Or go the other direction: stop trying to do fancy things and write simpler code that avoids errors.
#!/bin/sh
[ "${DEBUG:-0}" = "1" ] && set -x
set -u
foo="$( my-external-program | pipe1 | pipe2 | pipe3 )"
if [ -z "$foo" ] ; then
echo "Error: I didn't get any output; exiting!"
exit 1
fi
echo "Well I got something back. Was it right?"
if ! printf "%s\n" "$foo" | grep -q -E 'some-extended-regex' ; then
echo "Error: '$foo' didn't match what I was looking for; exiting!"
exit 1
fi
echo "Do the thing now..."
A lot of programs will either produce valid output as STDOUT, or if they encounter an error, not produce STDOUT. So for the most part you just need to 1) look for any STDOUT at all, and then 2) filter it for the specific output you're looking for. For anything else, just die with an error. If you need to find out why it didn't run, re-run with DEBUG=1.
Advanced diagnosis code won't make your program work better, but it will make it more complicated. Re-running with tracing enabled works just as well 99% of the time.
I actually tried rewriting this in Python, but gave up since Python's startup latency is atrocious if you have even a few imports (and using a socket to a pre-existing server is fundamentally unable to preserve enough process context related to the terminal). Perl would probably be a better fit but it's $CURRENTYEAR and I've managed to avoid learning Perl every year so far, and I don't want to break my streak just for this.
The Bash code is not only fast but pretty easy to understand (other than perhaps the header, which I never have to change).
I think you should be able to get rid of the is-oil part, because set -E was implemented last year
$ osh -c 'set -E; set -o |grep errtrace'
set -o errtrace
I'd be interested in any bug reports if it doesn't behave the same way
(The Oils runtime supports FUNCNAME BASH_SOURCE and all that, but there is room for a much better introspection API. It actually has a JSON crash report with a shell stack dump, but it probably needs some polish.)
(I'm sure this is lovely Bash, but for all the people who rejected Perl for its modem line noise vibe...what do ya think of this?)
As an aside, I actually wonder if Bash's caller() was inspired by Perl's.
There is also Carp and friends, plus Data::Dumper when you not only need the stack trace but also the state of objects and data structures. Which is something that I don't think Bash can really do at all.
Bash has associative arrays, but just POSIX shells like ash do not have associative arrays.
And, POSIX shell can only shift and unshift on the $@ array; so it would be necessary to implement hashmaps or associative arrays with shell string methods and/or eval.
Are you asking me to defend shell script syntax, or are you criticizing this except from a shell script?
The awk and printf are as obscure and unreadable as Perl, but still probably faster than just starting Perl.
Ironically, in terms of portability, it's probably more likely that awk and printf are installed than Python (or Perl). This application doesn't need Python in the (devcontainer) container, and nobody does sysadmin scripts with lua (which can't `export VARNAME` for outer shells) so shell scripting is justified though indeed arcane.
Getopt is hardly more understandable than a few loops through $@ with case statements.
I don't understand the relevance of other tools to "getting decent error reports in Bash"?
There are great logging (and TAP testing) libraries in Python, but that doesn't solve for debugging Bash?
Perl had numerous magic,
one/two letter variables and copious regexp use, which is what gave it the line noise insult. Wtf is $@ and $_? After spending enough time breathing Perl you got a feel and could actually read it, but take a break and forget all the things, and it's back to being inscrutable. Meanwhile, the bash above isn't totally inscrutable. It's not python clear, and I don't know what SHLVL is without looking it up, but the rest is clear enough (awk with NR, isn't self explanatory, but unfortunately for me I've used too much awk).
Yes. This also means that if you use a third-party shell library which uses “trap” internally (like shunit2), you can’t use “trap” in your own script at all.
Yeah but each only exists once. If you wanna clean up on ERR and also use that trap for debugging you need to implement some janky layer on top of the trap.
(FWIW my take away from issues like this is always: Bash is not a serious programming language. If you are running up against these limitations in real life it's time to switch language. The challenge is really in predicting when this will happen _before_ you write the big script!)
Take care that set -o pipefail will not work on older dash (including IIRC the current Ubuntu LTS), and neither will set -o pipefail || true if set -e is in effect. (For some reason that I’m too lazy to investigate, a failing set invocation will crash the script immediately rather than proceed into the second branch.) The best I could think of to opportunistically enable it was to use a subshell:
if (set -o pipefail 2>/dev/null); then set -o pipefail; fi
Or you can just target bash, I guess.
(I rather dislike shellcheck because it combines genuine smells with opinions, such as insisting on $(...) instead of `...`. For the same reason, with Python I regularly use pyflakes but can’t stand flake8. But to each their own.)
It is not that `set -e` is bad, it is that bash is a bit weird in this area and you have to know when things eat errors and when they don't. This is not really changed by `set -e`: you already had to know them to make safe code. `set -e` does not wave a magic wand saying you don't have to understand bash error control.
But having `set -e` is almost universally better for people who do not understand it (and I would argue also for people who do). Without it you are responsible for implementing error handling on almost every line.
As other have already said: this is one of those things that generally pushes me to other languages (in my case often Python), as the error handling is much more intuitive, and much less tricky to get right.
One of my major nitpicks with set -e isn't from bash... It's from poorly behaved programs you might run. Ones that don't exit with defined exit codes (non-zero on error!).
> Ones that don't exit with defined exit codes (non-zero on error!).
My favorite is `docker images`.
It outputs... not-empty when the output is empty. It returns zero (success) when output is not-empty and empty. So you have to both do error checking and parse its output before you know whether there are any matching images in the cache.
Contrast that with grep. You can do `if grep -q foo` or `if grep -qv foo`, you can do `if "foo" == "$(grep foo)"`, etc. Much more versatile and easy to use.
Then there's apps that report to stderr that some argument isn't used but then also exits with success without doing things. Then there's similar apps that do the same thing but... exits after also doing things. Ugh.
After some time you get to learn to test your apps before you write up scripts around them. That's a good thing though, it means that you know how to mix together a diverse toolset.
Well, maybe. But using `set -euo pipefail` here does not make it any worse as far as i understand? The script still does broken things, but the more correct way to be safe is not broken by set -euo pipefail
Is that a safety point of view? Is shell supposed to be input-safe? I may have limited shell skills but it doesn’t seems like it’s designed to be safe.
You can just do
should do the same thing. Also you should set "errtrace" (-E) and possibly "nounset" (-u) and "pipefail".or even use caller to print a full backtrace: https://news.ycombinator.com/item?id=44636927
That's neat, but if your bash script needs a backtrace it should not be a bash script. To each their own though.
My bash scripts don't "need" a backtrace, but it sure is nicer than not having one
FWIW, I've grown the following which handles a few more cases. For some reason I wasn't aware of `caller` ...
What the hell. This is cool and all but I'm looking at it as a signal I should move up one tier in language (eg: to Perl, PHP, Python or Ruby)
Or go the other direction: stop trying to do fancy things and write simpler code that avoids errors.
A lot of programs will either produce valid output as STDOUT, or if they encounter an error, not produce STDOUT. So for the most part you just need to 1) look for any STDOUT at all, and then 2) filter it for the specific output you're looking for. For anything else, just die with an error. If you need to find out why it didn't run, re-run with DEBUG=1.Advanced diagnosis code won't make your program work better, but it will make it more complicated. Re-running with tracing enabled works just as well 99% of the time.
> A lot of programs will either produce valid output as STDOUT, or if they encounter an error not produce stdout
Lots of programs produce nothing in the success case and only print in the failure case.
Ah yes, the old "just write non-buggy code, or at least code whose bugs are deterministic".
My solution just silently sits in the background for unexpected, unpredictable bugs.
I actually tried rewriting this in Python, but gave up since Python's startup latency is atrocious if you have even a few imports (and using a socket to a pre-existing server is fundamentally unable to preserve enough process context related to the terminal). Perl would probably be a better fit but it's $CURRENTYEAR and I've managed to avoid learning Perl every year so far, and I don't want to break my streak just for this.
The Bash code is not only fast but pretty easy to understand (other than perhaps the header, which I never have to change).
PHP maybe? Or in limited cases, AWK? But I'd definitely learn Perl, it's a gem.
PHP needs to be installed, Perl is dead and AWK is more limited than Bash
> PHP needs to be installed
There are some ways around this:
https://github.com/crazywhalecc/static-php-cli
For what little it’s worth, Perl is very much not dead.
Maybe try ruby, or you could use go (yeah, have to compile)
I think you should be able to get rid of the is-oil part, because set -E was implemented last year
I'd be interested in any bug reports if it doesn't behave the same way(The Oils runtime supports FUNCNAME BASH_SOURCE and all that, but there is room for a much better introspection API. It actually has a JSON crash report with a shell stack dump, but it probably needs some polish.)
>I'd be interested in any bug reports
What's the point? You can't fix them anyway
They meant Oil(s) as in fixing bugs in the bash compatible replacement that they author for the OP's 'is-oil' check.
I know
Setting PS4 gets decent error reports with `set -x` (and `set -x -v`; `help set`).
Here's an excerpt that shows how to set PS4 from a main() in a .env shell script for configuring devcontainer userspace:
This, too:(I'm sure this is lovely Bash, but for all the people who rejected Perl for its modem line noise vibe...what do ya think of this?)
As an aside, I actually wonder if Bash's caller() was inspired by Perl's.
There is also Carp and friends, plus Data::Dumper when you not only need the stack trace but also the state of objects and data structures. Which is something that I don't think Bash can really do at all.
There are no objects in bash. There are indexed and associative arrays and both can be iterated over like so:
or with the help of the keys: If you want to dump the data of any variable you can just use declare -p and you get something like this: What you can do, if you have a set of variables and you want them to be "dumped", is this:Let's "dump" all variables that start with "BASH":
Or one could do something like this: But the declare option is much more reliable as you don't have to test for the variable's type.Bash has associative arrays, but just POSIX shells like ash do not have associative arrays.
And, POSIX shell can only shift and unshift on the $@ array; so it would be necessary to implement hashmaps or associative arrays with shell string methods and/or eval.
Are you asking me to defend shell script syntax, or are you criticizing this except from a shell script?
The awk and printf are as obscure and unreadable as Perl, but still probably faster than just starting Perl.
Ironically, in terms of portability, it's probably more likely that awk and printf are installed than Python (or Perl). This application doesn't need Python in the (devcontainer) container, and nobody does sysadmin scripts with lua (which can't `export VARNAME` for outer shells) so shell scripting is justified though indeed arcane.
Getopt is hardly more understandable than a few loops through $@ with case statements.
I don't understand the relevance of other tools to "getting decent error reports in Bash"?
There are great logging (and TAP testing) libraries in Python, but that doesn't solve for debugging Bash?
There is at least one debugger for Bash scripts.
Perl had numerous magic, one/two letter variables and copious regexp use, which is what gave it the line noise insult. Wtf is $@ and $_? After spending enough time breathing Perl you got a feel and could actually read it, but take a break and forget all the things, and it's back to being inscrutable. Meanwhile, the bash above isn't totally inscrutable. It's not python clear, and I don't know what SHLVL is without looking it up, but the rest is clear enough (awk with NR, isn't self explanatory, but unfortunately for me I've used too much awk).
But trap doesn't "stack" (like e.g. defer in Go) so if you do this it's not available for other purposes like cleanup
Yes. This also means that if you use a third-party shell library which uses “trap” internally (like shunit2), you can’t use “trap” in your own script at all.
not sure what you mean, you can have separate ERR and EXIT traps that run independently.
Yeah but each only exists once. If you wanna clean up on ERR and also use that trap for debugging you need to implement some janky layer on top of the trap.
(FWIW my take away from issues like this is always: Bash is not a serious programming language. If you are running up against these limitations in real life it's time to switch language. The challenge is really in predicting when this will happen _before_ you write the big script!)
Why don’t all shells just do this?
Perhaps you underestimate just how many scripts are poorly written and part of your operating system.
For what it's worth, I think `set -euo pipefail` should be default for every script, and thoroughly checked with shellcheck.net.
Take care that set -o pipefail will not work on older dash (including IIRC the current Ubuntu LTS), and neither will set -o pipefail || true if set -e is in effect. (For some reason that I’m too lazy to investigate, a failing set invocation will crash the script immediately rather than proceed into the second branch.) The best I could think of to opportunistically enable it was to use a subshell:
Or you can just target bash, I guess.(I rather dislike shellcheck because it combines genuine smells with opinions, such as insisting on $(...) instead of `...`. For the same reason, with Python I regularly use pyflakes but can’t stand flake8. But to each their own.)
> such as insisting on $(...) instead of `...`.
Only one of those can be (sanely) nested. Why would you ever want to use backticks?
Looks slick. Works on older shells too
That's a pitfall №60
https://mywiki.wooledge.org/BashPitfalls#set_-euo_pipefail
I have never liked this statement of the problem.
It is not that `set -e` is bad, it is that bash is a bit weird in this area and you have to know when things eat errors and when they don't. This is not really changed by `set -e`: you already had to know them to make safe code. `set -e` does not wave a magic wand saying you don't have to understand bash error control.
But having `set -e` is almost universally better for people who do not understand it (and I would argue also for people who do). Without it you are responsible for implementing error handling on almost every line.
As other have already said: this is one of those things that generally pushes me to other languages (in my case often Python), as the error handling is much more intuitive, and much less tricky to get right.
One of my major nitpicks with set -e isn't from bash... It's from poorly behaved programs you might run. Ones that don't exit with defined exit codes (non-zero on error!).
> Ones that don't exit with defined exit codes (non-zero on error!).
My favorite is `docker images`.
It outputs... not-empty when the output is empty. It returns zero (success) when output is not-empty and empty. So you have to both do error checking and parse its output before you know whether there are any matching images in the cache.
Contrast that with grep. You can do `if grep -q foo` or `if grep -qv foo`, you can do `if "foo" == "$(grep foo)"`, etc. Much more versatile and easy to use.
Then there's apps that report to stderr that some argument isn't used but then also exits with success without doing things. Then there's similar apps that do the same thing but... exits after also doing things. Ugh.
After some time you get to learn to test your apps before you write up scripts around them. That's a good thing though, it means that you know how to mix together a diverse toolset.
Well, maybe. But using `set -euo pipefail` here does not make it any worse as far as i understand? The script still does broken things, but the more correct way to be safe is not broken by set -euo pipefail
What about for `/bin/sh`, i.e. posix compliant shells like dash?
Only set -eu is defined in POSIX and supported by dash.
> For what it's worth, I think `set -euo pipefail` should be default for every script, and thoroughly checked with shellcheck.net.
This
Automatically leaking the line number and command, even to stderr is not a sane default.
Is that a safety point of view? Is shell supposed to be input-safe? I may have limited shell skills but it doesn’t seems like it’s designed to be safe.
Why? It's not like you couldn't get the source of the running script anyway.
Because shells weren’t supposed to be doing complex logic. People use shells to do way way way more than they should.
beware of using this. any operations that returns non-zero exit status will cause immediately exit
For example `((n++))` where n is 0. I’m used to it enough that I usually remember to add `|| :` to the end of lines like this.