Racket->Rhombus: To Sexp or not to Sexp?
#racket #lisp #meta
Recently, I have had the pleasure of speaking with the venerable Matthew Flatt of Racket fame, who happens to be visiting NUS on sabbatical.
As part of his visit, he gave a brief seminar on one of the latest projects that he and some other Racket developers (racketeers?) have been working on - a little Racket language called Rhombus.
If you're not familiar with Rhombus, in short, it's effectively an indentation-based dialect of Racket that looks a bit like Python1:
;; Racket code (for/list [(x (list "a" "b")) (y (list "c" "d"))] (string-join (list x y) "")) ;; '("ac" "bd")
# Rhombus code for List: ~each x: ["a", "b"] y: ["c", "d"] x ++ y # ["ac", "bd"]
Whoa now! Hold on a second. Don't leave just yet!
Of course, I know that any self-respecting lispers in the audience are certainly wildly shaking their heads in disapproval. How could they not? Is this not a vile affront to anything that is holy and sacred? or is it?…
Well… while my initial reaction was the same, after I had the chance to spend some time hacking around with Rhombus code, its distinct style actually started to grow on me. While I'm not exactly a convert just yet, I can't deny that Rhombus code ain't without its own elegance and beauty, and I can't exactly say I disliked my time with the language itself.
In the rest of this blog post, I'll provide a quick introduction to Rhombus, look into what makes it different from other Lisps, and provide some reflections on my initial experiences with trying out Rhombus with a side-by-side Racket-vs-Rhombus case-study on a simple "real-world" problem that I was working on. Can a man learn to live without parentheses? Let's find out!
What is Rhombus? and what makes it different from other lisps?
Rhombus can be thought of as a sort of indentation-based Lisp, so the syntax, and its structure follows roughly what you'd expect:
debug:
"hello ~a"
bob
can be thought of as internally being represented as follows:
(debug (block
(stmt "hello ~a")
(stmt bob)))
But it's not quite that simple!
The idea of using indentation to encode structure is a pretty old one, and history is littered with the carcasses of many a Lisp that tried that route.
While many a prior Lisp has tried indentation as a replacement for parentheses, Rhombus actually opts for a slightly more nuanced approach, through a mechanism it calls Shrubbery.
Rather than purely using indentation for grouping, Rhombus' Shrubbery mechanism uses indentation plus select control characters to imbue the input lexical stream with additional structure. This means that Rhombus can support both nested blocks and infix operators quite seamlessly:
def f(x): x + x / 2
Then, following the spirit of its Lisp heritage, Rhombus macros can operate on Shrubbery, at an intermediate level between lexing and parsing, and thereby offer the user a large degree of freedom to tune the language to their liking.
As a result, while Rhombus code is no longer strictly adorned by parentheses like its Lisp brethren, it still has some pretty nice support for meta-programming:
;; Racket macro (define-syntax (debug stx) (syntax-case stx () [(_ msg arg ...) #'(printf msg (list (syntax->datum #'arg) arg) ...)] )) (debug "result: ~a" (+ 1 1)) ;; result: ((+ 1 1) 2)
# Rhombus macro expr.macro | 'encode($arg) $rest ...': let result: "" +& arg values('[$result, $arg]', '$rest ...') expr.rule | 'debug: $msg $arg ...': 'printf($msg, encode($arg) ...)' debug: "result: ~a" (1 + 1) # result: ('(1 + 1)' 2)
Case study on Parenthesis: tools for a more… civilised age?
So clearly, it seems you don't actually need s-exps to have Lisp-like extensible languages.
But, I like them!
Isn't a fear of parentheses merely an affliction that plagues the nascent Lisper?
Why would any experienced Lisp aficionado ever switch to Rhombus?
To investigate these questions in a slightly more scientific way, let's try write a simple grading script in both Racket and Rhombus (my original implementation of this script was in OCaml, and then later python).
In the rest of this section, I'll walk through Racket and Rhombus programs2 to perform this simple grading task, and in the process, we can explore how the features of these two languages influence the way one might code in them.
Problem definition
Our problem is as follows:
- Students can form teams of up to three for two assignments.
- Teams for each assignment should be distinct - i.e a student should have different team members for each assignment.
The data is provided as a collection of csv-files, describing the teams for assignment 1 and 2 using the student-ids of each team-member, and another that contains the class roster, mapping all student-ids to the respective student names and email addresses. The output of our program should specify whether there were any invalid teams and also print out various miscellaneous statistics about the teams.
Step 1: Reading CSVs
As all the inputs are provided as CSV files, it seems natural to start by creating a helper function to encapsulate this process. For bonus points, we can also hardcode in the path to the files in the function, as this can reasonably be expected to stay constant:
(define (read-input input) (let ([path (string-join (list "~/grading" input) "/")]) (call-with-input-file path (lambda (in) (csv->list (make-csv-reader in)))))
def open_csv_file(file): def path: "~/grading/" ++ file with_input_file( path, fun (ic): ic |> csv_reader |> csv_to_list)
As we can see, the Rhombus and Racket ended up looking fairly similar2, although the use of the string concatenation operator in Rhombus significantly reduces the code size.
The |> threading operator isn't a builtin
feature of Rhombus, but rather something that I quickly hacked
together as my heart pined for OCaml:
expr.rule | '$exp |> $f': '$f($exp)' | '$exp |> $f $rest ...': '$f($exp) $rest ...' | '$exp |> $f $rest ...': '$f($exp) $rest ...'
A rather interesting observation to make here is that while I could have implemented a similar threading operator in Racket, I opted not to, because I felt that the extra syntactic overhead required to write the macro would outweigh its benefits, while in Rhombus, the lighter syntax actually made me more inclined to use macros3.
Step 2: Collecting student teams
Using this helper function to read the input csvs, the next step is to extract the raw student teams data into a slightly more useful encoding for analysis, converting the list of csv records into a list of sets of students:
(define (construct-teams csv) (define (row->team team) (define students (map string-trim team)) (list->set students)) (map row->team csv)) (define assignment1-teams (construct-teams (read-input "assignment1.txt"))) (define assignment1-students (apply set-union assignment1-teams)) ;; same for asg 2...
def build_team(raw_team): def team: map(string_trim, raw_team) set.of_list(team) def asg1_teams: map(build_team, open_csv_file("assignment1.txt")) def asg1_students: apply(set.union, asg1_teams) # same for asg 2...
In the end, I wound up writing roughly the same code to extract the
teams. In my opinion, while I disliked the f(arg,...) syntax for applying functions, the Rhombus code was
a lot easier to parse at a glance.
A slight annoyance with Rhombus here was the fact that as commas don't reset the indentation level, I am forced to make all arguments for a function call to be vertically aligned, even when this leads to a slightly less pleasant layout.
Step 3: Building the student roster
As our listing of teams specify the team members by their student-ids, we'll need to also extract the class-roster in order to determine the names and emails of the students in each team:
(struct student-data (name email) #:transparent) (define (build-student-mapping) (define students (read-input "student-roster.csv")) (define mapping (make-hash)) (for ([student students]) (match-define (list name id email) student) (hash-set! mapping id (student-data name email))) mapping) (define student-mapping (build-student-mapping))
def build_student_mapping(): def mapping: MutableMap() for: ~each student: open_csv_file("student-roster.csv") val [name,id,email]: student mapping[id] := [name,email] mapping def student_mapping: build_student_mapping()
Here, we can see the use of Rhombus' for-loop syntax, which merges, folds, iterations and maps all into one unified looping construct4. Additionally, the syntax-sugar for mutable assignments also helps to keep the overhead for map-updates down, making the code easier to understand.
Step 4: Finding invalid teams
Preparations now out of the way, we come to the main meat of this program: the code to actually check whether any students have same team members between assignments.
The algorithm itself is pretty simple: iterate through all teams in assignment 1 and 2, and check that the size of any intersection of teams between assignments is always less than 1:
(define (find-invalid-teams) (define invalid-teams '()) (for ([asgn1-team assignment1-teams]) (for ([asgn2-team assignment2-teams]) (define overlap (set-count (set-intersect asgn1-team asgn2-team))) (when (> overlap 1) (set! invalid-teams (cons asgn1-team invalid-teams)) (set! invalid-teams (cons asgn2-team invalid-teams)) ))) invalid-teams) (define invalid-teams (find-invalid-teams))
def find_invalid_teams(): for values(invalid_teams=[]): ~each a1_team: asg1_teams ~each a2_team: a2_teams ~when set.count(a1_team *&& a2_team) > 1 a1_team :: a2_team :: invalid_teams def invalid_teams: find_invalid_teams()
Okay, so hands down, for this small example, the Rhombus code comes out to be leagues more concise and easy to grep than the Racket one, although, granted, some of the fault is my own inexperience with Racket, as the mutation in the Racket code can be eliminated by more judicious use of Racket's list libraries.
The Rhombus code also makes use of another OCaml-inspired macro to describe cons operations in a slightly more natural form:
expr.rule | '$hd :: $tl $rest ...': 'List.cons($hd, $tl $rest ...)'
The macro itself is a little bit nuanced - in order to capture the
right-associativity of cons, I've implemented the macro as a
transformation that rearranges the lexical stream '$hd :: $tl $rest ...' to the form 'List.cons($hd, $tl $rest ...)' - in other words, everything
after the double-colon actually gets placed inside the second
argument.
For the case of left-associative operators, Rhombus actually provides some simple syntax sugar for quickly defining them:
operator (a *&& b): set.intersect(a,b) operator (a *|| b): set.union(a,b) operator (a *~~ b): set.subtract(a,b) operator (a *^^ b): set.symmetric_difference(a,b)
These set-operators, while a little esoteric, have quite a compound impact, and actually end up not only simplifying this part, but every subsequent step as well.
Step 5: Printing output
Our last step is to print out the results - if there were any invalid teams, then for each invalid team, print out the details of each student in the team:
(define (print-team-details team) (define (print-student-details student) (define student-info (hash-ref student-mapping student)) (printf "\t~a: ~a, ~a\n" student (student-data-name student-info) (student-data-email student-info))) (printf "Team ~a:\n" team) (for ([student team]) (print-student-details student))) (when (> (length invalid-teams) 0) (println "Found invalid teams") (for ([team invalid-teams]) (print-team-details team)))
when List.length(invalid_teams) > 0 | printf("NOTE: found ~a INVALID teams.\n", List.length(invalid_teams)) for: ~each team: invalid_teams printf("Invalid team: ~a\n", team) ~each student: team val [name,email]: student_mapping[student] printf("\t~a: ~a, ~a\n", student, name, email)
When writing the Rhombus code, I naturally ended up inlining the helper functions - presumably, because the syntactic overhead was lighter, this time I intuitively felt the code would still be readable even without splitting it out into a separate function.
Finally, we can print out some summary statistics for the overall class:
(printf "Total students: ~a = ~a seen + ~a unseen\n" (set-count all-students) (set-count (set-intersect all-students all-seen-students)) (set-count (set-subtract all-students all-seen-students))) (printf "~a completed\n" (set-count (set-intersect assignment1-students assignment2-students)))
show_stats "Total roster ~a = ~a seen + ~a unseen": all_students seen_students (all_students *~~ seen_students) show_stats "~a completed": (asg1_students *&& asg2_students)
Once again, the lower burden to writing macros in Rhombus left me with more freedom to experiment, and so, when writing the Rhombus code, I ended up using an additional macro to simplify the printing code:
expr.macro | 'show_stats $text: $arg ...': values('printf($text ++ "\n", set.count($arg), ...)', '') | 'show_stats $text: $arg ...': values('printf($text ++ "\n", set.count($arg), ...)', '')
Conclusion: Sexp-o, ergo sum? Racket vs Rhombus
So. What did we learn? Well… I guess maybe, sometimes, once in a while, once in a blue moon, now and then, we might not need to use s-exps and parentheses in our Lisps?
More seriously, playing around with Rhombus was a great deal of fun. The language is really in a prototype stage at the moment, and the error messages aren't always super clear, but even at this point, you really do get the same wondrous feeling of empowerment to write your own language that is common to all Lisps.
In terms of critisisms, aside from the second-class editing support in GNU+Emacs compared to Dr.Racket, my main pain point while writing Rhombus was an increased difficulty in doing REPL-based development.
In particular, in other Lisps, I often find it easy to quickly prototype ideas by iteratively pasting sub-expressions (delineated by parentheses) into the REPL to analyze and debug what any given piece of code is doing. In contrast, when writing Rhombus code, as there are fewer lexical markers to denote the start and end of sub-expressions, I inevitably ended up shifting to a more mundane workflow, wherein I would just run the whole script in its entirety after each change.
For me personally, as a veritable OCaml connoisseur, this wasn't a huge deal-breaker as it's a workflow that I'm already quite comfortable with. However, looking forwards to the future of Rhombus, given the great emphasis many Lisp developers place on the virtues of their REPL-based workflows5, this likely a problem that may need to be addressed if Rhombus is hoping to make inroads on any existing lisp communities.