Asynchronous GeoTIFF reader

[weiji14]

Rewrite https://github.com/geospatial-jeff/aiocogeo in Rust!

We'll probably want to leave all the non-filesystem I/O to object_store, and focus on decoding IFDs asynchronously.

Help is most appreciated.

[kylebarron]

I started a prototype here: https://github.com/developmentseed/aiocogeo-rs

It's able to read all TIFF and GeoTIFF metadata and decode a JPEG tile, though general read/decompressor support is not yet implemented. I definitely plan to implement full-tile-decoding support, so that we can use it from Python efficiently.

It's not at all reliant on the tiff crate (it does use some enums that I could later copy over).

One question is how to marry both synchronous and async read APIs. One good reference is the parquet crate, which has both sync and async APIs to read Parquet.

[feefladder]

Hmmm.... I wanted to use tiffs in bevy_terrain, so modified the image/tiff crate to do async.
Some nice discussions on how to merge:

• https://users.rust-lang.org/t/what-to-pin-when-implementing-asyncread/63019/7
• https://users.rust-lang.org/t/how-can-i-use-the-code-written-on-read-write-seek-for-asyncread-asyncwrite-asyncseek-streams/85037/11

Basically: Put all non-io-related code (working on [u8]s) in a top-level module and then put (async)-io-related code in a submodule. What I did in image/tiff is to create a cursor over the byte array so I modified as little code as possible. I admit it's somewhat arcane, because I didn't know if the image/tiff folks actually wanted to maintain async support 😁. But now I see (busstoptaktik/geodesy#86) that dev efforts are concentrated here rather than georaster, I'll see if I can give it a shot....

[weiji14]

Oh wow, thanks @feefladder for all that work (and the PR at image-rs/image-tiff#245)!

I didn't know if the image/tiff folks actually wanted to maintain async support 😁. But now I see (busstoptaktik/geodesy#86) that dev efforts are concentrated here rather than georaster, I'll see if I can give it a shot....

Not sure either, but I'd definitely like it if most of the async code could be put in image/tiff, especially if there are no 'geo'-specific parts. If I'm not mistaken, image/tiff only supports full reads and not partial reads (e.g. it can't read specific tiles within a Cloud-Optimized GeoTIFF based on a bounding box) right now? There's some experimentation going on at aiocogeo-rs that might support partial reads at some point (based on original Python implementation at https://github.com/geospatial-jeff/aiocogeo/tree/0.3.0?tab=readme-ov-file#partial-read).

Eventually though, it would be great to have less fragmentation on different ways of reading GeoTIFFs in Rust, and have all the core implementations live upstream in image/tiff or here in georust/geotiff. But there might be a period where we'll need to try different styles and see what sticks 🙂

[feefladder]

Actually, I did some work on georaster based on the async stuff... lemme see if I can make that into a PR...bEDIT: ah it turns out my files didn't save properly...
image-tiff does support partial reads, it has a load_chunk(chunk_index: u32) function. Not sure if that is in release though 😁

[feefladder]

My pr over at image-rs/image-tiff#245 got rejected because it was too big and I didn't create an issue first. Then a proposal I had for more control (image-rs/image-tiff#250) there is no maintainer bandwith to support my proposed changes (also here). So I don't know if it is feasible in the near future to have async on top of image-tiff.

Anyone here have ideas for what is a good way forward?

[weiji14]

@feefladder, first off, I really wish I had your energy to do such deep technical work in the (Geo)TIFF Rust ecosystem. I can understand both your frustration as an open source contributor wanting to make meaningful changes, but also feel for the image-tiff maintainer who has to review a 1000+ line PR for a crate used by 47k+ of downstream projects...

Personally, I'd prefer not to fragment the TIFF ecosystem by having X different implementations in X repos, which is why I've suggested pushing things to image-tiff. That said, me and my colleagues at DevSeed are aware that the image-tiff maintainer is stretched rather thin, which is why the aiocogeo-rs repo Kyle mentioned above was created. The idea is to have an independent implementation for an async GeoTIFF reader in that aiocogeo-rs repo (that doesn't rely on image-tiff), and once that experimental work is stabilized, we will slowly upstream parts of it to image-tiff or geotiff crate as needed. This means a lot more duplicate work, but it can be good to have a proof of concept to show and get people excited about (and hopefully the image-tiff maintainers will be more receptive to a contribution that has been battle-tested in the wild).

Of course, I don't expect anyone to trust my obviously biased intentions on getting aiocogeo-rs working since it is currently hosted under the organization I'm working for. I'm aware @feefladder that you'd maybe prefer a different arrangement/license based on georust/meta#34 (comment) and some Discord discussions, and respect it if you'd like to carve out a different space for async GeoTIFF. Keen to hear your ideas on a way forward.

[kylebarron]

To add to this, I would love to figure out a good architecture such that we could use some of the decoding parts of image-tiff while managing our own data fetching. This is hard however because of image-tiff's architecture, and the fact that it lazily reads tiff tags on access instead of up front. (the reading and the decoding of the tags is combined).

But maybe there's a way to use aiocogeo-rs for the data fetching and tag decoding, but reuse image-tiff for the image decoding? That's what I was heading towards before I ran out of time (I'm doing a lot of work on open source rust vector projects)

[feefladder]

Thanks for the uplifting responses :)

I think the aiocogeo approach is a good one - having an async tiff reader specifically for COGs. That could have a smaller code/exec footprint than image/tiff without needing to worry about arcane tiffs (since being only for COGs is in the name). What I was currently thinking of is to put all lessons learned from image-tiff into such a reader, maybe slap on a writer as well. That would mean yet-another forked-and-maybe-rebranded-tiffreader. Which is exactly the problem we're trying to avoid here. What would be ideal is to have agreement between here and image-tiff on what direction image-tiff could develop, so that things made here can actually be upstreamed. What could be done as a middle ground between completely forking away and waiting on image-tiff, is to create a submodule (very c++ style) for image-tiff that we link to in our cargo_toml. However, there is quite a big change underway over at image-rs/image-tiff#242

This is hard however because of image-tiff's architecture, and the fact that it lazily reads tiff tags on access instead of up front.

How should it read tags? Read all images, including all tags Dos-style? I think it makes sense to read only the relevant tags for the current overview (thinking COG) and geo tags. Then, I dont't really see a way in which one could read the desired part of an image in less than 3 sequential requests (for which we know the overview we want to load). Doing some testing, I've found that finding the nth ifd normally fits within the (16kB) buffer from the first request. Then, when reading tag data (~60MB, further in the file), the pre-existing buffer is cleared, making further reads into other tag data slow. Now, to speed things up, I would Keep the first fetch around, since it contains all ifds, but then that is reader-implemented.

There is a slight problem also in the design there, since async doesn't mean concurrent. That is, if we make the decoder reader-agnostic (having <R: AsyncRead + AsyncSeek>) then - if I understand correctly - we cannot concurrently read tags without having multiple readers, since poll_read takes a &mut Self. What I did in image-rs/image-tiff#246 is

pub struct<R: AsyncRead + AsyncSeek> ChunkDecoder {
  image: Arc<Image>,
  reader: R
}

and then "cloning" the reader for concurrently reading multiple chunks - which is actually not really cloning, since it only clones the current buffer and not the pending read/seek -> is that too ugly? That's what ObjectStore circumvents.

At the end of the day, I do still think that image-rs/image-tiff#250 is a way forward, where the decoding functionality is exposed without dependence on a specific reader (or creating a mem_reader) <- I think there's still quite some design needed and would rather build on top of a fork where the needed functionality from image-tiff is carved out/discovered, rather than re-implementing. That would mean putting quite a large maintenance burden here (since image-tiff is likely to change).

1. Build on top of an Extensibility: mid-level API and harmonizing encoder/decoder structs. image-rs/image-tiff#250 fork, making minimal changes
2. aiocogeo does the async reading from object_store backend
3. georust/geotiff adds geo functionality which is backend-agnostic
   1. Maybe create a trait TiffDecoder in here for a backend and then implement the reader on top of that trait, so that we can easily switch out backends <- This actually doesn't work so well, since one is async and the other sync

Thanks @weiji14 for reading my meta and discord concerns! Indeed, I would like to be able to direct my (free time/thesis) efforts more in a direction I want to. Then, making a reader that secretly has these license terms that is then used here would seem a bit weird, would such a thing be accepted here (say, I PR a dependency for aiocog, having a goodness-clause)?

[HeroicKatora]

How should it read tags? Read all images, including all tags Dos-style? I think it makes sense to read only the relevant tags for the current overview (thinking COG) and geo tags.

Comparing to a recent development in png: it might be simpler for core data structures in image-tiff to support EAGAIN / ErrorKind::WouldBlock. That is, if the underlying doesn't have the necessary buffer filled and can not fill it synchronously, it signals to the decoder. Then the decoder updates some list of required data ranges, let's the caller schedule some form of progress on it, and then continue. Run this in a loop until decoding progresses completely. We might call this a sans-IO style but it's not forbidden to do synchronous IO so it's a little bit of a hybrid. This presents a solution to the buffer problem: we don't need to read tags sequentially, nor do we need to read everything up-front. With a large enough buffer quite a lot of values may complete at the same time, and the decoder could also signal multiple required ranges at once.

The hard design question here is finding the right interface to communicate the needs within the file, and properly control resource usage. (We want to avoid the need for buffers to have arbitrary size, the decoder must work with a constant bound of additional in-memory data besides the requested tags and decoded data for medical images and GeoTIFF). The interfaces are a little brittle since the control flow is quite inverted. Maybe there's a better way to write that as a trait with higher-ranked-trait-bounds now.