TL;DR of the TL;DR: Private, self hosted, GPS enabled, local search and offline routing map engine in under a gigabyte, on a $55 SBC, accessible anywhere.

Ok…I’ve been working on this for a while and it is SO satisfying to see it to completion.

I’m calling it “Pi maps” (because the Pi hosts it) but it’s really MapLibre GL JS, served locally by nginx. The Western Australia basemap is a single PMTiles archive, clipped from the Protomaps daily build to the WA boundary, zoom levels 0-15. The finished stack is about 770 MB.

770MB!

PMTiles is one of the bits I particularly like. Instead of running a tile server and maintaining millions of little tile files or a tile database, nginx just serves one archive with HTTP Range requests. Then MapLibre asks for the byte ranges it needs.

The map-serving path is:

Browser
  ↓
MapLibre GL JS
  ↓
PMTiles JS
  ↓
nginx HTTP Range requests
  ↓
WA.pmtiles

And the best part is that I needed no PostgreSQL, PostGIS or live calls to Google etc.

But wait…there’s more! Search and routing!

Rather than install something heavy, I put together a small search service using Python, from the OSM extract.

The source data is converted into a SQLite FTS5 database and the current index contains about 300,000 source records, for a whopping…78 MB. The runtime search service uses about 11 MB RAM at idle.

Driving routes are calculated locally using BRouter, which runs as its own container and nginx proxies the route API internally to it.

The routing dataset for turned out to be remarkably small - roughly 20 MB, with a single worker and a 128 MB Java heap.

The browser gets route geometry, distance, ETA and turn instructions without calling an external routing service.

TL;DR: The current stack is basically three containers:

• pi-maps - nginx + the MapLibre frontend + PMTiles
• pi-maps-search - Python + SQLite FTS5 local search
• pi-maps-brouter - local driving-route engine

After testing, the individual Maps components were around:

nginx/frontend       ~7.3 MiB
search service       ~11.2 MiB
BRouter              ~168 MiB
--------------------------------
total                ~187 MiB RAM

The complete thing - map, search database, routing data, local web assets, metadata etc ends up at roughly 770 MB.

So - a Raspberry Pi 4 is locally serving a zoomable map, POI/address search and actual driving routes for comfortably under a gigabyte of application data.

What’s more, I can use it remotely via Tailscale if I want and/or download the tiles directly to my phone and use it with something like GeoLibre or Atlas.

OSM / Protomaps data
        ↓
   Raspberry Pi
        ↓
 ┌───────────────┐
 │ WA.pmtiles    │ → map
 │ SQLite FTS5   │ → search
 │ BRouter       │ → directions
 └───────────────┘
        ↓
 Browser / phone / apps

It’s not yet a Google Maps replacement. Yet. Right now the routing profile is driving-focused; there is no live turn-by-turn rerouting, spoken navigation, traffic data, or giant commercial address database. But…I have ideas and too much free time (actually, I think I can solve the 4 of those 5 pretty simply).

Case in point: there was a cool technical problem in that the the POI polygon centroid can be a valid map location but NOT a valid driving destination. That was fun to solve.

The next thing I’m going to try is adding a wikipedia layer - or rather, a self hosted ZIM wikipedia layer - so that POIs actually resolve to clickable entries.

There. For once, I win.

PS: This was much easier to do that the whole “family chat” bullshit. Never work with children or animals.

PPS: Also, I may have broken the family chat IRC server, or at least, misunderstood how IRC uses connection resets. So…XMPP might win after all, much to my chagrin.