I'm not talking about on-prem but traditional hosted DC infrastructure migrating to AWS which is the market I sell myself to. On-prem is usually a mishmash of MSP managed and semi-competent admin crud that needs consolidation rather than stacks of kit. The workload impact profile is usually entirely different on AWS compared to infrastructure you built yourself which is where cost models fall flat on their arses. The last platform I moved back was 150k requests/second HTTP and they couldn't get the 90th percentile latency back down below 700ms on AWS because of the sheer shitty price/performance ratio. Back to a couple of stacked C7000s it went.
Incidentally my certification was a stab at the certification process which was so easy it is pointless. Also the trick on the market is to get someone else to pay for the training which you mostly just sit on ebay through buying test gear then pay for the exam then leverage that to increase the daily rate.
I came away more impressed with Azure after my 2 day visit to the Microsoft Tech Center, as were our software architects and machine learning guys.
Some items of note I found superior to AWS:
AutoML: Will test run your dataset against about 100 different deep learning models/algorithms to find the model with the best convergence.
Cosmos DB: More friendly to database admins and developers than DynamoDB, with more features (like SQL-like query available)
https://db-engines.com/en/system/Amazon+DynamoDB%3BMicrosoft+Azure+Cosmos+DB%3BMongoDBBot Service: Easily integrates with Azure cognitive services (voice to text, image ID/classification, faceID, translation and language detection, etc); build just one bot and deploy it automatically to webchat, Facebook, Line, Skype, Teams, Kik, Slack, etc. No customization for each channel needed!
https://dev.botframework.com/
Microsoft and Amazon are both very good at impressing people and the whole point of why they have you at their tech breifing centers is to impress you. They will be very pleased you have drunk the cool-aid.
Its horses for courses. I really like using Lambdas, and the API gateway. mostly what i'm doing is small stuff. What does concern me is that it locks me in too one particular vendor. Could i build what i'm doing with the same level of resilance and security for the price.
Your latter point is what worries me. Once you’ve glued everything into lambda you’re SOL if you want to move away. I only recommend the IaaS and portable services when I architect anything for the cloud.
Microsoft and Amazon are both very good at impressing people and the whole point of why they have you at their tech breifing centers is to impress you. They will be very pleased you have drunk the cool-aid.
Yes agreed... a thorough POC testing is needed to vet out the truth.
Its horses for courses. I really like using Lambdas, and the API gateway. mostly what i'm doing is small stuff. What does concern me is that it locks me in too one particular vendor. Could i build what i'm doing with the same level of resilance and security for the price.
Lambda's are fantastic, especially the way in which it integrates with all other AWS services. I thought about using open-source serverless in a docker container within EC2 for portability to other clouds or on-prem... but why? It'll just cost more and lacks seamless integration with all the other AWS services.
Forgot to mention Azure IoT and IoT Edge... seems to be more refined than AWS IoT/Greengrass. Certainly seems to have better acceptance from manufacturers, at least the ones I know about...
Your latter point is what worries me. Once you’ve glued everything into lambda you’re SOL if you want to move away. I only recommend the IaaS and portable services when I architect anything for the cloud.
Its ultra-cheap... but yes you get locked in...
its the same story.... pick 2 (cheap, transportable/resilient, performant)
It’s not even cheap though. The price curve is pretty flat and then goes straight up your butt.
I’ve seen a couple of startups nearly hose themselves with a finger in the air cost estimate.
Some items of note I found superior to AWS:
AutoML: Will test run your dataset against about 100 different deep learning models/algorithms to find the model with the best convergence.
You could read that also as "you'll train our AIs for free".
All the AI offerings are a joke so I wouldn’t worry. I know a big fintech that went through them all spending millions for basic risk management stuff and came to the conclusion a pigeon pecking on a yes/no button had the same outcome quality.
It’s not even cheap though. The price curve is pretty flat and then goes straight up your butt.
I’ve seen a couple of startups nearly hose themselves with a finger in the air cost estimate.
Well, yes overall its not going to be cheap, since you'll need miscellaneous services other than just lambda... our spend on lambda is less than 1% the cost of our spend on EC2, which is amazing considering how much we use lambda...
All the AI offerings are a joke so I wouldn’t worry. I know a big fintech that went through them all spending millions for basic risk management stuff and came to the conclusion a pigeon pecking on a yes/no button had the same outcome quality.
The canned AI model offerings from cloud providers are too generalized, so they won't give good results on everyone's specific datasets.
Our machine learning team regularly avoids using cloud services for any model training/tesing due to high cost.
Good guide for DIY deep learning setup here:
https://timdettmers.com/2019/04/03/which-gpu-for-deep-learning/
Your latter point is what worries me. Once you’ve glued everything into lambda you’re SOL if you want to move away. I only recommend the IaaS and portable services when I architect anything for the cloud.
To a point. The databases that the lambda connect to could be easily migrated to. MySQL = MySQL. The object store ( S3 ) coudl be done with something else. I'm using Python for my lambdas. I'd have to investigate what kind of API you could stick in front of them, to make them lambda like. At least for my use case, the costs are not crazy. I'm in the league of a few thousand requests per day, not 100,000/sec
I do have a large print of a sticker I saw hanging in my cube.
This one
There are no utilities, just someone else's generator.