Comment by vachina
Comment by vachina a day ago
The consumers are domestic EU so you don’t really need the reach and availability of the big 3.
Comment by vachina a day ago
The consumers are domestic EU so you don’t really need the reach and availability of the big 3.
You go multi-region. Multi-cloud is extremely expensive, both in terms of data and functional equivalence.
Bare metal is pretty much the same story: you can host it at different providers, but scaling that and maintaining coherence between data centers is not an easy feat as it might sound.
And seriously now, no sane provider is willing to cover your losses if they go do down. On the other hand, it's not a secret this is not happening and you can take this into account in your risk management strategy.
After years and years, Amazon now has an offering to shield you from when us-east-1 goes down. Funny, no?
If you’re going to go multi-region and take the latency hit may as well go multi-provider no?
Multi-region within the same provider won’t shield you against unknown shared dependencies on a single point of failure (AWS console auth still relies on credentials being checked in a single region if I remember right).
And yes fully agreed that maintaining consistency between active-active regions (whether cloud or bare-metal) is super hard and not worth it for most deployments. Active-standby with point-in-time-recovery and an acceptable data loss window is much easier - when one region is confirmed down, someone throws a switch and the standby becomes active.
> no sane provider is willing to cover your losses
Agreed, but thats why all those who justify the 10-90x premium of the cloud over bare-metal are full of it - that premium is not actually worth it.
> it's not a secret this is not happening
Maybe for you it’s not a secret? Literally every thread tries to justify cloud reliability and their resulting markups. Well if it’s that reliable they’d put their money where their mouth is.
> If you’re going to go multi-region and take the latency hit may as well go multi-provider no?
No. If you go multi-region, you use the same tooling, same terraform modules and logic and so on. There's little plumbing needed to make it work. And latency wise this is not an issue in most cases, since most of the requests are covered by the CDN anyway. And you don't have to duplicate everything.
If you go multi-cloud you need to learn a whole new set of systems. And that is expensive. Both in terms of operating and people - because you will need more.
> Agreed, but thats why all those who justify the 10-90x premium of the cloud over bare-metal are full of it - that premium is not actually worth it.
You get charged a premium for convenience. And a high enough chance you don't have downtime.
> Maybe for you it’s not a secret? Literally every thread tries to justify cloud reliability and their resulting markups.
Cloud is reliable if you are willing to spend some money to benefit from that reliability and convenience.
---
Besides this, another thing where cloud saves you money is compliance. They have all the right attestations in place to make your audits go easy. If you self-host on bare metal, you're going to spend a lot of time to be compliant with various regulations. Maybe if you're a small company, you don't have that much compliance you need to do. But once you grow a little, those immutable Stackdriver logs are a godsend when you're asked to prove logs have not been tampered with.
> If you go multi-cloud you need to learn a whole new set of systems.
Isn't the whole point of Kubernetes that you don't need to do this? Also, if you want to know for certain that you aren't "vendor locked", running on two clouds is a constant test of that fact. That is to say, have a stack that can deploy to Kubernetes then have two clusters in separate clouds.
It's not just geographic regions around the globe.
It is also the wide array of services -- well integrated into their primitives of security, authentication, governance, monitoring and logging, etc
Is there a EU cloud provider that provides -- even if limited to EU geography -- the equivalent of Blob Storage + Azure Data Lake Storage + Azure Data Factory or Fabric + Microsoft Foundry with native access to OpenAI and Anthropic models?
Having used both worlds: a lot of the provided features come with strong vendor lock-in, and in most cases that not, with slightly stronger “local” engineering you can reach the same targets and needs locally.
The more I work (started coding 40 years ago, and data engineering 25 years ago), the more I favor designs that are less coupled to cloud features.
If you do so, the offering in the EU just as it is now is well enough to scale.
In short: more computer science, less delegating to cloud operators, stronger designs.
The irony is that EU education is still broader and more grounded in fundamentals, compared to US one that has become increasingly skills-oriented.
I also prefer to design solutions that are portable and platform independent, cloud providers simplify and hide something to you, it has a cost (not just money) that you cannot quantify on long term and that's clear for who has experience in both worlds.
This is sane advice but maintaining strong internal engineering and IT teams is not everone's cup of tea; even organizations gthat intend and try to do this cannot achieve it -- and furthermore the big cloud operators have spent millions of dollars spreading the gospel that the only correct way to survive and build is to migrate to cloud and use their lego building blocks.
A lot of the justification for moving away from own datacenters and heavy in-house engineering teams actually makes sense to organizations of many shapes and sizes. If the core business is not technology -- it is hard to stay invested in a in-house cloud-agnostic engineeringc capability.
The world has more or less accepted this reality and adopted the services like S3 and managed databases, and tight integration with the likes of Microsoft Entra/Purview/Sharepoint/O365 etc for ready-to-use business integrations. Prying organizations away from taht convenience is going to be hard. But the current environment cretaed due to lack of trust in US could the strongest motivator ever to push businesses and nations in that direction. I reckon it will be a long and painful process though.
Until two years ago, I did not need Google Dataflow as a very specific example. But then new business requirements came in and there were two options:
- develop something internally and support it
- use a cloud provider offering, fire it up and forget about it
The choice was pretty straightforward.
You can get all of that in the EU via scaleway, Ionos etc. for example.
I don't know what you mean by native access to frontier models. Who has native access to these frontier models?
Availability ain't worth shit unless the compensation for missing said availability is anywhere near the business losses caused by it. "Credit on your bill" doesn't count (and you're not even likely to get that since they can just lie on their status page and pretend everything is fine).
Cloud is convenient but don't expect any kind of availability you can actually rely on. If you actually need that, you're gonna have to go multi-cloud or self-managed bare-metal at multiple providers anyway.