Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

CLI tool to concatenate or merge pmtiles #105

Open
daniel-j-h opened this issue Dec 14, 2023 · 13 comments
Open

CLI tool to concatenate or merge pmtiles #105

daniel-j-h opened this issue Dec 14, 2023 · 13 comments

Comments

@daniel-j-h
Copy link

We've been working more and more with pmtiles recently and we're coming across various use cases, such as concatenating multiple pmtiles into one. At the moment we're simply loading multiple pmtiles in our frontend but ideally we could bundle up multiple pmtiles into a single pmtile to put on a static host and be done with it.

Usage could be as simple as

pmtiles cat output.pmtiles input0.pmtiles input1.pmtiles ..
@bdon
Copy link
Member

bdon commented Dec 15, 2023

There are a few details that complicate doing this correctly.

  1. If the if the tiles contained in each input pmtiles have no overlap, this is very easy to do
  2. If there is tile overlap, we can use the ordering of the inputs to determine priority: input0.pmtiles has priority. However, this may not work correctly for low zooms: imagine merging two raster tiles that cover Europe and South America. The Zoom 0 Tile is only correct if it contains the merging of the pixels.
  3. If there is tile overlap for vector tiles, and the layers are disjoint, we can deserialize the overlapping tiles and combine their layers or even directly concatenate the protobufs.
  4. If there is tile overlap for vector tiles AND they have overlapping layer names, then we need to deserialize both tiles and then do a merging of their layers which requires rewriting the dictionary encoding of the keys/values.

In a perfect world we could accomplish all of the above with go-pmtiles, but it's preferable to first address the use cases people actually need (which do you need?). A major factor in determining these is whether it complicates the go-pmtiles dependencies. For example 2) above would require adding image processing libraries for PNG, JPG etc which will probably need a native dependency (we don't want any).

@rriemann
Copy link

I have a Jekyll blog that I also use for travel blog posts. In the yaml meta data I can put a GEOJSON object. Then, a plugin loads currently mapbox with mapbox cloud data. I want to move the maps to selfhosting.

I do not want to have map data covering the entire world on high zoom levels, but instead only files for the area that I cover in my travel blog.

In my mapbox gl style file, I need to
(a) change for each geojson object the source to the corresponding pmtiles file. I do not know currently how I can pass through the file name from my javascript code to my style.json without inlining the entire style.json
(b) use only one pmtile file that is the concatenation of all pmtile files of the blog.

From what I read, (b) is so far not possible and difficult to implement. Hence, I would need to find a way how I can do (a) in the most elegant way.

@bdon
Copy link
Member

bdon commented Dec 18, 2023

@rriemann Is each region displayed on a separate page? If so, you can re-use only the layers key of the style.json and then dynamically swap the sources values in JavaScript to accomplish (a). Just changes the url key of the source name to point at the pmtiles.

(b) will be difficult to implement as one big pmtiles if there are any low zoom tiles that overlap among the regions, since we'd have to do an individual tile merging for it to look correct.

@daniel-j-h
Copy link
Author

In a perfect world we could accomplish all of the above with go-pmtiles, but it's preferable to first address the use cases people actually need (which do you need?).

Thank you for this detailed reply! We came across the following use cases recently

  1. Some datasets like the Copernicus DEM come in multiple files (e.g. 1° grid) and we can get one pmtiles file per grid-tile easily. But ideally we'd merge them at the end into a single DEM pmtiles file again. Adjacent and disjoint.
  2. We generate raster pmtiles for the same quadkey over multiple years. At the moment we create e.g. 2020.pmtiles 2021.pmtiles and so on. Ideally we'd merge them into a single pmtiles file where the years are represented as layers.

There are workarounds for now e.g. to merge the mbtiles files before converting to pmtiles but a pmtiles cat would simplify things and provide a convenient way to handle use cases like the ones described above.

@bdon
Copy link
Member

bdon commented Dec 20, 2023

I don't know of any smart MBTiles mergers that would accomplish the correct per-tile merging above - have any pointers?My suggestion for 1) is to build a single pmtiles from a VRT mosaic over the 1 degree grid cells. They won't be truly disjoint if there are any lower zoom levels generated, as those will almost certainly have some common tiles.

Use case 2 is 3) from my post above which is relatively easy to accomplish. The simplest way is to do direct PBF concatenation, relying on vector_layers to enforce that there is no layer name overlap. If there is layer overlap the program should exit with an error. Then the algorithm is to iterate over tileIDs from archives A and B, taking into consideration deduplication/RLE (a little tricky), unzip both tiles, concatenate the protobufs (respect input ordering, probably) and re-zip the output tile

@coogle
Copy link

coogle commented Jan 21, 2024

I would be interested in this feature as well, specifically when building raster tiles in the cloud it would be nice to be able to do something like "this .pmtiles was from 0-11" and merge it with one that was from 12-18, etc. Given the cost of computing power, it would be useful to be able to not essentially waste-work if for some reason you had to try to recover from a particular point in the multi-day process of building the rasters.

@bdon
Copy link
Member

bdon commented Jan 21, 2024

It ought to be relatively easy to implement pmtiles merge a.pmtiles b.pmtiles for case 1 above, throwing an error once any overlap in tiles is detected. Sounds like this would handle a good number of valid application.

@cldellow
Copy link

cldellow commented Jan 28, 2024

I'd be interested in this feature, too - specifically case (3) described in #105 (comment)

I've used tile-join to similar effect, but based on felt/tippecanoe#10, I wonder if its support is more rudimentary than what's in go-pmtiles.

@bdon
Copy link
Member

bdon commented Jan 29, 2024

I'm thinking we do this:

pmtiles merge INPUT_1.pmtiles INPUT_2.pmtiles INPUT3.pmtiles -o OUTPUT.pmtiles

  • if the input is raster, it will fail if there is any overlap in addressed tiles
  • if the input is vector, it will fail if there is any overlap in stated vector_layers in the metadata (we would need to trust the metadata to be correctly describing the tiles). Otherwise the merge will byte-concatenate the layers

We can maybe add a --overwrite-overlapping-tiles option that will accept overlaps and choose the relevant tile from the first of the archives in order. I don't have any concrete plan to implement true tile-level merging.

minzoom/maxzoom issue

Finally there is an outstanding issue of how to deal with merging archives that do not have matching max zooms. If I merge one tileset that has max 14, and another that has max 15, my archive in total has a max of 15, but when I look at the 14 area in most browser and zoom into 15 I get nothing. This is a limitation of how maxzoom is expressed in most clients like Leaflet, OL, MapLibre. I don't see a smart way to rectify other than also enforcing maxzoom is uniform across all merged tilesets.

JSON metadata

We need to update the binary header like bounds, minzoom, maxzoom, # addressed tiles to be accurate after a merge. But if the JSON metatadata has keys like vector_layers or tilestats we may need to modify those. Since vector_layers is a part of the specification we can figure out what the combined layer set is. But tilestats may just need to be deleted, and any application-level JSON metadata could become incorrect after a merge operation.

cc @ZeLonewolf

@ZeLonewolf
Copy link

Thanks, this feature would help me for a few obscure development use cases when I want to demonstrate a change in OpenMapTiles and style it with Americana, the the examples I have are in different parts of the planet. For example, I'd currently like to render a sample that has Massachusetts, Yukon, and Spain to proof-out an OpenMapTiles concept (and hence I'm rendering it on openmaptiles-tools. In my admittedly one-off use case, I really don't care what happens in the case of tile collisions.

@prusswan
Copy link

To address the issue of tilesets with uneven/variable max zoom level (e.g. protomaps/PMTiles/discussions/321), I see a use for merge + extrapolate(?). Assuming the original tileset has a maxzoom of 15-16:

  1. extrapolate z15 tiles into a full set of z16 tiles
  2. extract z16 tiles (partial coverage) from the original tileset - this is already supported
  3. merge results from 1 and 2 (with 2 taking priority)

This is not a common usecase but I feel it makes more sense to "fix" the data, instead of coming up with hacks for different renderers.

@Patukas
Copy link

Patukas commented Mar 4, 2024

My use case:

I am working on a raster dem map that needs a detail of zoom 13 in one area but needs only zoom 10 in the other area.
If a render all the map in z13 is 100G, too much cause storage limitation. Solution:

-Create one area in z13 and the other in zoom 10 and then merge.

@mgibbs189
Copy link

@bdon @daniel-j-h 👍 for pmtiles merge

it's preferable to first address the use cases people actually need (which do you need?)

Case 1 (no overlaps) would be an excellent start. Especially useful for larger datasets split into chunks.

Related: protomaps/PMTiles#403

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

9 participants