digging into their GH repos surfaces https://protomaps.github.io/PMTiles/?url=https%3A%2F%2Fr2-pu... which is at least more detailed if not quite a GMaps killer

Edit: And I missed parents link to https://app.protomaps.com/downloads/small_map which is actually better.

It's important to me that I showcase the system working as advertised - you are free to download the 110GB file from that static storage with no restrictions. Unfortunately, the demo is slow because I'm hosting it on Cloudflare R2 which has no outgoing bandwidth fees. Before, it was on DigitalOcean Spaces, which is low latency but would cost me $1-2 out of my own pocket each time someone clicked the link.

The Lambda/Workers integration as described in the other comment is the preferred solution to this for developers operating this at planet scale. Of course, the static storage solution is perfect for small to medium-sized areas, even on GitHub Pages.

Can you expand on this a little? Cloudflare is supposed to be pretty competent as CDN and with caching.

https://protomaps.com/docs/cdn/cloudflare

However, it is an optional acceleration layer on top of the PMTiles access pattern; and it doesn't make a compelling demo, it's just a Z/X/Y layer which is how map tiles have worked for the past 20 years.

I have my own tiny personal project for maps renderer using Leaflet. For now I use 3 sources for tiles (Google, OSM and Esri) but having more sources would be very cool.

To expand a bit -- There are adapters for mapbox(gl/libre) and leaflet, and so on that will allow you to use a .pmtiles directly from the client with a range request supporting server, like s3 or nginx.

I've used protomaps basemaps to generate an Ireland and EU basemaps for a project from OSM data. I still haven't quite figured out the tuning for which features should be shown, combined, or excluded yet, and the generated layer names are sort of, but not quite compatible with similar base layers + styles from mapbox (which uses the same base data, but a different, proprietary conversion). This part takes a reasonable skill as a GIS person and some taste in the design of the feature symbology.

The downside of doing this on S3 directly is that you might wind up publishing a link to a 100 gig file that would cost $10 if someone just downloaded it. Lambda invocations seem cheap compared to that risk. OTOH, throwing it on a hetzner box is quite reasonable.

This is the focus of most development for the next few months - like you said, it's also a matter of aesthetics and context, like if the map is underlying another PMTiles data layer it should be lower contrast then if it's overlaid with pin markers. The end goal is to have a flexible basemap solution adaptable to most apps but you are free to roll your own proprietary style on top of the open source tiles if you'd like (there are already companies doing this)

I've got a couple of goals, or at least evaluations for goals --

1) Have a stripped down basemap like Carto Light, but with different emphasis for the "bicycling friendly" roads, and selectively higher contrast/emphasis for those minor through roads that discourage car traffic but are perfect for a bike. (e.g. Irish boreens, those 1 lane or double track "paved" L roads that are everywhere outside of the major cities.) I find that I simply can't see the minor roads on apple/google maps when I'm out unless I'm zoomed in to the point that I can barely see a network. (fading eyesight and super low contrast).

2) Another lighter base map, but with transit focused features to compliment some of the open data in Ireland around transit -- the routes and timetables and so on. I'm aiming for a site that can be 100% statically hosted but will show the routes in a more friendly manner than Bus Eireann's site.

3) If it goes well, I might be doing this in a more commercial context for some clients who are currently using mapbox for their basemaps but due to political concerns need to insert in different names/boundaries for disputed areas/features. We'd love to be able to self host, but quality is a major concern there.

The bugs that I'm seeing in the 1 and 2 cases are things like handling the look of freeway interchanges/flyovers at close zooms, consistently getting rivers to show at appropriate zooms because they're into two separate feature types depending on the width, with the narrow bits getting dropped. And the general mismatch between the styling I'm used to from the mapbox converted layer names/feature types and what's coming out of pmtiles/basemaps. The feature bits look like tippecanoe coalesce/drop densest features tuning, but I'm also looking at just dropping out entire feature sets to drop the size of the tiles. That should help the coalesce/dropping behavior as well.

OSM is a freeform dataset and not a cartographic product, so most of the basemap work now and in the future is on getting good results of this transformation for 200+ countries. The mismatch vs. existing maps must exist because we need to ensure to downstream commercial users that all end products of the map generation are openly licensed - that's why this is being pursued as an independent, self-funded project with support from GitHub Sponsors: http://github.com/sponsors/protomaps

With 512MB RAM these generally complete in under 100ms so the unit costs of Lambda are within a few multiples of Cloudflare Workers (.5USD / million invocations)

Do you have any tips for doing so without renting memory in the cloud?