Using Serverless for unpredictable traffic to for TV fundraiser event

Yeah, we literally used it as a stripe proxy and to serve frontend. All data was served in stripe and then queried using Stripe Sigma, their flavour of Presto.

Can tell you the total cost from RTE does Comic Relief, which was the same size, was about 16 euros to host the whole thing. This would be about the same, however is on RTE's cloudflare account, so not easy to get breakdown.

Workers were just a good fit not to have to host anything, due to the unpredictable number of donations that we would be getting through. Deployment and development is also really easy.

Spikey, unpredictable traffic, serverless is the way.

Sorry about the title.

>Terrible title for a fun little case study

I figured it was one of those AWS invoice shock stories from "serverless" and "millions of Euros".

When I saw the title I honestly was thinking that someone managed to convert millions of viewers into a million Euro AWS bill...

It's like night and day in my (small) experience. Cold start is VERY fast.

So Lambda & (L@Edge) is a general purpose computing system. They run a normal process in a programming language of your choice, then feed it event data, take the response out and continue the pipeline. They might keep the process running for reuse sometimes. If your process is being reused, it’s “warm” and works as fast your code. If it’s “cold”, your process start time, Lambda finding an execution slot, downloading your code there etc all add up. Good news is that if you’re doing multi core stuff or specialized code it’ll work fine (now including special Intel AV-something instructions).

Cloudflare workers took a different approach - they run V8 node-ish processes on all their edges, and you submit your code as JS or WASM that fits their interface (which is based on the service worker standard). They then execute it as a V8 isolate, which is blazing fast because the V8 process is already up and running, it’s just running JS/WASM code, which it knows how to do pretty well. Even the cold start is pretty small, because the JS/WASM tends to be pretty small and Cloudflare knows how to cache stuff.

They’re both very different systems, but both have their uses.

Until they couldn't, because their Stripe account was being rate-limited, for presumably entirely non-technical reasons

In this scenario we did have an account manager at Stripe, however fortunately and unfortunately the show brought in somewhere in the region of 7X the initial target.

Servereless stampeding a payment processor like Stripe as opposed to something like RDS are two different kettles of fish, as am sure you can imagine, stripe have very high capacity and there is 1:1 relationship to the user unlike DB writes, where there may be many per transaction / action.

I have don't it both ways in the past, queueing payments and processing later. Problem is with doing that you cannot immediately return card errors, which has a large effect on conversion. With something like Amazon it is fine to do that, as users inevitably come back to buy the products they want, for donations not so much.

And half the comments on are from people who didn’t read the article.

This isn’t about getting rich quick but setting up a system to respond to high load that comes in bursts.

The authors setup a system to handle bursts of traffic driven by a popular tv show. Normally the traffic is zero so in this case lambda makes a ton of sense.