A concrete example: video upload and processing. Your frontend handles uploads, then you can offload conversion to AWS while you continue to handle other requests.
Seems like a poor example - if you're building a product, you can easily afford the couple hours it takes to set up an EC2 image and autoscaling, dump a work item in SQS, and pick it up on an EC2 spot instance. And if you're doing video processing, you really really want to use a more efficient language than Javascript (like C) to handle the video processing. Combine the two of these and you'll get roughly a 100x cost saving over dumping a JS function into Amazon Lambda.
I see this being most useful if you have a one-off analytic you need to write against some big data in S3 or RDS. For one-off scripts dealing with the raw AWS APIs is just useless overhead, and the expense of running the script will be negligible.
Having done this a few times from scratch for various companies, there a ton of moving pieces for almost any processing pipeline. Being able to scale that pipeline without writing the ops code to make it happen is actually magical. I'm not saying everyone should jump out and use this, but it takes a lot of work to:
(a) measure each of the points of your service
(b) deploy your code in an automated manner
(c) deploy your monitoring in an automated manner
(d) make sure your code is under supervision
(e) setup alerting on the monitoring
(f) scale up / down and within price constraints as needed
(g) repeat this for all supporting services (queue, db, etc)
(h) write your actual application code
The potential to handle certain classes of problems via SQS/SNS/S3 pipelines is pretty alluring. You still have to do configuration, but the bet is that the configuration necessary for the SQS/SNS/S3/Lambda pipeline is far lower than that necessary to setup random autoscaling Celery, Resque, or random JMS/AMQP system on top of Ubuntu with Chef/Puppet/whatever.
Cool! As someone with more experience than me in this, would you mind responding quickly to these points? I will give my personal opinions but if you can trump with more info that would be cool too. :)
1. I agree that JMS sounds like a hassle but is that really necessary? I would think that you can batch process data on an EC2 instance, then pick it up in your local code directly using AWS APIs... not sure.
2. I am not so familiar with the Lambda system but I'm also not sure how it would scale db as necessary (item "g" in your list) thus overall processing time would still be bottlenecked by other resources (database IO, for example), no?
I agree with your points but in all these cloud-compute scenarios I always wonder "Are we trying to reach a theoretical limit of fastest-possible computation, or just reach some reasonable saturation point close to the natural bottlenecks/throttles of our system integrations?".
3. Having been burned a few times now by over-optimizing when considering cloud I would probably now first consider just picking a slightly oversized EC2 instance and throwing some high-performing code onto it (Java, C++). Dynamic languages + auto-scalable resources (though I'm talking about web hosting in particular now) seems to drain clients wallets more than anything. At this point I'd actually recommend anyone with new web infrastructure to just buy a static instance and write optimized Java rather than trying their hand at auto-scaling Ruby/Python/Node. Do you notice a similar issue with your clients regarding code optimization vs. auto-scaling?
This is a cool thought exercise I suppose, the idea of within a program throwing a particular function to an offsite parallel-compute engine. I just imagine it will complicate platform integration & bleed money given that it is in js (plus API headaches abound, as you mention)... If the same concept of dynamic provision/deploy/process/collect could all be done from a Java/C++ app I'd be very intrigued, but I suspect it is already near possible or achieved trivially through AWS APIs, albeit it probably with a manual deploy involved.
I think I remember this concept in Matlab back from when I did some research in grad school -- basically an instance of Matlab can be setup as a compute server, and the parallel processing functions of Matlab code on other computers can portion out work to it. This is the ideal model in my mind.
1. Write high-performance code in any language with some function that should be happening remotely in parallel.
2. Configure AWS to auto-provision the resources necessary.
3. Execute code, have it behave as if it is all running locally.
Really all these things can already be achieved with SOA, RMI, message queues, etc., the trick is just in making it transparent to the programmer so there is no deploy step involved. With the right spec it could even become platform agnostic (change small config file somewhere to target different cloud platform... would be nice to see a JSR about that in the near future!).
This is largely the architecture behind MapReduce or Hadoop Streaming. Write high-performance code in any language with an isolated parallelizable function, configure some cloud to auto-provision workers that run that function repeatedly on millions of records, execute code & pretend it's local.
Combine the two of these and you'll get roughly a 100x cost saving over dumping a JS function into Amazon Lambda.
And then move it off EC2 onto dedicated hardware and you'll see another ~30x cost savings.
Running a permanent transcode cluster on EC2 would be rather insane.
Hetzner rents you i7's for $50 per month, the EC2 equivalent (c3.8xlarge) costs $40 per day.
Yes you can cut EC2 cost with spot-instances, but at least in our case that would still have been significantly more expensive than just renting some scrap metal.
If you need cheap, disposable compute for semi-predictable loads then the Hetzner flea market (yes, they really have one!) is hard to beat on bogomips per dollar.
Yeah, dedicated is definitely the way to go if you can afford the ops staff. Actually, if you're really big and can afford the hardware engineers, building your own DCs and computers is the way to go. I've seen the profit margins of some major cloud providers; they're insane. About the only industry more profitable are the telecoms.
I was targeting my comment toward a startup that'd likely be building a product like the OP suggested, though, with maybe a dozen devs who are all generalists. Setting up a bunch of EC2 instances to pull videos out of SQS/S3 and run them through ffmpeg is something an ordinary full-stack developer can do in a morning, and scaling it up just involves clicking a button. Running and scaling a dedicated server farm reliably usually needs a dedicated ops person to keep it all working.
With the additional cost that you actually need to upload the raw video, which you don't need if you just do it locally without Lambda. Or do I misunderstand this?
I see this being most useful if you have a one-off analytic you need to write against some big data in S3 or RDS. For one-off scripts dealing with the raw AWS APIs is just useless overhead, and the expense of running the script will be negligible.