`h2::Error` hides IO error

[konstin]

In uv, we want to retry streaming downloads on connection reset errors. These errors are common e.g. in GitHub Actions when downloading e.g. a 20MB file.

Our goal is to have a streaming download-and-unpack. We have a loop in which we try the streaming download-and-unpack, and if it fails, inspect the error chain on whether it looks like a retryable network error (https://github.com/astral-sh/uv/blob/22f80ca00d3f5af7d087154e998c886bbea8cee1/crates/uv-python/src/downloads.rs#L945-L993). This is easy for status code errors, where we get a reqwest::Error at the top of the error chain, but during the stream, the error may be nested in several layer of other crates involved in the unpacking, see the example error below.

We want to specifically retry io::ErrorKind::ConnectionReset. But when trying err.downcast_ref::<io::Error>() we don't get anything, as the h2 error wraps the IO error, which is not apparent from the debug output.

Tar(
    TarError {
        desc: "failed to unpack `/home/konsti/.local/share/uv/python/.temp/.tmpBHaUXj/python/bin/python3.12`",
        io: Custom {
            kind: Other,
            error: TarError {
                desc: "failed to unpack `python/bin/python3.12` into `/home/konsti/.local/share/uv/python/.temp/.tmpBHaUXj/python/bin/python3.12`",
                io: Custom {
                    kind: Other,
                    error: reqwest::Error {
                        kind: Decode,
                        source: reqwest::Error {
                            kind: Body,
                            source: Error {
                                inner: ErrorImpl {
                                    kind: Body,
                                    cause: Some(
                                        Error {
                                            kind: Io(
                                                Kind(
                                                    ConnectionReset,
                                                ),
                                            ),
                                        },
                                    ),
                                },
                            },
                        },
                    },
                },
            },
        },
    },
)

I'm also interested in ways to make our transient network error detection (https://github.com/astral-sh/uv/blob/f88aaa8740824b03be36c7ff106a139151452165/crates/uv-client/src/base_client.rs#L983-L1067) more reliable in general, without going through reqwest's error chain.

[seanmonstar]

It's possible to get the IO error from h2::Error::get_io().

Note that reqwest and especially hyper do not consider the error source chain "stable", so any code downcasting any of the source errors should be resilient to the order changing, or the internal version upgrading.

[konstin]

Would it be possible to get some stable error view, or would you accept PRs for an API such as Error::is_likely_transient?

With both our users and our own CI, we found retries in network operations to be essential. We're seeing a lot of ConnectionReset and similar errors that we need to catch, and ideally we wouldn't rely on implementation details of the error chain for it.