Man, that makes no sense to me if it's just aws cost -- very basic back of the envelop calculating assuming you had $500k to spend per year on cpus on aws:
A c5d.2xl (8 cores) on aws is about $1875 per year reserved, $500k per year would buy you over 2100 cores; if you only did 100 req/s per core, you'd be able to handle something like 213k req/s average.
Just the alb costs on lambda at this request rate would cost you $10MM annually... 20x the cost of doing it on VMs.
>>> 500e3 / (1875.0/8) * 100 * 60*60*24*365 / 1e6 * 1.51
10158796.8
$500k to spend
$1875 per box / 8 cores per box
100 req/s
60*60*24*365 seconds in a year
1e6 requests per $1.51 of alb spend
Staggering the amount of work you can do with just the amount of money you saved and it seems impossible to me that lambda was able to account for this?
I'll color the picture in a little more, maybe that will help make sense of it.
We have a unique workload that includes very large bursts of work that need to be worked on ASAP. Spinning up instances on demand is not fast enough, so we had a lot of servers sitting idle. Lambda allows us to execute this same work for much cheaper because we aren't paying for idle CPUs. Lambda also allows us to execute at much higher concurrency than ec2 would, for the cost. 15000 Threads spread across 100's of IPs is easy and cheap on lambda, not so much on EC2.
We went from ~250 customers per cluster(our term for a set of dedicated workload servers) to 2600(haven't pushed it past there yet). The work that was done on those instances is slowly being moved to Lambda and all new work/features are added to Lambda directly. This has effectively allowed us to drop a few clusters and also move to lower resources per instance.
FWIW we don't use ELB for Lambda, they come in via API Gateway, to Lambda/Step Functions directly or into an SQS queue in the case of workloads that have rate limits. For Lambda based workloads it breaks down something like 56% of cost is Lambda, 34% of cost is Step Functions and 9% is API Gateway.
A c5d.2xl (8 cores) on aws is about $1875 per year reserved, $500k per year would buy you over 2100 cores; if you only did 100 req/s per core, you'd be able to handle something like 213k req/s average.
Just the alb costs on lambda at this request rate would cost you $10MM annually... 20x the cost of doing it on VMs.
Staggering the amount of work you can do with just the amount of money you saved and it seems impossible to me that lambda was able to account for this?