Ramblings on Cloudflare cost

[eddy-geek]

Very happy to see the project is growing, and that Cloudflare R2 is now generally available.

I think I could try hosting my PMTiles there, after reading Why deploy Protomaps on a CDN? and Protomaps on Cloudflare, but pricing is scary...

1. Workers - each tile will be 1 worker request,
   a. free plan: limits: 1K/min & 100K/day
   b. Pro plan (5$/month): 10M/day free then 0.5$/M
2. R2 - when there is a cache miss, the worker makes a range read request to R2. this is Class A, 1M/month free then 4.5$/M.
   Cache is not controllable so it's safe to assume that at hobby scale, there will be no reuse between user sessions. still, assuming average user loads 1K tiles/session, that could mean 1000 sessions free per month.
3. Storage - free 10 GB/month then $0.015/GB-month

So, if users start scraping tiles -- as is the common bad-practice for raster tiles -- bill can sky-rocket.
There is no budget limit, only Billing Notifications that you have to hope you read quickly (and that require Pro plan).

On the plus side it seems there are various rate limiting options, that are probably worth a shot:

• old billable Rate Limiting
• newer free WAF Rate Limiting

So it seems for a hobby project (low scale so low cache reuse opportunity), if ZXY compatibility is not needed, using only R2 directly is fine and would be less costly? The main issue I see is the bad concurrent-range request browser support (as per #16 / https://cesiumjs.s3.amazonaws.com/caching/index.html ).

Also, IIUC, smarter hilbert-aware batching of tile request would not help when workers are involved, only for direct protomaps-js → R2 requests -- which could make this option more cost-effective and alleviate the issue above.

Am I right? Do you have feedback from running your workloads, and demos like https://protomaps.github.io/protomaps-themes/examples/maplibre-basemap.html ?

[bdon]

• you can set CACHE_MAX_AGE in the Workers script and choose 1 week or 1 month. a 50% cache hit rate is a reasonable guess at longer cache ages.

• Range reads are Class B operations, which are $0.36 per million.

• 1000 tiles per session is pretty unusual, I would estimate an average of 20-50 tiles per user session, of course it depends on your application.

• The possibility of users scraping tiles is present for any public-facing API, even if you use a hosted SaaS.

Using R2 or S3 directly without the CDN is the recommended deployment method for hobby projects. Range requests work fine concurrently on R2 because it uses HTTP/2, but the latency is poor right now (can spike to 1-2 seconds).

Batching tile requests would affect only the direct-to-storage option, correct. That could be implemented in a modular way at the Source level of the jS library.