I think this article is supposed to commend Python as a better choice than Bash? I think the author forgot to make a convincing argument for that, though.
> The point of bash-bashing is to reduce use of the shell.
That's a tautology. I get it, we're supposed to reduce the use of bash ... but why?
> Without much real work, it’s easy to replace shell scripts with Python code.
But you're writing the same thing again? There needs to be more reason than this.
> The revised code is easier to read and maintain, runs a little faster, and can have a proper unit test suite.
Three claims. Any data to support it?
Some of the examples of "bash is bad" are also not convincing to me. Here's an example:
> An example of a shell obscurity is the way the current working directory is set. The cd command is clear enough, but in the presence of sub-shells using (), can make it difficult to discern a stack of nested shell invocations and how the working directory changes when the sub-shells exit.
So...why do you have "a stack of nested shell invocations" that all need to be working directory aware? And how would this be any better in Python? (It could be better in either Python or Bash but either one still requires the developer to avoid simple mistakes.)
I don't know. I don't get it. I feel like the author had the unfortunate experience of knowing more Python than Bash and inheriting someone else's (who also didn't have much Bash experience) crappy Bash code.
In my experience, it trips me up no more than Python or Fabric do. I have used Fabric a lot too. At another job, we built all of our pipeline tooling in it.
I somehow keep returning to Bash, though. I enjoy it.
Interesting, I used fabric many years ago but thought it had been deprecated. However, it still seems remote focused. Is that a problem for a shell replacement?
I've written extensive programs in Haskell, and yet I also have many Bash scripts.
I go back and forth. For the longest time I ignored Perl and just wrote scripts in C. Then I learned Perl, then Python. I prefer Ruby for no better reason than it's less boring. These scripting languages share a common strength: nested hash tables make everything easier. Still, I'd revert to Bash, explaining that if you can't accomplish the same thing with virtual text files in Bash you don't really understand Bash.
Nevertheless, switching from Bash to Ruby is like taking off a heavy backpack after a long day hiking. Who needs a joint when the weight is off your shoulders?
There isn't any "data" on these things. It's a qualitative, holistic assessment of the risk-benefit tradeoff between two choices. Like much of life in general, in fact.
Beware of Python devs writing Python shell scripts. They're used to using all the brand new (backwards incompatible) language features as soon as they come out. The vast majority have no consideration for a 10 year old PC running a 10 year old OS. The python3.x they write will not run.
Bash, surprisingly, gets backwards incompatible language features too. But the vast majority of bash developers are writing for all computers, not just computers with software from the last 3 years.
Python language could work if you stuck to an old target. But python culture is a problem and makes this very unlikely. What python is, and what python devs deal with, changes constantly. That doesn't work for shell scripts where stability is king.
Yeah, but nobody's targeting Python 3.0; I think the point was that the Python crowd is likely to exclusively target current versions and lean on 3.6 and under being officially EOL.
If you are running a 10-year old version of python 3, and you try and install some recently written python 3 application, it will never work - you'll be stuck in dependency hell, even getting pip to run may be a challenge due to outdated certs or TLS versions.
I'm fully in favor of replacing shell scripts with Python3 scripts wherever possible. In fact that's part of my job. That being said, the author of this article is using confusing / wrong terminology to discuss bash.
This article says that, "The [bash] shell isn't a complete programming language". While I agree that bash lacks any real data types besides strings, bash is a Turing-complete programming language[1], so it's theoretically possible to write _any_ program in bash that can be written in Python. It might be a hell of a lot uglier and lack things like imports, modules, etc., but it can be done.
For example, there is an implementation of an HTTP daemon written purely in bash[2]. Once again-- terrible idea, but great execution.
I was hoping someone finally had a library or something to make python good at the things shell does well (mostly, running stuff and piping output around). If only.
This really just reads like someone who doesn't know shell and does know python, but thinks the problem is the tool. They continually describe shell with phrases like "obscure and difficult-to-predict" while talking through things that are either obvious with experience or from context, and then write something different in python (i.e. something you could write in BASH that does something different) that's not less obscure but is only using python obscurity.
> An example of a shell obscurity is the way the current working directory is set. The cd command is clear enough, but in the presence of sub-shells using (), can make it difficult to discern a stack of nested shell invocations and how the working directory changes when the sub-shells exit.
Er... yeah? So you expect PWD to be a global and it was actually a local?
Or more explicitly:
> One of the shell’s ickier features is that variables tend to be global. There are some exceptions and caveats, however, that lead to shell scripts that are broken or behave inconsistently.
That's not a shell feature, that's just how people frequently write it - you can make functions and declare your variables local if you want. I can easily write a python script with all my variables at the top, too.
> I was hoping someone finally had a library or something to make python good at the things shell does well (mostly, running stuff and piping output around). If only.
I haven't heard of xonsh either till now. After looking it over, it seems like a perfect integration of sh with Python. There's a good chance I'll give an honest shot at migrating from zsh to xonsh this week.
I agree with your sentiment. People like to complain about tools. The nice things about bash (and even sh) are that you can combine a bunch of simple tools together quickly because the syntax doesn't get in the way of it. The language designers didn't intend for people to be writing thousands of lines of scripts; they intended them to be using them locally and focused on interactivity. So naturally the syntax is nice for interactive work. It's a stretch to pretend that working with python paths is easier than working with shell paths, just like it's a stretch to say your bash machine learning library is better than a python one. The author should let things excel at what they're good at and stop pretending everything needs to be unified.
> I was hoping someone finally had a library or something to make python good at the things shell does well (mostly, running stuff and piping output around). If only.
Even if you have a great library that abstracts away shell like things, you’d still probably be better served with a shell script. Shell scripts aren’t supposed to be big. They are quick one offs that solve immediate problems. This means that they are also spread around the filesystem. I’m of the opinion that one shouldn’t install libraries into the system Python path. I also tend to make heavy use of virtual envs. Setting up a venv for each off scripts doesn’t seem like a great plan.
For more centralized workflows, I could see Python shell replacements being helpful. Especially project specific scripts where the overhead of managing a venv is negligible.
But realistically replacing the majority of bash scripts? I just don’t see it.
After going over it again, I think I can be more precise: What annoys me is that it's inconsistent to the point of being disingenuous. Like... You want to say error handling is hard to do right? That's fine, it is. You wanna write "Checking the status of programs using $? can be accidentally left off, leading to inconsistent behavior" and then in the same article write in explicit exception handlers, and act like somehow those can't be forgotten!? Get lost.
> I was hoping someone finally had a library or something to make python good at the things shell does well (mostly, running stuff and piping output around). If only.
Good news, they did! Even better news: it doesn't require python, it's cross platform, and supports an easy-to-read verbose style for scripts but a quick-to-write terse style for interactive usage. It's an absolute joy to use for nested data structures, everything is an object and it has full native tab completion.
Unfortunately it came out of Microsoft so a lot of people haven't given it the time of day yet. It's honestly one of the joys of modern CLI usage though.
I wrote this to simplify just that for my own tools; the subprocess module that comes with the batteries has a very general interface, i think that it is a bit complex for a quick script.
My objective was to get an abstraction, for a one line process run and extraction of the result, similar to what we had in Perl5 with the system library function. https://perldoc.perl.org/functions/system
Also the shell is impractical, when it comes to slightly more complex programs. There is a limit on what you can do with pipes. Maybe that's the reason why perl is that flexible, as they tried to bridge both realms: Perl had to be useful as a replacement for the quick shell like script, and to be useful as a general purpose programming language.
We had some problem, roughly, "run a tool on all YAML files in the directory" or something. The first attempt that led to a production outage:
(It is late, bash examples are approximate. Don't sweat the syntax.)
the_tool *.yaml
There weren't any YAML files, so argv[1] was literally "*.yaml", which then wasn't found, leading to errors, etc.
shopt -s nullglob
the_tool *.yaml
(Don't even get me started on how there's shopt -s and set -o and WTF.)
There weren't any YAML files, so argc was == 0, which the program considered an error. Mea culpea, perhaps.
AN_ARRAY=(*.yaml)
if [[ < length of the array is not 0 > ]]; then
the_tool "${AN_ARRAY[@]}"
fi
This failed, b/c bash can't represent an empty array! (And we run -u, as one must, if one wants reasonable behavior.)
So finally the contorted,
AN_ARRAY=(*.yaml)
if [[ "${AN_ARRAY+x}" = x && < length is not 0 > ]]; then
the_tool "${AN_ARRAY[@]}"
fi
To say "you're holding it wrong" is just folly. Bash is worthy of the same treatment that "A Fractal of Bad Design" gave PHP, and it has no place in anything that wants to call itself engineering.
Use it on the command line in your day to day, fine; but if it's getting automated, jump to at least Python.
On the contrary, I am 100% open to suggestions that BASH sucks. (Personally, I would have picked the inadequacies of `set -e` as my favorite ... unfortunate aspect.) My objection was very narrow: this specific article does a poor job at saying why shell sucks, and a worse job still at presenting Python as a superior alternative. That's not to say that good reasons to use something else don't exist, and it's not even to say that Python can't be that alternative. (Your sibling comments suggesting xonsh and https://amoffat.github.io/sh/ paint a fairly compelling way to actually move to python without losing the advantages of shell, for instance.)
>To say "you're holding it wrong" is just folly. Bash is worthy of the same treatment that "A Fractal of Bad Design" gave PHP, and it has no place in anything that wants to call itself engineering.
If you have a problem that involves "run a command with a list of parameters but not if the list is empty" and you don't immediately think of `xargs -r`, then consider that it's because you don't understand the tools, not because the tools are bad.
This is reasonable in general, though in this case, I think this will replicate the problem of the second Bash version, as the problem there is with the_tool requiring at least one argument.
While dethanatos's point about Bash's default behavior with a non-matching glob is on target, the remainder of the example demonstrates a problem in the_tool, which will need to be worked around regardless of how you invoke it.
> If you have a problem that involves "run a command with a list of parameters but not if the list is empty" and you don't immediately think of `xargs -r`, then consider that it's because you don't understand the tools, not because the tools are bad.
`-exec` would run the command for each file separately, which isn't what the original problem wanted, and I didn't want to assume that it would be equivalent.
Bash is amazing at being expressive and allowing for a ton of common things sysadmins / Devops / power users need to do with just one or a few lines of code vs a “proper” language that might require libraries, compiling stuff, or just a ton more SLOC + time debugging. If you don’t know what a glob is going to match, try it first. Understand the semantics. If you need portability, do that too, but a ton of what I end up using bash for is along the lines of slicing up poorly formatted inconsistent lists and running tools against stuff, possibly in conjunction with args from those lists, and for that it’s amazing. Stuff like jq helps a lot, and tools like xmlstarlet before that were great too. I remember trying to rewrite a script that was getting long and AWKward in Python and getting pretty frustrated with how explicit and error-prone it was there. Sure, the result was probably more robust, but it came at the expense of an order of magnitude more dev / debugging time. That’s not always necessary; sometimes “good enough” is good enough, and usually bash is good enough.
But it's like, I have to turn my head around and stop doing one thing (globs), and switch over to another thing (find, xargs). The language has baited me down one way of doing it, when I shouldn't be doing it that way. I'm well aware of xargs (but not the -r flag). I typically avoid xargs, as our code has to¹ work on macOS & Linux, and xargs is one of those utilities whose available options differ between the two for even simple use cases. (Such as here, for the "does it exec on no args?".)
But it still proves the point: the equivalent Python is just so much more straight-forward, easier to read, and not apt to shoot me at a moments notice, and doesn't have bugs. A reader only familiar with Python's syntax can read and understand Python, and can write Python that is correct. Not so with bash, you have to understand, "oh, no, globs will screw you here, you need this flag from this utility" — no, it's lunacy.
Like, I can go down the road you're suggesting (and in the past, I did) and keep learning the nooks and crannies and intricacies of bash, each one adding to the nightmare fuel that is my understanding of shell. I can keep tweaking the script around various failure modes, letting each line fail in prod 3 different ways. Or, I can just re-write it in Python, or Rust, or any other sane language and have it never fail again, for reasons that would also either apply to bash, or apply and have worse consequences than bash. (E.g., the Python would raise, bash would just plow on with it.)
Your response is exactly the same as my response before I saw the light on PHP, C, etc.: "you're holding it wrong". No, it's a bad tool that guiles the user into holding it wrong. There are better tools that are easier to hold.
¹b/c despite all work being done for Linux, devs get MBPs, b/c the company can't manage more than exactly 1 type of machine, for everyone.
> Is it really right to not process any files just because one of them just disappeared?
No idea what you mean here, the glob can expand to one or many.
> In general, situations like this suffer from race conditions and time-of-check/time-of-use problems.
I mean, that's a problem in every version presented so far, one that I don't think can be fixed externally without something extra like hardlinking to a new directory, because "the_tool" apparently can't handle missing files and they can go missing when "the_tool" starts up. That's why I didn't even bother.
What happens if someone runs `the_tool this-file-does-not-exist.yaml` or the yaml file is removed by some other means before you get a chance to process it? Does your world also come tumbling down in these cases? If yes, please stop blaming Bash for it.
Personally, I never use bash for scripting. I use the system's scripting shell, Debian's Dash or NetBSD Ash. Rhetorical question: If Bash was great for shell scripting, why doesn't the OS project choose it for scripting. When I compile the OS kernel and userland, the project (NetBSD) uses shell scripts not Python. Not suggesting the shell is necessarily better but I am suggesting it works and people understand it.
There are many ways to run a command on every .yaml file in a directory. Personally I like using xtrace. Something like:
echo cd the_directory > 1.sh
echo "test -f *.yaml" >> 1.sh
ls -1 *.yaml|sed 's/^/the_tool /' >> 1.sh
chmod +x 1.sh
less 1.sh
sh -ex 1.sh
As mostly a day-to-day shell user who automates some things, I do not use Python. Never had an outage.
> To say "you're holding it wrong" is just folly. Bash is worthy of the same treatment that "A Fractal of Bad Design" gave PHP, and it has no place in anything that wants to call itself engineering.
is this a cloud AMD64 opinion that ignores the 2 trillion devices which Gartner predicts will never run python?
I recently finished a ~600 lines ash script and the kids at work which just graduated complained that *"Y u no python?" For them it was difficult to understand because apparently I avoided cat/grep/dirname/basename/etc and used shell syntax wherever possible to avoid shelling out. (the thing I did would only run when system reaches a certain load)
Everyone forgot that the target hardware has only busybox (ash) and was massively resource constrained. My only options were to implement what I was doing as an eBPF, in C, or shell (ash). The shell script took 4 days (and was refactored/optimized several times within that period), the C (or eBPF) program would have taken ~10 days and the eBPF which is the best solution is unmaintainable once I leave the company.
from a performance perspective I take a bash script any time. Also python programs over time have a risk of eventually pulling in a dozen dependencies.
Judging by the reaction of the kids at work, what you did is already unmaintainable.
Plus, if you are running shell out of necessity on a constrained system, that doesn't mean shell deserves to be a programming language where there's actual choice.
how so? there is no python or perl or ruby etc on the target. what would have been a maintainable option other than /bin/sh ? C/C++ ?
> where there's actual choice.
again what choice? :) as I said there is no other solution than what I've given. the system literally doesn't have anything else regardless what "I think" it should support.
You asked how bash is performant. I responded with reasons its performance is good in many applications. Here”s one more: bash starts up faster than other interpreters (e.g., Python), and startup time can often be a significant factor in overall performance.
This tool of yours can not cope with files not existing, and causes production outages? Yet somehow you manage to blame bash for this? The outcome would not have been different if systemd or python started your tool.
In fact, handling that types of errors is what bash is suitable for. "If file is readable, process it" is not an unusual construct. Please do take care however to avoid races if this directory is a spool directory that can be populated at any time, in which case it is useful that rename() is atomic, sometimes an additional working directory is required. Again, this is how the operating system is designed and nothing the shell can gloss over.
I have seen many problems from previously unknown race conditions when concurrency is increased. Scripts have to be kept composable and trivial to understand in order to guard against that.
To be honest, I would argue to remove any script that looks like the last example above. It is not clear what side effects of that bizarre array construct are useful to the system. Keep scripts as simple and readable as possible, a trivial truth perhaps, applicable to every language, but one that needs to be reiterated until the end of our profession.
I mean, I think there's something to that. Bash has a lot of arcane rules; I personally never write an `if` statement right, what with the `[[` vs `[` and having to surround the condition with whitespace. I find that bash scripts carry a million little gotchas that will cause variables to not evaluate/expand the way you'd expect.
Say what you will about Python, but it's much easier to reason about what a script will do as compared to bash (unless the bash is _extremely_ simplistic).
As much as I appreciate Python, I much rather use Perl as a replacement for complex shell scripts. Perl 5 is ubiquitous, it has been stable for the best part of the last 20 years and it's unbeatable for all those tasks that involve heavy text processing.
Python is good for complex, structured applications, but shell scripts are not that. Shell scripts are glue, and there's arguably no better glue than Perl.
I first learned the shell, and pretty much stuck to it conservatively avoiding bash extensions.
Later, I learned perl (around version 4)
my first impression of perl was annoyance.
$foo = 123;
seemed silly because there was a dollar sign on the left side of the assignment.
and... $foo, @foo, $', $_, s/abc/def/;
all the syntax seemed needlessly cryptic and reinforced the "perl is a write-only language" idea in my mind.
But then something happened.
At some point all of these idioms went away as I became fluent and could think in perl.
Because there were many ways to do something in perl, I found that it
become VERY easy to get an idea in my head out into working code.
So perl was to me my most expressive language.
But what I've noticed is that many people haven't overcome the syntax barrier
and perl looks like line noise to them. More worrying is the fact that everyone
can get their ideas into perl, but because people solve problems so differently
their perl can be vastly different from what I am accustomed to.
So a few years later I learned python. The indentation requirement was a minor
annoyance, but that was quickly overcome by the consistency and visual structure
it added.
A more major annoyance was that many things were harder to implement in python.
Regular expressions is a huge one - they are a major feature of perl and I had
learned to use them liberally.
Yet I persisted and with help from the many "batteries included" with python, I
was easily writing portable maintainable scripts in python.
And they were still readable after 6 months.
Now I write only small shell scripts, and past a certain size all other scripting
is in python. It's quite easy to write meaningful scripts in python. (I've pretty much lost my fluency in perl via python -- I fumble around now looking at or modifying perl scripts)
by the way argparse is quite easily my favorite import in python.
In my experience, bash makes the simple things simple and the hard things (like complex pipelines, validating arguments, or “pure bash”) really hard. YMMV.
Python makes everything medium difficulty which can be a big win depending on the size of automation required.
I myself kind of went full circle. I started mostly python, then started to really embrace bash, mainly because I joined a company where everyone else used it a ton, got pretty decent at it. Even wrote basically a primitive container orchestrator in Bash because my PM didn't want to use anything "newfangled". Then I joined a startup at basically the ground floor, and it's almost entirely pure python, with some Makefiles to automate common operations. All that to say, I have a lot of experience with both paradigms.
And frankly...bash kinda sucks as a programming language and environment. It's a footgun factory. Even when I was most fluent, I constantly had to deal with nested quotes, escaping, stringly types, empty variables, implicit behavior.
Python's biggest weak point is definitely its packaging and dependency ecosystem... but it at least has one. Bash doesn't even have modules. And it's gotten MUCH better over the years, with pyproject.toml, poetry, actual version resolvers, pipx makes managing venvs for cli tools way easier.
"but bash is everywhere" says many folks. yeah, and? There was a point in time where it wasn't. Python is seeing way more market penetration.
Xonsh (python based shell) and plumbum (python library which makes pipe-operating and subprocessing easier) are both great.
I once worked for a company that had a product which had a fairly complicated and time-consuming build process. Someone had written a distributed build process using the traditional CMD.exe shell scripting language. It was running multiple processes in parallel, and they communicated by writing and reading text files. Later they added PowerShell scripts. All in all, there were tens of thousands of lines of PS and BAT scripts. It was quite a nightmare.
The team was supposed to maintain this mess. We decided to create a compiler that translated .BAT scripts to C#. The idea was that the C# code would be easier to understand and modify. I left before it was complete, so I'm not sure how that worked out. But last I heard, they were pleased that some parts of it now ran thousands of times faster.
Im dealing with some python3 crap today on my main laptop.
Primarily, because of a library I compiled and installed, I'm getting other nasty side effects form it. Primarily that protonvpn-cli isnt now running becasue of some broken library.
I've never had bash break like that. Quirks, sure. But never this sort of brokenness.
I had a similar problem with the YouTube plugin for Kodi recently. Turned out to be some weird Python 3.10 regression that took over a month to fix.
For ProtonVPN you can also use the OpenVPN config files they provide to avoid any Python client issues. Ironically I have a Bash script written around that to generate individual ovpn files as needed from the zip for myself.
mkdir -p would create the directory without error. checking if a directory exists is trivial before creating it. Python throws an error if the path given to os.mkdir exists too.
This article is way too long to digest, so i should stop now.
Perhaps the argument should be, rewrite hacky things in bash as programs and build them into your apps?
I've written a lot of bash scripts in my time, even some 10,000 liners. Christ - I have probably written over 100k lines of them. The big ones (by big I mean say over 100 lines) all started life when I thought they would remain under 100 lines.
I'm a professional programmer by trade. Admitting to starting a program I know would be over 100 lines in shell script would be close to an admission of incompetence, because technically shell script is one of the worst computer languages out there. So I'd never admit to it. But it's so damned portable, and ubiquitous and the batteries it comes with (the 'nix cli tools) are to powerful and complete, it makes irresistible at times.
Where this article falls down is they are recommending Python3. Python2 would be fine. But in 'nix environments, where file names and configuration files can't be treated as text (Unicode) because there is no well defined system encoding, Python3 manages to introduce more rare bugs than shell script. It encourages you to treat everything a text, the dies ignominiously when decoding the 'nix byte stream (file name or whatever) fails. (On Windows where everything is UTF16, this isn't a problem - for local files. It remains a problem for data from external systems, like the internet.)
Pull off making a programming language less reliable than shell was a mean achievement - but the Python3 devs did it. Hats off to 'em.
>This article is way too long to digest, so i should stop now.
Lol one of the top problems with having "scripts" in python, their reliability-lifetime" (a.k.a as execute-and-forget) is very limited. Packages are outdated, api changes etc etc
I have some bad news for you, from a long, long, long time CS/EE/Sysadmin person:
I can probably run my bash script on every linux I ever touched.
I can probably have that python code break on half of the linux boxen I use. This one does not have module X installed. This one is too old, this one is too new, this python is not holding its mouth just at the right angle to work.
Yeah, bash bash all you want. I have pulled out scripts decades old and run them. I can not say the same for other things I have written.
... Hilariously, I think shell might actually win that; neither is installed by default, but you get some sort of bash/sh with any of interix, WSL1/2, git, cygwin, or mingw.
isn't WSL like a hypervisor that runs a guest? I'm clueless about Windows but when I start a WSL it is usually some form of Debian or Ubuntu that I can download from the Microsoft "store". Therefore isn't the shell you have whatever comes with that guest?
yeah python (and before perl with cpan) had these issues that maintaining a consistent system is more effort than writing the damn thing in the first place. I loved Perl but tried to avoid modules like the plague. And python is almost worse because the same mentality (avoid modules) was abandoned by most of the community so now you get incompatibilities with what the OS provides and what gets installed via pip.
Perhaps it's unfair to pick at python for this when many of these issues are not the language but because people picking python when they should have just used shell script.
I have the same issue with bash scripts. One host has jq, another is missing it. One host has yq version 2 with completely different syntax than yq version 4.
If you switch from Linux to *nix where you introduce BSD variants (like macOS) it gets even worse. Python dependency management leaves much to be desired but bash has none.
What are standard tools and shell features? Bash + GNU? Those might be consistently installed on common Linux distros but that doesn't work across *nix systems (Linux, macOS, BSD). That probably won't work against less-common Linux setups like anything embedded (busybox or container images) either
"Moving furniture by throwing it in the back of a pickup truck is obviously bad. Here's how to do it with a forklift and 18-wheeler instead, because this is definitely better."
Places that I've implemented shell/utility scripts in Python tend to do so because their other application code was also in Python and Python has much more testing support than bash scripts will ever have.
That's why I've done it before, to write tools like custom log rotating / clean-up stuff; another project that comes to mind was a data pipeline that used some Python regex stuff to clean up non-printable characters and ugly stuff from text files before importing data. Both of these projects had lots of tests to verify that the code did what we wanted it to do.
I decided a while ago to write all my future scripts in Python whenever possible instead of sh or bash due to the overall better syntax and power of Python.
Typer is one of the better packages for running Python scripts from CLI. Fabric & Invoke are pretty good too.
To run any bash commands in Python I have a Python script that creates a temporary bash file with the code to execute, makes it executable, runs it with the commands in the built-in subprocess package, then deletes the file. This makes any complex bash commands with piping and whatnot runnable straight from Python.
This way I don't have to learn all the different bash-based Python commands from pathlib and whatnot that throw unexpected errors, and I can run them using pure bash syntax from Python. But then you also get all the benefits of Python's looping syntax, classes etc.
OK, I'll bite... The author would complain about my python3 in the same way and conclude it would be better to write it in my Bash.
Without questioning the "why" of the script, sans my customary morning coffee, and without having tested the code below.
Maybe I'd do something like this.
For the small price of function invocation overhead, the nice benefit of doing it this way is that one can `source` the functions into one's Bash shell and use each one as a standalone Unix tool, complete with tabtab completion, pipelines etc.
#!/usr/bin/env bash
stop_app() {
pkill ${1:?Fail. App name required.}
}
today() {
date +%Y%m%d
}
analyse() {
local analyser=${1:?Fail. Analyser script name required.}
local out_dir=$(printf "results_%s" $(today))
while yaml_file
do local out_file="${out_dir}/summary_$(basename yaml_file).txt"
python3 ${analyser} ${yaml_file} > ${out_file}
printf "%s\n" ${out_file}
done
}
exec_analytics() {
local analyser=${1:?Fail. Analyser script name required.}
local source_dir=${2:-"~/Documents/ExtFin-EFS/smoke/"}
find ${source_dir} -type f -name *.yaml |
sort -r |
analyse ${analyser} |
tail -1
}
update_current() {
local latest_outfile=${1:?Fail. Provide latest output file.}
ln -sf ${latest_outfile} "current.txt"
}
And maybe one can invoke it like...
stop_app "whatever_app" && (
trap "rm -f /tmp/module_design_analytics_outfile" 0 HUP TERM PIPE INT
if exec_analytics module_design_analytics.py |
tail -1 > /tmp/module_design_analytics_outfile
then update_current $(printf /tmp/module_design_analytics_outfile)
echo "Done"
else echo "Oops. Something went wrong."
fi
trap - 0 HUP TERM PIPE INT
)
The only real alternatives to bash are shit-tier languages like python and perl.
I will never use a programming language which has significant white space, specially if I may have to view and edit the scripts in a remote terminal with vi.
137 comments
[ 4.4 ms ] story [ 168 ms ] threadPrevious discussion: https://news.ycombinator.com/item?id=14998213
> The point of bash-bashing is to reduce use of the shell.
That's a tautology. I get it, we're supposed to reduce the use of bash ... but why?
> Without much real work, it’s easy to replace shell scripts with Python code.
But you're writing the same thing again? There needs to be more reason than this.
> The revised code is easier to read and maintain, runs a little faster, and can have a proper unit test suite.
Three claims. Any data to support it?
Some of the examples of "bash is bad" are also not convincing to me. Here's an example:
> An example of a shell obscurity is the way the current working directory is set. The cd command is clear enough, but in the presence of sub-shells using (), can make it difficult to discern a stack of nested shell invocations and how the working directory changes when the sub-shells exit.
So...why do you have "a stack of nested shell invocations" that all need to be working directory aware? And how would this be any better in Python? (It could be better in either Python or Bash but either one still requires the developer to avoid simple mistakes.)
I don't know. I don't get it. I feel like the author had the unfortunate experience of knowing more Python than Bash and inheriting someone else's (who also didn't have much Bash experience) crappy Bash code.
Shell is great for one liners to run commands, cancerous for anything more complex. Fabric combines the best of both worlds: shell for one liners, Python for more complex logic https://docs.fabfile.org/en/2.6/getting-started.html#addendu...
I somehow keep returning to Bash, though. I enjoy it.
I go back and forth. For the longest time I ignored Perl and just wrote scripts in C. Then I learned Perl, then Python. I prefer Ruby for no better reason than it's less boring. These scripting languages share a common strength: nested hash tables make everything easier. Still, I'd revert to Bash, explaining that if you can't accomplish the same thing with virtual text files in Bash you don't really understand Bash.
Nevertheless, switching from Bash to Ruby is like taking off a heavy backpack after a long day hiking. Who needs a joint when the weight is off your shoulders?
Otherwise they'd just be plain old compiled "languages."
There isn't any "data" on these things. It's a qualitative, holistic assessment of the risk-benefit tradeoff between two choices. Like much of life in general, in fact.
(Sorry, Oingo Boingo)
Bash, surprisingly, gets backwards incompatible language features too. But the vast majority of bash developers are writing for all computers, not just computers with software from the last 3 years.
Python language could work if you stuck to an old target. But python culture is a problem and makes this very unlikely. What python is, and what python devs deal with, changes constantly. That doesn't work for shell scripts where stability is king.
Python3 is just under 14 years old.
This article says that, "The [bash] shell isn't a complete programming language". While I agree that bash lacks any real data types besides strings, bash is a Turing-complete programming language[1], so it's theoretically possible to write _any_ program in bash that can be written in Python. It might be a hell of a lot uglier and lack things like imports, modules, etc., but it can be done.
For example, there is an implementation of an HTTP daemon written purely in bash[2]. Once again-- terrible idea, but great execution.
[1]: https://en.wikibooks.org/wiki/Bash_Shell_Scripting#A_few_not...
[2]: https://github.com/avleen/bashttpd
edit: added newline to separate references
It's not pure bash, as it relies on netcat or socat plus some other external programs (ls, tree, cat, date).
There does exist a pure bash httpd though. It relies on a loadable builtin (from the bash tree though not built by default).
https://github.com/dzove855/Bash-web-server
This really just reads like someone who doesn't know shell and does know python, but thinks the problem is the tool. They continually describe shell with phrases like "obscure and difficult-to-predict" while talking through things that are either obvious with experience or from context, and then write something different in python (i.e. something you could write in BASH that does something different) that's not less obscure but is only using python obscurity.
> An example of a shell obscurity is the way the current working directory is set. The cd command is clear enough, but in the presence of sub-shells using (), can make it difficult to discern a stack of nested shell invocations and how the working directory changes when the sub-shells exit.
Er... yeah? So you expect PWD to be a global and it was actually a local?
Or more explicitly:
> One of the shell’s ickier features is that variables tend to be global. There are some exceptions and caveats, however, that lead to shell scripts that are broken or behave inconsistently.
That's not a shell feature, that's just how people frequently write it - you can make functions and declare your variables local if you want. I can easily write a python script with all my variables at the top, too.
xonsh (https://xon.sh/)
Been using it for a few years now. It's worth it.
I've only kicked the tires on both, but plumbum:
https://plumbum.readthedocs.io/en/latest/
and sh: https://amoffat.github.io/sh/
both feel pretty good in that regard.
For more centralized workflows, I could see Python shell replacements being helpful. Especially project specific scripts where the overhead of managing a venv is negligible.
But realistically replacing the majority of bash scripts? I just don’t see it.
If not, there are copycats like Ruby and Python . . .
Good news, they did! Even better news: it doesn't require python, it's cross platform, and supports an easy-to-read verbose style for scripts but a quick-to-write terse style for interactive usage. It's an absolute joy to use for nested data structures, everything is an object and it has full native tab completion.
Unfortunately it came out of Microsoft so a lot of people haven't given it the time of day yet. It's honestly one of the joys of modern CLI usage though.
https://pypi.org/project/subb/
I wrote this to simplify just that for my own tools; the subprocess module that comes with the batteries has a very general interface, i think that it is a bit complex for a quick script.
My objective was to get an abstraction, for a one line process run and extraction of the result, similar to what we had in Perl5 with the system library function. https://perldoc.perl.org/functions/system
Also the shell is impractical, when it comes to slightly more complex programs. There is a limit on what you can do with pipes. Maybe that's the reason why perl is that flexible, as they tried to bridge both realms: Perl had to be useful as a replacement for the quick shell like script, and to be useful as a general purpose programming language.
(It is late, bash examples are approximate. Don't sweat the syntax.)
There weren't any YAML files, so argv[1] was literally "*.yaml", which then wasn't found, leading to errors, etc. (Don't even get me started on how there's shopt -s and set -o and WTF.)There weren't any YAML files, so argc was == 0, which the program considered an error. Mea culpea, perhaps.
This failed, b/c bash can't represent an empty array! (And we run -u, as one must, if one wants reasonable behavior.)So finally the contorted,
To say "you're holding it wrong" is just folly. Bash is worthy of the same treatment that "A Fractal of Bad Design" gave PHP, and it has no place in anything that wants to call itself engineering.Use it on the command line in your day to day, fine; but if it's getting automated, jump to at least Python.
If you have a problem that involves "run a command with a list of parameters but not if the list is empty" and you don't immediately think of `xargs -r`, then consider that it's because you don't understand the tools, not because the tools are bad.
While dethanatos's point about Bash's default behavior with a non-matching glob is on target, the remainder of the example demonstrates a problem in the_tool, which will need to be worked around regardless of how you invoke it.
find has -exec, xargs isn't necessary.
No?
But it's like, I have to turn my head around and stop doing one thing (globs), and switch over to another thing (find, xargs). The language has baited me down one way of doing it, when I shouldn't be doing it that way. I'm well aware of xargs (but not the -r flag). I typically avoid xargs, as our code has to¹ work on macOS & Linux, and xargs is one of those utilities whose available options differ between the two for even simple use cases. (Such as here, for the "does it exec on no args?".)
But it still proves the point: the equivalent Python is just so much more straight-forward, easier to read, and not apt to shoot me at a moments notice, and doesn't have bugs. A reader only familiar with Python's syntax can read and understand Python, and can write Python that is correct. Not so with bash, you have to understand, "oh, no, globs will screw you here, you need this flag from this utility" — no, it's lunacy.
Like, I can go down the road you're suggesting (and in the past, I did) and keep learning the nooks and crannies and intricacies of bash, each one adding to the nightmare fuel that is my understanding of shell. I can keep tweaking the script around various failure modes, letting each line fail in prod 3 different ways. Or, I can just re-write it in Python, or Rust, or any other sane language and have it never fail again, for reasons that would also either apply to bash, or apply and have worse consequences than bash. (E.g., the Python would raise, bash would just plow on with it.)
Your response is exactly the same as my response before I saw the light on PHP, C, etc.: "you're holding it wrong". No, it's a bad tool that guiles the user into holding it wrong. There are better tools that are easier to hold.
¹b/c despite all work being done for Linux, devs get MBPs, b/c the company can't manage more than exactly 1 type of machine, for everyone.
Is it really right to not process any files just because one of them just disappeared?
In general, situations like this suffer from race conditions and time-of-check/time-of-use problems.
No idea what you mean here, the glob can expand to one or many.
> In general, situations like this suffer from race conditions and time-of-check/time-of-use problems.
I mean, that's a problem in every version presented so far, one that I don't think can be fixed externally without something extra like hardlinking to a new directory, because "the_tool" apparently can't handle missing files and they can go missing when "the_tool" starts up. That's why I didn't even bother.
There are many ways to run a command on every .yaml file in a directory. Personally I like using xtrace. Something like:
As mostly a day-to-day shell user who automates some things, I do not use Python. Never had an outage.is this a cloud AMD64 opinion that ignores the 2 trillion devices which Gartner predicts will never run python?
I recently finished a ~600 lines ash script and the kids at work which just graduated complained that *"Y u no python?" For them it was difficult to understand because apparently I avoided cat/grep/dirname/basename/etc and used shell syntax wherever possible to avoid shelling out. (the thing I did would only run when system reaches a certain load)
Everyone forgot that the target hardware has only busybox (ash) and was massively resource constrained. My only options were to implement what I was doing as an eBPF, in C, or shell (ash). The shell script took 4 days (and was refactored/optimized several times within that period), the C (or eBPF) program would have taken ~10 days and the eBPF which is the best solution is unmaintainable once I leave the company.
from a performance perspective I take a bash script any time. Also python programs over time have a risk of eventually pulling in a dozen dependencies.
Plus, if you are running shell out of necessity on a constrained system, that doesn't mean shell deserves to be a programming language where there's actual choice.
how so? there is no python or perl or ruby etc on the target. what would have been a maintainable option other than /bin/sh ? C/C++ ?
> where there's actual choice.
again what choice? :) as I said there is no other solution than what I've given. the system literally doesn't have anything else regardless what "I think" it should support.
Typically, time spent interpreting control flow in a bash script is dwarfed by time spent in the programs called by the script.
> It is literally interpret by lines
It's just splitting lines into words to build up commands. It's not compiling C++.
> starting a whole process per if statements..
In the common pattern `if command1; then command2; fi`, if command1 is a builtin (e.g., `test` or `[`), then bash does not create a subprocess.
In fact, handling that types of errors is what bash is suitable for. "If file is readable, process it" is not an unusual construct. Please do take care however to avoid races if this directory is a spool directory that can be populated at any time, in which case it is useful that rename() is atomic, sometimes an additional working directory is required. Again, this is how the operating system is designed and nothing the shell can gloss over.
I have seen many problems from previously unknown race conditions when concurrency is increased. Scripts have to be kept composable and trivial to understand in order to guard against that.
To be honest, I would argue to remove any script that looks like the last example above. It is not clear what side effects of that bizarre array construct are useful to the system. Keep scripts as simple and readable as possible, a trivial truth perhaps, applicable to every language, but one that needs to be reiterated until the end of our profession.
It absolutely can, I have no idea where you’re getting this idea.
I mean, I think there's something to that. Bash has a lot of arcane rules; I personally never write an `if` statement right, what with the `[[` vs `[` and having to surround the condition with whitespace. I find that bash scripts carry a million little gotchas that will cause variables to not evaluate/expand the way you'd expect.
Say what you will about Python, but it's much easier to reason about what a script will do as compared to bash (unless the bash is _extremely_ simplistic).
I wouldn’t want to worry about python environment when dealing with OS post-install scripting for instance.
Python is good for complex, structured applications, but shell scripts are not that. Shell scripts are glue, and there's arguably no better glue than Perl.
I first learned the shell, and pretty much stuck to it conservatively avoiding bash extensions.
Later, I learned perl (around version 4)
my first impression of perl was annoyance.
$foo = 123;
seemed silly because there was a dollar sign on the left side of the assignment. and... $foo, @foo, $', $_, s/abc/def/;
all the syntax seemed needlessly cryptic and reinforced the "perl is a write-only language" idea in my mind.
But then something happened.
At some point all of these idioms went away as I became fluent and could think in perl.
Because there were many ways to do something in perl, I found that it become VERY easy to get an idea in my head out into working code.
So perl was to me my most expressive language.
But what I've noticed is that many people haven't overcome the syntax barrier and perl looks like line noise to them. More worrying is the fact that everyone can get their ideas into perl, but because people solve problems so differently their perl can be vastly different from what I am accustomed to.
So a few years later I learned python. The indentation requirement was a minor annoyance, but that was quickly overcome by the consistency and visual structure it added.
A more major annoyance was that many things were harder to implement in python. Regular expressions is a huge one - they are a major feature of perl and I had learned to use them liberally.
Yet I persisted and with help from the many "batteries included" with python, I was easily writing portable maintainable scripts in python.
And they were still readable after 6 months.
Now I write only small shell scripts, and past a certain size all other scripting is in python. It's quite easy to write meaningful scripts in python. (I've pretty much lost my fluency in perl via python -- I fumble around now looking at or modifying perl scripts)
by the way argparse is quite easily my favorite import in python.
https://github.com/Mylab6/PiBluetoothMidSetup
Of course I could of done this in bash, but Python is just so much cleaner.
Python makes everything medium difficulty which can be a big win depending on the size of automation required.
And frankly...bash kinda sucks as a programming language and environment. It's a footgun factory. Even when I was most fluent, I constantly had to deal with nested quotes, escaping, stringly types, empty variables, implicit behavior.
Python's biggest weak point is definitely its packaging and dependency ecosystem... but it at least has one. Bash doesn't even have modules. And it's gotten MUCH better over the years, with pyproject.toml, poetry, actual version resolvers, pipx makes managing venvs for cli tools way easier.
"but bash is everywhere" says many folks. yeah, and? There was a point in time where it wasn't. Python is seeing way more market penetration.
Xonsh (python based shell) and plumbum (python library which makes pipe-operating and subprocessing easier) are both great.
The team was supposed to maintain this mess. We decided to create a compiler that translated .BAT scripts to C#. The idea was that the C# code would be easier to understand and modify. I left before it was complete, so I'm not sure how that worked out. But last I heard, they were pleased that some parts of it now ran thousands of times faster.
Primarily, because of a library I compiled and installed, I'm getting other nasty side effects form it. Primarily that protonvpn-cli isnt now running becasue of some broken library.
I've never had bash break like that. Quirks, sure. But never this sort of brokenness.
For ProtonVPN you can also use the OpenVPN config files they provide to avoid any Python client issues. Ironically I have a Bash script written around that to generate individual ovpn files as needed from the zip for myself.
mkdir -p would create the directory without error. checking if a directory exists is trivial before creating it. Python throws an error if the path given to os.mkdir exists too.
This article is way too long to digest, so i should stop now.
Perhaps the argument should be, rewrite hacky things in bash as programs and build them into your apps?
I'm a professional programmer by trade. Admitting to starting a program I know would be over 100 lines in shell script would be close to an admission of incompetence, because technically shell script is one of the worst computer languages out there. So I'd never admit to it. But it's so damned portable, and ubiquitous and the batteries it comes with (the 'nix cli tools) are to powerful and complete, it makes irresistible at times.
Where this article falls down is they are recommending Python3. Python2 would be fine. But in 'nix environments, where file names and configuration files can't be treated as text (Unicode) because there is no well defined system encoding, Python3 manages to introduce more rare bugs than shell script. It encourages you to treat everything a text, the dies ignominiously when decoding the 'nix byte stream (file name or whatever) fails. (On Windows where everything is UTF16, this isn't a problem - for local files. It remains a problem for data from external systems, like the internet.)
Pull off making a programming language less reliable than shell was a mean achievement - but the Python3 devs did it. Hats off to 'em.
Lol one of the top problems with having "scripts" in python, their reliability-lifetime" (a.k.a as execute-and-forget) is very limited. Packages are outdated, api changes etc etc
I can probably run my bash script on every linux I ever touched.
I can probably have that python code break on half of the linux boxen I use. This one does not have module X installed. This one is too old, this one is too new, this python is not holding its mouth just at the right angle to work.
Yeah, bash bash all you want. I have pulled out scripts decades old and run them. I can not say the same for other things I have written.
Exactly THIS !
Perhaps it's unfair to pick at python for this when many of these issues are not the language but because people picking python when they should have just used shell script.
If you switch from Linux to *nix where you introduce BSD variants (like macOS) it gets even worse. Python dependency management leaves much to be desired but bash has none.
"Moving furniture by throwing it in the back of a pickup truck is obviously bad. Here's how to do it with a forklift and 18-wheeler instead, because this is definitely better."
That's why I've done it before, to write tools like custom log rotating / clean-up stuff; another project that comes to mind was a data pipeline that used some Python regex stuff to clean up non-printable characters and ugly stuff from text files before importing data. Both of these projects had lots of tests to verify that the code did what we wanted it to do.
Like testing (Bats sucks).
Like string manipulation (use other utils or even languages callable from Bash).
Like complicated logic. Just don't do that in Bash.
Bash is for (glorified) one-liners. And for I-can't-believe-you-did-that-in-Bash-ers.
Typer is one of the better packages for running Python scripts from CLI. Fabric & Invoke are pretty good too.
To run any bash commands in Python I have a Python script that creates a temporary bash file with the code to execute, makes it executable, runs it with the commands in the built-in subprocess package, then deletes the file. This makes any complex bash commands with piping and whatnot runnable straight from Python.
This way I don't have to learn all the different bash-based Python commands from pathlib and whatnot that throw unexpected errors, and I can run them using pure bash syntax from Python. But then you also get all the benefits of Python's looping syntax, classes etc.
Without questioning the "why" of the script, sans my customary morning coffee, and without having tested the code below.
Maybe I'd do something like this.
For the small price of function invocation overhead, the nice benefit of doing it this way is that one can `source` the functions into one's Bash shell and use each one as a standalone Unix tool, complete with tabtab completion, pipelines etc.
And maybe one can invoke it like...I will never use a programming language which has significant white space, specially if I may have to view and edit the scripts in a remote terminal with vi.
I am also not too interested in using a programming language which looks the same before and after encryption: https://www.goodreads.com/quotes/tag/437174-perl-the-only-la...