summarize: Don't panic on broken pipe

[sharnoff]

I hit #31 while piping summarize to head. The original issue was in part due to prettytable panicking. That's since been upgraded to a version with the fix, but some of the println!s that happen after table.printstd() will also panic from the broken pipe:

$ summarize summarize ... | head
...
thread 'main' panicked at library/std/src/io/stdio.rs:1165:9:
failed printing to stdout: Broken pipe (os error 32)
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

To fix this, change the println!s to writeln!s, and ignore any errors.

Fixes #31.

[bjorn3]

I think you should just exit the process on a broken pipe rather than keep writing data to nowhere.

[sharnoff]

Makes sense -- implemented a catch-all "exit on failure" in 59876d9.

Do you think it should special-case broken pipe to return a visible error otherwise? or better to just give up

[bjorn3]

Do you think it should special-case broken pipe to return a visible error otherwise?

Yes, I think it should still show an error and exit when an error other than a broken pipe occurred.

[sharnoff]

Fair enough, implemented in 75fde69

[Kobzol]

What do you think about just configuring SIGPIPE to abort the process? For example like this. I think that's a more universal solution than patching a few print lines.

[ChrisDenton]

Non-unix platforms may not have signals (e.g. windows).

[Kobzol]

Right, but I thought that this issue only happens on Unix? Does Rust also somehow ignore sigpipe (or something similar) on Windows? Is that even a thing?

[ChrisDenton]

Closing a pipe on Windows just causes further writes to fail. The println macro will panic on any failure to write so it will panic on a closed pipe.

[Kobzol]

I see. Well, hard to say how many people redirect CLI commands into pipes on Windows with summarize 😆

[ChrisDenton]

Sure, I don't know how much of an issue it is in practice.

Incidentally rustc has a safe_print macro for handling this across platforms. It still panics but with a payload that tells rustc to treat it as a normal (albeit fatal) error rather than an unexpected panic (this works because rustc uses catch_unwind to catch all panics).

[bjorn3]

Why does println!() even produce a panic message for broken pipes? Can't it just silently panic?

[ChrisDenton]

"silently panic"?

[bjorn3]

std::panic::resume_unwind(Box::new(SomeMarkerType)) will unwind without producing any panic message.

[ChrisDenton]

Ah. Likely because nobody has proposed it to libs-api yet.