implementation, often downloading the docker image just-in-time instead of having it ready to be deployed on multiple machines.
gpu-friendly base images tend to be larger (1-3g+) so that takes time (30s - 2m range) to create a new Machine (vm).
Then there’s “spin up time” of your software - downloading model files adds as long as it takes to download GB of model files.
Models (and pip dependencies!) can generally be “cached” if you (re)use volumes.
Attaching volumes to gpu machines dynamically created via the API takes a bit of management on your end (in that you’d need to keep track of your volumes, what region they’re in, and what to do if you need more volumes than you have)
Interesting. We have this discussing this kind of services (offloading training) over the last several days [1] [2] [3]. Thinking on the opportunity to compete with top cloud services such as Google Cloud, AWS, and Azure.
Is there any configuration to keep alive the machine for X seconds after a request has been served, instead of scaling down to zero immediately? I couldn't find it skimming the docs
Machines are both dumber and more powerful than you'd think. Scaling down means just exit(0) if you have the right restart policy set. So you can implement any kind of keep-warm logic you want.
Can fly run cog files like replicate uses? Would be nice to take those pre packaged models run them here with the same prediction API
Maybe cos it's replicate they might be hesitant to adopt it but it does seem to make things a lot smoother
Even with lambalabs' lambdastack I still hit cuda hell
https://github.com/replicate/cog
They're a "real" cloud provider (with their own hardware) and not a reseller like Vercel and Netlify. So this isn't _that_ surprising. AWS economies of scale do allow them to make certain services cheap but only if they choose. A lot of time they choose to make money!
AWS is one of the most expensive infrastructure providers out there (especially anything beyond the "basic" services like EC2). And even though AWS still has some globally-notable uptime issues, "nobody ever got fired for picking AWS".
I mean from hearsay of people which had to work with AWS & Google Cloud & Microsoft Azure it seems to me that the other two are in practice worse to a point they would always pick AWS over them even through they hate the AWS UX.
And if it's the best of the big 3 providers, then it can't that bad, right ..... right? /s
As a person working in a startup which used AWS for a while:
*AWS is expensive, always, except if magic*
Where magic means very clever optimizations (often deeply affecting your project architecture/code design) which require the right amount of knowledge/insights into a very confusing UI/UX and enough time evaluate all aspects. I.e. it might simple not be viable for startups and is expensive in it's own way.
Through most cheaper alternatives have their own huge bag of issues.
Most important fly.io is their own cloud provider not just a more easy way to use AWS. I mean while I don't know if they have their own server centers in every region they do have their own servers.
AWS isn’t the cheapest so how is that a surprise? They are a business and know how to turn the right knobs to increase cash flow. GPUs for AI is one major knob right now.
This is amazing and it shows that Nvidia should be the most valuable stock in the world. Every company, country, city, town, village, large enterprise, medium and small business, AI bro, Crypto bro, gamer bro, big tech, small tech, old tech, new tech, and start up want Nvidia GPUs. Nvidia GPUs will become the new green oil of the 21st century. I am all in and nothing short of a margin call will change my mind.
