Once upon a time, while fiddling with a log processing shell pipeline, it hit me that the UNIX Way is a Surprisingly Functional Way, and so Functional Programming (FP) and Bash must be a natural fit. They fit. The world was never the same again.
Now I believe it so much, that I will go on a limb and assert that it is highly inappropriate to write imperative-style Bash when we can just as easily write supremely functional Bash. Why? Because it makes for supremely better Bash (more reusable, more composable, more scalable, more enjoyable).
Yes, I truly believe.
Yes, I'm sane.
No? Well OK, humour me…
Obligatory mea culpa
Because I'm am going to lazy-stream this in N parts. Because my original post was growing to "never gonna ship" size. But not because "Bash ain't a real programming language". (Besides, in our post-reality world, we get to make our own reality.)
Nothing here will be novel.
- I've not invented anything that follows. There are too many influences to enumerate fully. I'll provide references as I go along.
- I expect to revise, correct, add to this series as I learn more over time.
- Code in the post assumes Bash 4+, because that's what I've been using over the last 8-odd years.
Your Mileage May Vary.
- The Bash code will be both message and medium. but I will describe general FP thinking, and Bash sure isn't the only medium. Please replicate solutions in your favourite language (in a UNIX-like way)!
- I won't go crazy with Bash-isms, so the ideas and most of the code should generalise to most UNIXy shells. That said, I haven't used other shells to make equally confident claims about FP in them.
FP suffuses my very being.
- And I'm fairly confident the FP ideas will translate broadly, because I use them all the time; in my code (Clojure, APL, Ruby, JS, Python, SML…), in my designs for logging systems, infra-as-code systems, CI/CD systems, as well as designing human/communication workflows for teams.
- But it's possible I've lived my life all wrong.
An outline of the "N" parts
This (zeroth) post is about why the UNIX Way is the way of functional-style design 1.
N more posts are brewing, with examples and techniques. Likely one per topic:
A rad example from 1986 to motivate the rest of the series.
Deep-dive into bash functions and function design techniques
- Using functions to craft one's own Bytes-sized UNIX tools
- Using them interactively like regular UNIX tools
Pipelining all the things
- How we automatically get map / filter / reduce / early termination
- Automatic streaming (regular pipes, tee, named pipes etc…)
- Ways to do pipeline-friendly domain design, and to translate that into pipeline-friendly functions.
Avoiding manual state management with intelligent use of:
- Variables, scopes, program invariants
- Command substitution
- Process substitution
- Templating with heredocs and herestrings
- Trickshots with things like
- Reasonable uses of pattern matching
Environment isolation in detail
- Lists and sub-shells
- in Pipelines
- Session portability
Designing idempotent / restart-friendly solutions
- Because things can and will fail.
- Mainly because I haven't had to 2 write parallel Bash, but it will be fun to mess with.
Maybe sundry topics like associative arrays (Bash-only), job control, namespacing, metaprogramming, flame-bait like "pipes are monads" etc.
Prelude: Seeing the UNIX tools philosophy as a functional design philosophy
The many remarkable aspects of UNIX Nature were discovered over half a century worth of versions, revisions, disasters, and reincarnations. While many avatars of UNICES and UNIX-likes have come and gone, the UNIX Way (articulated by the 1990s) has thrived through the ravages of time. Here it is, embodied in the form of the UNIX Tools Philosophy.
Most importantly, do one thing, and do it well (just like a function).
Consume and emit plain data (just like a function).
Output the same data format as is received at input (formerly only plain lines of text, but now also structured literal data like JSON, EDN etc.)
Don't be chatty (i.e. avoid side-effects, again, just like a function).
Be line-oriented, which design choice turns out to be naturally streaming, with automatic support for map/filter/reduce, which we will use a lot.
Favour universal composition via standard interfaces like file descriptors, standard IN/OUT/ERROR, line-orientation, and UNIX pipes (quite monadic, an argument for much later).
Be as general-purpose as possible for wide reusability, in any context. This pushes tools away from imposing internal structure on input data, as also from maintaining persistent or shared internal state.
Ideally have sane behaviour like environment isolation, idempotence, etc.
Last but not least, when out-of-the-box solutions are not good enough, it encourages us to detour to building our own tools. And these can be simple Bash functions, usable interactively at the command line, just like full standalone programs!
Of course, practice can diverge from the ideal, but not by too much (many tools have to work with stateful objects like files and sockets, some may rely on lock-files, some should be idempotent but aren't, others may grow to do more than one thing and do everything badly etc.). Besides, not even Haskellers escape this reality, so there.
It stands that the UNIX Way strongly encourages us to create laser-focused, composable, purely functional, data-flow oriented programs that we can remix at will into surprisingly powerful solutions with surprisingly little ceremony.
This Way has proven to be very useful at scales several orders of magnitude apart; from in-program 1-liner functions, to 1 kilobyte tools, to operating systems, to planet-wide distributed systems. This unreasonable effectiveness is why UNIX People have long valued these values.
"Screw that, show me your code"
Sorry! I feel ya… no code, no dice. Here is some of my FP-style Bash. I plan to crib liberally from these to illustrate the posts-to-come.
bash-toolkit: a "Swiss Army Toolkit" of functions I've been accumulating over the years.
oxo: a retro-style noughts and crosses game in Bash (and it speaks!).
Next up: Shell ain't a bad place to FP: part 1/N in which we take apart Douglas McIlroy's famous pipeline from 1986, to motivate the rest of the series. "Take apart" in the sense of "Design is about taking things apart.". A most respectful sense.
May the Source be with us.
There was the one time I could have, at a $DAYJOB, but I was quite green, and had deadline, and it was a one time log analysis thing, and I a large EC2 box to waste, which I hogged for half a day, and came away stunned that my crappy shell pipeline chewed through ~600 GiB (gzipped) without crashing anything.↩︎