On error the decoder should emit everything it decoded instead of swallowing it

[orium]

On error the decoder should emit everything it decoded instead of swallowing it. In particular, the case I'm interested in fixing, is when the gzip footer is not present. Currently the last "chunk" is never emitted.

I'm working on a PR right now.

[NobodyXu]

Does that mean we should actually advance the output in #451 while skipping advancing of input?

[orium]

Does that mean we should actually advance the output in #451 while skipping advancing of input?

Yes, but that's not sufficient: we also need to make sure the Decoder returns Ok(()) before returning an error on finish(). For instance gzip already consumed the input before returning "unexpected EOF" if the footer is not present. I'll update #454, but I'm seeing another difference in behavior between 0.4.32 and 0.4.39 that I need to investigate.

[NobodyXu]

Thank you for your work!

That definitely makes sense

[orium]

Actually, I think it's better to not advance neither the input nor the output like I've like done in #451. That keeps a reasonable invariant: the input consumed matches the output emitted. What I want to solve in this issue is the case where the Decoder iterates multiple times and decodes a bunch of data. It then find an error and that data already decoded is lost because we simply return a Ready(Err(e)). So in #454, if I find an error and there some stuff was already decoded, we return that we read that successfully and then fail in the next call to Decoder::do_poll_read()

[NobodyXu]

I will review the PR once it's ready and I will have a look at it t see if it makes sense

Without some context and testing, it's really hard to tell just by looking at API