- Sedky Aboushamalah
Have you ever wondered how much a full-stack serverless application costs; Especially at scale? Or when it starts to cost money?
When designing the initial prototype of Serv, I knew we had to use serverless architecture because of the success we found with Ajtima:
- fast developer velocity and ramping
- Easily automatable CICD pipelines
- no infrastructure to manage (duh)
but one of the other things I took for granted was the free tier of AWS; ie "How much can we get for free"? And that's what I'm here to tell you about.
What will you get out of this reading ?
- The awesome-ness of free tier providers such as AWS, Google Firebase, Netlify, etc
- The stack we used when building a multi-tenant restaurant ordering app
Let's continue using Serv as the subject of the case study. Serv is a platform that lets restaurant patrons scan a QR code, browse a menu, order items, and pay for them.
We'll run the numbers at 3 different scales:
Having been a solutions architect for a few years, I hate to continue to preface everything with the horrible phrase "it depends", but it is true; YMMV depending on what tools your app is using, and how your users interact with your app.
With that said, Here are the generic numbers specific to the Serv use case you would see at the end of this guide. Continue if you want to see this number broken down by each service in the stack:
|Users||Running Monthly Costs|
For 100 daily users, we don't pay a penny. We are completely leeching off of free SaaS platforms like AWS, GitHub, Netlify, and others.
For 10,000 daily users, or 300,000 unique monthly users & sessions, we pay $2.77 USD per MONTH. That is an absolutely, ridiculously low number, which only serves to highlight the absurd volumes that AWS is accustomed to dealing with. I mean, their pricing is laid out by the millions.
For 1 million unique daily users, or 30 million monthly users, it would only cost us $435.25 USD a month to support all of our infrastructure. How much does your app net you per user?
However, it's not all roses. It's like when Black Friday deals get you in the door to spend 10 bucks on a 20 dollar shirt, only to have you pay $1,000 for a full price TV while you're there. Vendor lock is diffiuclt to climb out of.
Each Service in the Stack
Serv has two interfaces. An end-user client for restaurant partons to place and pay for orders.
And a Dashboard view for restaurant staff to manage orders, menu, etc.
This is just React that gets compiled into HTML/JS/CSS. We can use any host for that. I've been using Netlify & GitHub for years, and never paid a penny (Thank-you Netlify & GitHub). Netlify has a direct integration with GitHub that allows you to build your dev/staging/production website on pushes and pull requests.
So far, our server bills are $0/month. Because Netlify doesn't charge for traffic, only for build hours in the CD pipeline. Yes, for
1,000,000 users of our front-end app, we don't pay a penny.
Back end API functions
The back-end is all in AWS, so slightly more difficult to calculate.
Serv uses the following:
- AWS Lambda functions
- AWS Appsync
AWS Lambda is invoked whenever somebody makes a payment, which is once a day per user. We use Lambda to avoid storing sensitive tokens in the front end.
1 daily user = 30 monthly requests per user
So our new running cost including Lambda
|Daily Users||Running Monthly Costs|
Wow. That's right. To serve 1 MM users per day, or 30 MM requests a month, Lambda costs us....
To serve 10,000 users a day, it costs us a whopping NOTHING.
AWS Appsync is a managed GraphQL API that sits on top of Lambda/DynamoDB/other data sources.
Their pricing page lays out what is included in the free tier:
250,000 query or data modification operations 250,000 real-time updates 600,000 connection-minutes
let's run the numbers. During a session ie single unique daily user:
- Average group of customers is 5 people at one table
- A user creates one order and receives 4 per session
- Users spends 5 minutes ordering and paying.
|Daily Users||Monthly Groups||Connection Minutes p/m||Query and Data Modification Ops p/m||Real-time Updates p/m||Total Cost|
|100||600||15,000 / $0.00||3,000 / $0.00||15,000 / $0.00||$0.00|
|10,000||60,000||1,500,000 / $0.072||300,000 / $0.20||1,500,000 / $2.50||$2.772 USD|
|1,000,000||6,000,000||150,000,000 / $11.952||30,000,000 / $119.00||150,000,000 / $299.50||$430.452|
Note, we didn't add our data transfer charges, but it's a small minority of the overall Appsync charges
Math things: (can skip)- Daily Groups = Daily Users / 5 - Connections Per group
= 5 mins * 5 users = 25 mins per group (4.00 per million operations) - Real Time Updates per group = 5 x 5
= 25 real time Updates per group ($2.00 per million updates)
Our running totals so far:
|Users||Running Monthly Costs|
Still really affordable. I mean,
10,000 users is a whole community. And to support that across all our front-ends & APIs, we barely have to pay more than a shot of espresso PER MONTH.
1,000,000 unique daily users, it costs us
$0.000435 per user. It's easy to come out ahead in this age of PAYG.
Let's continue with an example that really hurts the bank this time.
Text-Messaging on AWS SNS
We (initially) decided to use Phone number One-Time-Code sign up & sign in to make the ordering process as simple as possible. We used an awesome AWS Cloud Formation that spun up all the necessary lambdas and SES topics.
Each unique user will enter their phone number and receive an SMS verification code once per session.
The price of each text message sent by AWS Pinpoint/SNS is
Oh and hey, we've left the free tier! Pinpoint is not included.
t = Daily Users t X 30 X 0.00581 = AWS SNS cost per month
|Daily Users||Monthly Cost|
Whoa, what happened? Our costs just increased from leech level to beyond Enterprise.
Are we screwed? Even if we look at the industry-leading Twilio's pricing, we see that SMS is expensive. Maybe we can use email confirmation, or we can roll our own tools; We have other options.
But that's beside the point; The point is to highlight that (unlike Serv) you need to be careful, and can't expect all services to be affordable or you will be surprised month end.
What if you need to use a non-managed service or have your own APIs to run? We can run these workloads on EC2, FarGate, EKS, or one of the other compute services.
For Serv, we ran Tyk APIM and tried a few different compute services, but ended up settling on EC2 for the affordability.
This one is much more complex to calculate, as you have to figure out how many nodes you need at various RPS benchmarks. Here's a guide I wrote for Tyk to help us do that.
Long story short, it is extremely affordable or FREE to bootstrap your next application on AWS. As a matter of fact, we built Ajtima on Google's FireBase, which is another platform to run full-stack applications on, and found it equally affordable.
Does that mean it's free across the board? No. Some tools are more expensive than others. But all-in-all, a good developer picks the right tool for the job anyway.
Finally, is this blog all-encompassing? No, of course not. For example, I left out data storage costs, and DynamoDB costs which were behind the Appsync APIs. But they were besides the point; My point was to illustrate the generosity of PAYG Cloud platforms, not get into the weeds of each service.
Would you want to see more in-depth technical architecture diagrams to showcase each component in our tech stack or how it all fits together? Let me know, drop your comments below or reach me via contact options in my footer to start a conversation!