5 ms·

I getting it from Jo Beda, the CTO of heptio (now VMWARE). More specifically , from his definition of cloud native. A cloud native system is a system that mana

I getting it from Jo Beda, the CTO of heptio (now VMWARE). More specifically , from his definition of cloud native. A cloud native system is a system that manage its resources via an API. The vendor locking in this case is the cloud API. I.e. if I want a resource (e.g. a VM) , I need to use the EC2 api. With kubernetes, I write to the kuberentes API which is open source, cloud vendor natural. However, and this the key to my comments - kubernetes is the platform of platforms. It is an API platform and not just an API. In the future (some would say near future), all of the cloud services will be offered as kubernetes APIs (via a set of domain specific CRD). I am currently working on the AI platform. As for the big machines comments. I absolutely agree. However, I would like to ignore the actual hardware (banning some sort of HA that would require at least 2 machines).

ethbro 7y ago
> In the future (some would say near future), all of the cloud services will be offered as kubernetes APIs (via a set of domain specific CRD) If (big if) some standardized variant of kubernetes "wins." K8s was transparently Google's strategic shot at forcibly commodifying AWS. But everyone has their own agenda. Google, Microsoft, and (to a lesser extent) Amazon could survive in a world of fully commodified compute. Give their compute assets are essentially surpluses from running their other businesses (advertising, software sales, and online retail). So none of them have a strategic imperative to derail k8 evolution. But managed, specialized cloud services are lucrative to all of them (especially Amazon). So their best-case isn't a fully commodified future.
- dodobirdlord 7y ago
  > Give their compute assets are essentially surpluses from running their other businesses (advertising, software sales, and online retail). This is mythology. It is mostly not actually true.
  - ethbro 7y ago
    Source? Do any of the major clouds publish their base, first-party load?
    - dodobirdlord 7y ago
      Consider the ever-expanding region lists of cloud providers, and it becomes clear that the vast majority of cloud infrastructure was purpose-built to support their cloud offerings.
      - ethbro 7y ago
        Yet at the same time, expanding their own use. I doubt very much the first deployment of Gmail looked anything like its footprint today. More regions doesn't necessarily stem solely from customer demand. I expect the vast majority does, but without numbers, that's just a guess.
BaronSamedi 7y ago
That seems to be trading one type of lock-in for another. The problem with that approach is that now you have added a highly complex layer into your architecture that you need to maintain, integrate, and secure. I think it more likely that in AI, Amazon and Microsoft will continue to add more powerful cloud services making home-grown kubernetes solutions even less relevant.
- streetcat1 7y ago
  So, in any case you will have the complexity. With CRD and operators you might be able to automate it away. An AI service is much more than submitting a CSV file to a cloud and getting a model back, or even a web end point. You need: 1) watch for changes in the input data. 2) watch for model decay. 3) watch for latency. 4) retrain (preferably automatically). 5) redesign your pipeline and imputation methods when the data changes. 6) watch for problems in the input data during serving 7) Scale out or Scale down. 8) Be optimal with the training resources (for example, train on spot instances). 9) Canary and shadow release of model 10) Data and model versioning. Not to mention privacy and price. What happen today is that there are no offering of alternative to the cloud services, so there is no point of reference.

ethbro 7y ago

> In the future (some would say near future), all of the cloud services will be offered as kubernetes APIs (via a set of domain specific CRD) If (big if) some standardized variant of kubernetes "wins." K8s was transparently Google's strategic shot at forcibly commodifying AWS. But everyone has their own agenda. Google, Microsoft, and (to a lesser extent) Amazon could survive in a world of fully commodified compute. Give their compute assets are essentially surpluses from running their other businesses (advertising, software sales, and online retail). So none of them have a strategic imperative to derail k8 evolution. But managed, specialized cloud services are lucrative to all of them (especially Amazon). So their best-case isn't a fully commodified future.

dodobirdlord 7y ago
> Give their compute assets are essentially surpluses from running their other businesses (advertising, software sales, and online retail). This is mythology. It is mostly not actually true.
- ethbro 7y ago
  Source? Do any of the major clouds publish their base, first-party load?
  - dodobirdlord 7y ago
    Consider the ever-expanding region lists of cloud providers, and it becomes clear that the vast majority of cloud infrastructure was purpose-built to support their cloud offerings.
    - ethbro 7y ago
      Yet at the same time, expanding their own use. I doubt very much the first deployment of Gmail looked anything like its footprint today. More regions doesn't necessarily stem solely from customer demand. I expect the vast majority does, but without numbers, that's just a guess.

BaronSamedi 7y ago

That seems to be trading one type of lock-in for another. The problem with that approach is that now you have added a highly complex layer into your architecture that you need to maintain, integrate, and secure. I think it more likely that in AI, Amazon and Microsoft will continue to add more powerful cloud services making home-grown kubernetes solutions even less relevant.

streetcat1 7y ago
So, in any case you will have the complexity. With CRD and operators you might be able to automate it away. An AI service is much more than submitting a CSV file to a cloud and getting a model back, or even a web end point. You need: 1) watch for changes in the input data. 2) watch for model decay. 3) watch for latency. 4) retrain (preferably automatically). 5) redesign your pipeline and imputation methods when the data changes. 6) watch for problems in the input data during serving 7) Scale out or Scale down. 8) Be optimal with the training resources (for example, train on spot instances). 9) Canary and shadow release of model 10) Data and model versioning. Not to mention privacy and price. What happen today is that there are no offering of alternative to the cloud services, so there is no point of reference.