Map makes unnecessary repeated requests against a source's tile

[daniel-j-h]

Hey folks, I ran into this issue when shipping an overview map displaying vector tiles with a single tile on zoom level zero.

Meaning

• I have a vector tile source with a single tile encoding the whole world on zoom level 0; there are no other tiles
• I restrict minzoom and maxzoom in maplibre's vector source, both set to 0
• When the page loads the vector tile gets loaded once via the network (so far so good)
• When I now zoom in and out, I can see unexpected additional requests against the very same vector tile

This is unexpected to see many more network requests against the very same tile at zoom level 0. I expected the tile to get loaded once and then the map over-zooming the encoded geometry as we zoom in and out without having to make the same request again and again.

It seems like the network requests are cached by the browser but not by the map. This is an issue in cases caching does not work e.g. as in protomaps/PMTiles#272 and ideally the map would not make network requests it doesn't need to make.

---

Here is a self-contained reproducible example with the demotiles for simplicity
https://gist.github.com/daniel-j-h/27f614132e4293bf5de1d3a613a225b7

Here I set minzoom and maxzoom to zero
https://gist.github.com/daniel-j-h/27f614132e4293bf5de1d3a613a225b7#file-map-html-L38-L39

Here are the corresponding docs

• https://maplibre.org/maplibre-style-spec/sources/#minzoom
• https://maplibre.org/maplibre-style-spec/sources/#maxzoom

Note: The same behavior is visible when hard-coding a different zoom level than zero, I just thought this would be easiest to reproduce.

maplibre-gl-js version: 4.1.3 (latest)

browser: Firefox and Chrome on Ubuntu 20.04 LTS both affected

Here is a visual example of what I wanted to use this for: shipping an overview world map where zoom level 0 is perfectly fine and no other zoom levels are needed:

and here you can see the map making the very same network requests over and over again as we zoom in and out

[HarelM]

When using the developer roll with cache disabled I'm not sure you get the desired effect.
Also I'm not sure what are the cache control of the tiles you are serving.
This is not to say that there isn't a bug here, but there is some code that takes care of caching inside maplibre.
Maybe the fact that you are using the same zoom level for min and max "confuses" the code.
Feel free to investigate this. 😃
The main code is in source cache and ajax classes.

[daniel-j-h]

I checked the dev tool disabled cache but that's not it. Here you can see the map still making multiple requests to the single vector tile. The request and response headers are identical by the way and caching by ETag and max-age tag looks fine, too.

Independent of browser caching what I don't understand is why over-zooming needs to make requests at all when minzoom and maxzoom is both set to zero. As in: there should only ever be a single network request to this vector tile.

[daniel-j-h]

Another data point: if we set minzoom to 0 and maxzoom to 1 and zoom in we can see similar browser requests in the network tab that are not explainable. As in: if I zoom in all the way and out again, I get 30+ requests to the same tiles while zooming.

[HarelM]

Can you create a jsbin/stackblitz/codepen reproduction of this issue?
Sorry for nagging...

[daniel-j-h]

Ah I thought the self-contained reproducible gist I shared above would work, but sure I can copy & paste it into a jsbin for you here you go: https://jsbin.com/bufiyejase/1/edit?html,output

Open the dev tools network tab and for example zoom all the way in, then zoom out again. Notice how maplibre makes repeated requests against the 0/0/0 tile while zooming. This is unexpected.

[HarelM]

0 might be a bit confusing due to world copies, so I used 4,4, which still has this issue.
I'm not sure I fully understand the code in here:

maplibre-gl-js/src/source/source_cache.ts

Line 667 in 5086ec2

_updateRetainedTiles(idealTileIDs: Array<OverscaledTileID>, zoom: number): {[_: string]: OverscaledTileID} {

I also think this bug exists for a long time without anyone noticing.

Inside _addTile there's a call to the worker thread to get the relevant tile, which when zooming to zoom 5 seems to get the behavior wrong.
I need to see if this is just when setting the min and max zoom to be the same or a general issue in the code.

Feel free to investigate too as it might take me a while to find the time to dig into this.

CC: @msbarry as I know you are looking at this weird code...

[msbarry]

I've noticed this behavior as well. I think the vector tile worker does actually need to re-process the tile since the layout properties need to be recomputed for each overzoom level, it shouldn't need to re-fetch. Since SourceCache only caches the entire fetched+parsed result it doesn't make any attempt to skip downloading the same data again.

This is on my radar to look into with the source cache improvements, it's very similar to avoiding duplicate fetches for raster-dem tiles used across terrain, hillshade, and countours. I think we want a layer between SourceCache (which is a bad name, it's actually staging the tile pyramid required to render the current scene) and the raw tiles we actually fetch that deduplicates and caches requests.

Some related issues:

• currently source data loading/loaded events are identical for overzoomed tiles and fetched tiles, should we have separate events (or a way to differentiate source data events), for actually loading data from a source vs. when we finish parsing a fetched tile?
• is fetching then parsing in the web worker necessary vs. fetching the tile in the main thread and transferring to the worker? It's easier to deduplicate requests if we do the latter

[HarelM]

Yes, passing the fetched data is a performance overhead.
Service worker might be a good fit for managing the cache stuff, I'm not sure if we can share it with regular workers, but it might be an interesting direction to pursue.

[msbarry]

Yes, passing the fetched data is a performance overhead.

Can you elaborate? We pass the fetched data back from the worker to the main thread... If it's an array buffer then it should do a 0-copy transfer with the built-in transferrable utilities?

[HarelM]

You'll need to wait somehow on the main thread for the fetch to complete instead of sending the message to the worker and continue doing other stuff.
Might not be interesting, but worth thinking about.

[wagewarbler]

Can confirm that I've also observed this same behavior. The behavior with source_cache is to preserve the most recent parent for cross-fading with new tiles as you zoom in for a more fluid transition between zoom levels. If you narrow the range of zoom levels down, I would expect this to "flatten" the tile cache pyramid so that your last parent is the max zoom, but I prob need to give the code another pass with an eye out for this specific behavior. This is something my team needs as well so will likely dig into it soon if someone doesn't get to it quicker.