Key Features
Tech Stack
The algorithms do divide the map up into chunks that are themselves divided up and so on, but not on the strict geographical basis a quadtree uses. You might not want to divide Manhattan in two for routing purposes, even if the 74th longitude line runs straight through it.
[1] https://turing.iem.thm.de/routeplanning/hwy/esaHwyHierarchie... [2] https://publikationen.bibliothek.kit.edu/1000028701/14297392...
FWIW, Valhalla already does a version of this: it partitions the routing graph into hierarchical tiles and runs with multiple hierarchy levels (highway / arterial / local) specifically to keep search + in-memory working set smaller on long routes: https://valhalla.github.io/valhalla/tiles/
The "quadtree tile unlock" mental model is a nice way to think about it though - if you have a favorite paper / implementation that leans harder into the tiling aspect, I’d love a pointer. I’m currently focused on packaging + offline data consistency, but routing performance on constrained edge boxes is definitely a core constraint I care about.
Right now the offline geocoder in the demo does place/street-level search + reverse, but street + house number ("Main St 12") isn't supported yet. It's explicitly on the near-term roadmap: richer geocoding output with house numbers and (optionally) street/area geometry instead of just centerpoints.
Photon is built on Elasticsearch (Java) - so it tends to mean a heavier index + higher RAM/CPU expectations and more moving parts. That's fine on a beefy server, but it is a rough fit for the "drop-in appliance on small edge/on-prem boxes (amd64/arm64) + simple ops" goal.
Corviont's geocoder is intentionally "boring": a single SQLite file + an HTTP service, built from Nominatim-derived data. Fast startup, low RAM, easy to ship per-region, and it stays consistent with the rest of the stack.
That said - if there is demand for a "server-grade geocoder option" for people already comfortable running Elastic, I am not opposed to offering it as an alternative profile. The default is just optimized for constrained edge hardware and minimal moving parts.
Have you (or anyone here) actually run Photon on edge-class hardware? If you have real-world numbers, I'd be interested in seeing them. When I add house-number search, Photon might be an easier route than enhancing my current approach.
Let's start it with index of whole Spain, 2.4gb download, 4gb on disk: https://gist.github.com/dkourilov/e243270684b5973f1fac005c78...
I'd say it's pretty usable to run a EU-sized country or several US states on any commodity PC. For embedded devices, it really depends what is the device. On Raspberry PI it should be fine for batch geocoding, realtime (typeahead) will definitely be lagging.
Those numbers look pretty reasonable. I’ll keep Photon in mind and, as I get time to benchmark different approaches on a few representative regions/hardware, I’ll use the results to decide what the best way forward is - and I’ll publish the numbers when I do.
photon is just a single process and opensearch runs inside it (but you can run photon and opensearch separately). Saying "Java" means more memory is in general wrong as the underlying technology Lucene is heavily memory optimized.
One place Corviont is trying to differentiate is the update story for edge/fleet deployments: the goal is a signed, resumable regional dataset updater (verify manifest -> atomic swap -> reload/rollback) so boxes in the field can stay fresh without manual rebuilds or "re-download the world" updates. Headway (at least from a quick skim) looks more like "bring your own data / regenerate when needed," which is totally fine for servers, but fleets usually need something more automated.
If you've seen Headway (or similar) handle incremental/regional updates well, I'd love to learn from it - updater design is the big missing piece I'm validating demand for.
Indeed there is currently no incremental update in Headway, and deployments are largely an exercise left to the reader.
For maps.earth (a Headway planet deployment), I typically rebuild the world, and then do a blue/green deployment.
I guess the one exception is for transit routing. We have individual transit zones small enough to fit into memory, which can be deployed incrementally. There’s nothing really built in about it - just another level of indirection via our “travelmux” service which redirects your routing queries to a different backend depending on mode and region.
I am trying to learn from real deployments as I design Corviont's updater for edge boxes (bandwidth caps, maintenance windows, unreliable WAN, atomic swap + rollback).
When you say transit "zones" are small enough to deploy incrementally - what is the actual artifact per zone (roughly what format), and what sizes do you typically see?
And when a transit zone dataset changes, how do you roll that out safely - do you restart/reload the backend that serves that zone, or do you bring up a new backend/version and then flip travelmux to point at it?
So it’s not really incremental with respect to the existing transit zone deployment. I just mean I can redeploy a single transit zone with the latest GTFS without having to touch the other transit zones, tileserver, geocoder, etc.
Deployment/rollback is handled by k8s config.
I ask because I've been looking to self host some sort of map tile server and they seem to have database in the hundreds of GB.
Maptiler has a bunch of datasets, anywhere from 385MB to 527GB but the OSM dataset is only 70GB. (MBTiles format)
Also it’s not one giant tile DB - there are 3 datasets:
- map tiles (PMTiles)
- routing tiles (Valhalla tiles)
- geocoder index (SQLite)
For small countries like Austria/Slovakia, each is typically hundreds of MB.
If I may have a feature request, I’d like to have only some of the features turned on - in my case it would be just the reverse geocoder (so I could skip the map and routing data download and storage).
Right now I have my own reverse geocoder for https://weathergraph.app which downloads OSM dumps and builds in-memory KD tree for lookups. Surprisingly, the whole world can fit in 3-4 GB of RAM, and service starts in 90 seconds on a cheap VPS, no database needed.
Re: Weathergraph - thanks for the details. Since you already run whole-world reverse geocoding on a single server, that's a bit different from Corviont's current regional/fleet packaging (where you ship only the area you need to each edge/on-prem deployment). A "world geocoder-only" pack could still make sense - but it's a different distribution/update story than my default.
For your use case, do you want reverse results at the city/region/country level - or do you also need street/house number detail? That choice mostly determines how heavy a world geocoder-only pack needs to be.
I didn’t realize how many use cases are there - you guessed right that for my use the whole world, but only up to city level, is quite enough.
In my todo list I have adding landmarks, so that I could eg. show a nearby mountain when outdoors and far away from city, but then deciding which POI types to include gets much trickier.
This is also why I lean on Nominatim-derived data for Corviont - it already condenses a lot of this into a single importance/ranking model. I have not tried a whole-world build though, so I am not sure how well that behaves at planet scale.
Why not create a "builder" repo, where people could generate their own local datasets by a bounding box?
Re: a bbox "builder" repo - it's an interesting idea. I could see it going two ways: (a) you want to run a bbox builder yourself, or (b) you want a simple way to specify a bbox so the dataset pack can be produced for you.
I started with the "ship a known-good pack" approach because the build pipeline is the messy part, and I want deployed boxes to stay simple/reproducible.
For your use case, which did you mean - run the build locally, or "draw/paste a bbox and get back a ready-to-run pack"? And would bbox be OK, or do you prefer admin boundaries (country/state/city)?
I get why you created the monaco pack, it's a nice demo and fast to download and run. I would rather choose a big city where potential users live (london, nyc, the bay,?), but maybe there are technicalities that make that more complex.
Drawing box on a map in browser and generating the tiles, routing and geocoding db sounds quite heavy for backend compute. There was a project wiht a website that could generate a tileset (pmtiles file?) from box drawn on a map, but I can't find it now. There was a limit to the box size and if it was over some threshold, you had to have premium, or contact sales, or sth, I thought it was protomaps, but no tool like that there now.
Anyway really nice idea, I will follow your progress for sure!
Bounding box input is a really interesting idea. For my product, the best use would probably be to automate the request/fulfillment flow if this ever picks up. Letting someone draw/paste a bbox makes the request unambiguous and could be automated end-to-end.
Btw, the pack build process isn’t that compute-heavy. It’s not instant, but as a rough data point: on my laptop, a small country like Austria/Slovakia is on the order of ~1 hour to build the three artifacts (tiles + routing tiles + SQLite geocoder).
If you read the bottom of the page they plan to sell it. Not sure how that works when all the software/data is open.
It is built on open-source components and OSM-derived data, so the obligations are mostly (a) keeping third-party software licensing/attribution intact, and (b) OSM/ODbL compliance - clear attribution, and if you distribute derived OSM databases, meeting the ODbL share-alike requirements for those database artifacts.
I am handling this via attribution in the UI/docs and a licensing reference bundled with the dataset that points to a public licensing/attribution repo containing the third-party license texts and details: http://github.com/corviont/licensing
What I am selling is the packaging + ops side: ready-to-run region packs and (eventually) a signed updater for fleets.
I've played around with OSRM, and Nominatim, etc, but had to do some trickery to run on a raspberry pi.
(For anyone interested in running some of these kind of things on a pi, I talk about it generally here, I need to post an update with more info at some point. http://blog.onaclovtech.com/2025/02/general-purpose-to-speci...)
One goal of Corviont is to avoid the on-device pain you hit with OSRM/Nominatim: the region pack is built once elsewhere, and the edge box mostly just serves prebuilt artifacts (PMTiles tiles, Valhalla routing tiles, and a SQLite geocoder index based on Nominatim).
In practice, requirements scale with region size and traffic. For larger regions the main constraint is usually SSD storage plus enough RAM headroom for routing/cache. I also picked Valhalla partly because it generally has a smaller RAM footprint than OSRM at serving time (OSRM is extremely fast but tends to be more memory-hungry).
If you have 1-2 concrete suggestions on what makes it feel that way (copy/layout/typography), I am happy to improve it.
Not affiliated with Hacker News or Y Combinator. We simply enrich the public API with analytics.