Best practices for classifying transient reqwest::Errors?

[dbolduc]

Hey, I own some client libraries that are built on top of reqwest. Let's say I have an idempotent request. I attempt to make it and receive a reqwest::Error. What logic should I use to determine if the error is retryable (e.g. a transient connection error, or some incomplete frame).

I see #2811, and am aware of the retry module, but I need to handle the retry loop (backoffs, total number of attempts, etc.) in my layers. So that will not work for me.

We are currently using this logic, but our customers report it is too conservative (googleapis/google-cloud-rust#5864).

fn should_retry(e: &reqwest::Error) -> bool {
    if err.is_connect() {
        return true;
    }
    match err.status() {
        Some(code) => is_retryable_code(code),
        None => false,
    }
}

I have also found a reqwest-middleware crate, with this logic, which seems more complete: https://github.com/TrueLayer/reqwest-middleware/blob/614b9474f6bec85c8660e4d52b8d9f12f8359229/reqwest-retry/src/retryable_strategy.rs#L138-L184

Would you recommend using reqwest-middleware's implementation?

Do you have any other suggestions on best practices here?