We run an ad platform that resizes, crops to the important section, transforms with text, and more with billions of requests per day. It's very cacheable, horizontally scalable, and CPU is cheap.

The main scaling challenge here is serving and bandwidth, both of which are handled by their CDN. I expect 5 figure costs for that but the rest of the platform is very expensive.

Wonder if they could use S3 bucket with a lambda to resize. I’m sure their images are a long tail so certain images likely have way more traffic than others. At a previous company where we were also doing a lot of imgix traffic we did something like this but with a go service but now I’d just offload to s3 and use lambda to avoid Devops..

Why would it depend on request rate rather than upload rate? Just pre-scale several variants on upload. Then run an MR job every now and then and delete the variants that haven't been touched for more than e.g. a month. Scale those on demand, store recents and frequently accessed pre-scaled. This is literally a few days of work.

Maybe some caching would help.