Pin protobuf<4

[ashnair1]

Tests were failing due to the recent release of protobuf 4.21.0. See protocolbuffers/protobuf#10051

Error propagation: protobuf > tensorboard > pytorch_lightning > torchgeo

[adamjstewart]

Thanks! Hopefully we can remove this soon.

[adamjstewart]

Looks like 4.21.0 was yanked so maybe the protobuf devs understand that this was a bad idea: https://pypi.org/project/protobuf/#history

If you rerun the tests on your other PR it might pass without needing this PR. Not sure if protobuf or tensorboard will get a new release first.

[ashnair1]

Not sure what's going on with this test failure. I could add the protobuf entry under pip in the environment.yml but I'm not sure that's the issue here.

Per the discussion in the linked issue, it might be a good idea to pin protobuf=3.21.1 since it seems the devs still plan to push a 4.20.1

[adamjstewart]

Yeah, definitely pinning to <4 is a good idea until our deps fix themselves. This issue affects just about every library that depends on protobuf (tensorboard, torch, and likely others).

[adamjstewart]

I also have no idea why the Windows tests are failing. We don't use environment.yml in those tests so it can't be related to that.

[adamjstewart]

Can you rebase and also add this to requirements.txt?

[ashnair1]

Ran into a failed coverage test (no idea why that happened). Pushed again to see if error reappears.

[adamjstewart]

No idea what's going on with coverage. You have all the same commits as main, so it's not like new code has been added and the absolute number of lines of coverage have changed...

Also, you'll probably want to add protobuf to the ignore list in .github/dependabot.yml so that dependabot doesn't try to update it.

There is a new release of protobuf now that hasn't been yanked, but there's also a new release of pytorch-lightning that pins protobuf to an older version, so we're technically good for now. PyTorch has also pinned to older protobuf and 1.12.0 seems poised for a release soon. Hopefully we won't need this for long.

[ashnair1]

There is a new release of protobuf now that hasn't been yanked, but there's also a new release of pytorch-lightning that pins protobuf to an older version, so we're technically good for now. PyTorch has also pinned to older protobuf and 1.12.0 seems poised for a release soon. Hopefully we won't need this for long.

Wouldn't we then need to update the min version of pytorch-lightning and/or pytorch? Seems like a big jump from 1.3 to 1.6.5 and 1.7.0 to 1.12.0 respectively.

[adamjstewart]

No need to change the minimum versions, TorchGeo is still compatible with both the older and newer releases. Someone might need to manually install an older version of protobuf to get PyTorch or tensorboard working, but that's not a TorchGeo issue.