Comment by ac360 | Hacker News Reader

Comment by ac360 | Hacker News Reader

Of interest:

The following real life scenario below portrays an administrative API layer that runs in EC2, and how it compares to what the raw hardware costs would be in Lambda. It purposefully excludes anything to do with requests (CloudFront+ELB for EC2 and API Gateway for Labmda).

EC2 scenario: 2 c3.large instances, general purpose 15GB SSD (need two for HA, span AZs) 1 Yr All upfront reserved pricing on both instances

Lambda scenario: 1042 memory size Workload running on EC2: 16000 requests a day to an endpt. 200 ms ave = 3200000 ms total

3200000 / 100 ms = 32000 segments of 100 ms

EC2 Cost: 1084(yr)/365 = $2.96986301369863/day

Lambda Cost: 32000 * 0.000001667 = $0.053344/day

The cost difference is so large because this workload is not a very high EC2 utilization scenario. We need extra servers for multi AZ HA and to handle bursts. Lambda handles both of these for you.

This also does not include the huge costs savings of no server management (security patches etc) that go into EC2 or Docker containers.

Using two c3.large instances to serve 16,000 requests a day (~11 requests a minute) is akin to using a muscle car to commute a quarter mile - you're vastly over-provisioned for such a simple workload.

This kind of workload could easily be handled by a pair of t2.micro instances. This would cut your costs for dedicated servers to around $0.40 a day.

Plus, you could then take advantage of persistent in-process memory caches and warm services, helping to speed up your responses even further.

Using tools like Lambda come with a number of advantages, but you give up a lot of control at the same time. Perhaps this isn't a problem for most side projects, but for business critical software I'd be very leery of handing that much control over to someone as uncommunicative as Amazon.

"The cost difference is so large because this workload is not a very high EC2 utilization scenario. We need extra servers for multi AZ HA and to handle bursts."

The $0.40 cost per day I calculated includes multi-AZ HA, and your bursts would have to be fairly significant to overwhelm two t2 instances which are accumulating CPU credits 95% of the time - say around 600 requests per minute for a few hours.

Of course, this is all conjecture since I can't actually view your workloads, but we've been performing this exercise for our own services, and the t2 series of servers is remarkably capable: especially given their cost. The original administrators thought (or were lead to believe) that we really needed the raw horsepower that c*.large+ instances offer... and they didn't in most cases.

Routers, crud API wrappers, even some disk persistance applications all qualify as "non-CPU intensive".

Reply on news.ycombinator.com