Protomaps: Self-hosting map tiles has gotten cheap and easy
bruceroettgers.eu
bruceroettgers.eu
> What is OpenFreeMap?
> OpenFreeMap lets you display custom maps on your website and apps for free.
> Using our public instance is completely free: there are no limits on the number of map views or requests. There’s no registration, no user database, no API keys, and no cookies. We aim to cover the running costs of our public instance through donations.
> What is the tech stack?
> There is no tile server running; only nginx serving a Btrfs image with 300 million hard-linked files. This was my idea; I haven’t read about anyone else doing this in production, but it works really well.
You'll also see upselling from MapTiler on the pages you're required to attribute along with upselling in the docs https://openmaptiles.org/docs/host/maptiler-cloud/
Self-hosting ended up being relatively cheap and easy too. They publish the whole planet as a single PMTiles file every week, and I just copy that to R2 and serve it through a small Worker, basically the Cloudflare route the article mentions.
Few people store data on a server as tiles though it might be served as tiles. Aside from the added storage size, you dont want to have to update 20 levels tiles every time one piece of data is added or removed.
If you are optimising for cost and read-traffic, such as OpenFreeMap, shifting that complexity to a preparation step makes a lot of sense.
I just think the postgis approach works better at both ends of the scale. For large users its a mature, well maintained (by others) and heavily optimised (again by others) system. For small users the benefit of not having to worry about rebuilding individual tiles especially when data updates is a no brainer.
Openfreemap probably has the use case where its pretty close to comparable (set up for osm data so global, large data, very clear use case and feature set to design for, you can update via cadence, enough of a user base to support development, the sell of it all being "one package" is positive for some users, etc.). Whenever i see protomaps, versatiles, tilemaker, maptiler etc. Ive never seen a major use case I can buy into them for.
If you are into this stuff, you can view PMTiles from the CLI with "gloss", or try to drag-and-drop on its web interface [1]. It seems large tiles (>128M) and some raster formats are failing; plus it would be nice to load from URL there. Will fix that...
(AI warning). ntmaps [2] is a Golang library that lets you use different tile providers, deployable in terminal or browser. ntplaygrounds [3] uses that and nt3d and ntactor to make a city simulation. Data is pulled from OpenStreetMap in those.
[1] https://nimbleterminal.com/gloss/app.html [2] https://nimbleterminal.com/ntmaps [3] https://nimbleterminal.com/ntplaygrounds
The official PMTiles viewer [1, 2, 3] is also pretty good.
[1]: https://pmtiles.io/
Cloudflare, generally, is amazing for this, actually. It serves my personal hiking map for about $2/mo, using R2, D1 and Workers.
Building the map itself is free via GitHub Actions.
It's bonkers how the map world has changed over the last 20 years.
1. The app is a remix app, workers do server-side rendering, so that a vanilla `curl` shows the actual payload, not just a script tag that loads the app.
2. I'm a great big nerd, so I recently added the ability to post reviews/photos/GPX traces to your AT protocol PDS. The app has a Worker cron job that subscribes to Bluesky's Jetstream for events in its namespace, and indexes these into D1 in order to answer certain API calls like "get traces in this area" or "get reviews for this hike". So the app itself isn't the primary source of truth, a person's PDS (which they control) is. The app just maintains a cache/index that can be rebuilt.
The real "too good to be true" might be GitHub Actions building a 100GB+ PMTiles file for free
Bunny's volume tier costs $5/TB served, i.e. $1 gets you 2~6.7 million tiles served, probably closer to 3.
But Cloudflare also allows public R2 buckets so we could potentially ignore the Workers cost, i.e. $1 gets you 2.8 million tiles served. I think caching just works in front of R2? So you get significant gains if your visitors are mostly looking at the same areas and not zooming in all the way to their house every time.
So they come out pretty evenly or with Cloudflare a bit on top depending on the scenario. Bunny does has the advantage of easy data replication across regions (I don't know of a good way to achieve this with R2, although D1 has some kind of replication feature in beta).
Both are still far, far cheaper than almost any vector tile subscription I've come across. And, of course, more expensive than self-hosting (at least at the hobbyist level).
[0]: https://bruceroettgers.eu/blog/benchmarking-r2-head-requests...
and we have a CF worker than turns them into rasters on the fly if that's how people want to get them
Hence Protomaps. Hosting costs almost nothing.
The main problem though is that once you host your tiles, why not put what you want inside ?
Tune Tilemaker's processing.lua to your envy. Spend hours designing maps and building a Typescrpt dynamic Maplibre style.
Maptiler is pretty good company, maybe they would support the project.
We now generate our tiles every week for 2 €, and could soon generate them daily for .5 €
the fun/hard parts of the project weren't really scale dependent, although downloading source imagery and building the tile pyramids did take literally weeks to run. actually, the hardest part was getting demosaicing right, e.g. making the quilt seamless. filtering out cloudy and smokey/hazy tiles while picking alternatives from the same time of year so groundcover & crops are in a consistent growth stage so you don't end up with checkerboard patterns or cloud silhouettes randomly across your map. and there were other problems that came up more or less just because of how & when the satellite took the images that require exposure correction and more filtering
I also have a pelias index with 170M records for search and self-host valhalla for routing. the open source mapping ecosystem has really impressed me the more I learn with each project
Here is a commercial one: https://cloudless.eox.at/preview#map=14.00/-80.77673438/-33....
I really like their data set. For self hosting, vector map tiles are small enough that storing it in a bucket won't break the bank. I think you need somewhere around 100-150GB or so for protomaps. A few dollars per month goes a long way. However Mapterhorn's elevation data set is something like 11 TB or so. That's going to cost some money to store somewhere. Satellite imagery is similarly expensive to host. And you typically want both.
There is of course an asymmetric relation with what you store and what you serve. Ideally, you store data for the whole planet. But you probably are only going to serve a tiny portion of it. This favors hosted setups where you pay for what you use via a CDN and share the storage with others. A shared service also relieves you from the burden of having to keep that data up to date. The main issue with that is commercial map services tend to overcharge for more intensive usage to the point where self hosting actually becomes attractive despite this.
The main issue is that the number of requests really goes through the roof if you start getting tiles beyond certain zoom levels, which pushes you into the expensive price tiers. With most providers you pay per request and you can easily generate hundreds of thousands of requests just panning around at low altitude over a map, which means you download lots of small tiles. That's how we hit the Maptiler limits. They have high resolution elevation data and exploring the Alps is a lot of fun. We hit 800K requests per day without any real users using our system even. This was just us playing with the data. We switched to Mapterhorn and cap the zoom level at zl-12 for now. Since it's a free service I want to be a bit mindful of not hammering their service.
I'd be interested in reasonable commercial solutions that won't cost me 300+$ per month that would allow me to remove zoom level limitations. At those price levels, self hosting becomes a possibility of course. Which suggests to me that somebody should be able to undercut that price and still make a decent profit.
> VersaTiles is a completely FLOSS stack for generating, distributing, and using map tiles based on OpenStreetMap data, free of any commercial interests. VersaTiles is a free and open alternative to commercial map services like Google Maps or Mapbox. It is designed for newsrooms, NGOs, developers, and public institutions who need reliable maps without vendor lock-in. VersaTiles is fully self-hostable, requires no API keys, charges no usage fees, and does not track your users.
With Mapbox or Google Maps I wouldn't even have considered building this game. Protomaps made it a no-brainer
Thanks to Brandon and the OSM contributors. This is what made the idea affordable for a side project.
After evaluating maplibre vs the somewhat-abandonned protomaps-leaflet library, I decided to go with the latter as I don't really need the fancy 3d gpu and custom layers functionality of maplibre. Plus leaflet is releasing a long-awaited v2 right now.
Dirt cheap to run, theme-able, easy to use.
https://developers.cloudflare.com/cache/reference/range-requ...
https://github.com/pf-wikis/mapping (using maplibre-gl-js and some custom Java code to manage layers, tile generation, and PoI collation from the associated MediaWiki)
https://github.com/pf-wikis/mapping-data (using kart)
[0] aportee.ca
dktue's branch has all the perf PRs applied to it: https://github.com/dktue/tilemaker/tree/perf/planet-3min
The app will periodically download the configured map regions and store the pmap files [1]. The Go backend then directly serves them to the frontend [2].
[1]: https://github.com/jo-m/cartomancer/tree/a02f042173b5ba1a7d9... [2]: https://github.com/jo-m/cartomancer/blob/a02f042173b5ba1a7d9...
It costs me something like $2/month iirc, and it made tile querying faster over protomaps' free hosted endpoint.
p - proxy - tp8ez dot bunny dot run
so I closed that window.
What else you got?
Thanks for the explanation.
Protomaps sounds interesting.
My experience with it is ... ok. It works. But going directly to cloud storage (s3 in my case) is unpredictable latency and neither the browser nor most servers really cache HTTP range requests granularly, so every tile is another high-latency round trip. Not worth it - I ended up putting Martin tile server in front of it.
I assume someone could write a protomaps adapter in WASM with some smart caching? Until then, Standard GET requests are just easier to cache and perform much better in practice. The downside being you have to run a server or statically generate files.
One thing I really like is that the data is static. This isn't useful for some applications, but for my use, namely blogging about my travels, it's perfect. I wouldn't come back and edit the text or pictures to reflect subsequent events, so why would I edit the map? Being able to take ad hoc snapshots of the map at a point in time is perfect.
It works amazingly on GrapheneOS, Android Auto, and Desktop app via Flatpak.