I did the development using Geiser, an Emacs package for interactive Scheme development. It can connect with a running Scheme, and you can evaluate code from within Emacs, have completion based on the currently available bindings, etc.
The coolest part of this is being able to reevaluate function and class definitions while the program is running, and seeing the effects immediately. In this sense, GOOPS (the Guile Object Oriented Programming System, inspired by the Common Lisp Object System) is really cool in that you can redefine a class and the existing instances will automatically be updated to reflect the new class definitions (unlike, say, Python, in which if you redefine a class, it's an entirely new class, and instances of the old class will have no relationship with the new class).
One thing I realized is that you have to adapt your program a bit if you want to make the most of interactive development. For example, in Guile's web framework, there is a procedure (run-server handler), which starts up the web server and calls handler to serve each request. My request handler was a procedure named
handle-request, so my code called
(run-server handle-request). The problem is that this way
run-server will be called with the value of
handle-request at the time we started the server, and subsequent redefinitions of
handle-request while the server is running will have no effect on the running server. Instead, I ended up writing something like:
(start-server (lambda (request body) (handle-request request body)))
I.e., instead of calling
handle-request directly, the server will call the anonymous function which, when called, will call
handle-request. In this way, it will use the value of
handle-request at each time the anonymous function is called, so it will see changes to the definition.
Another thing to take into account is that there may be some bindings you don't want to redefine when you reload a module, e.g., variables holding program state (in my case, a variable holding the blog object). For that, Guile provides a
define-once form, which only defines a variable if it doesn't already exist.
One gotcha I encountered was when using parameters, the Scheme world equivalent of dynamically-scoped variables. Parameters in Guile have per-thread independent values, and since the web server and the REPL run in different threads, they may see different values for the same parameter. I ended up not using parameters anyway for other reasons (more on that later).
Guile comes with a web framework of sorts, though it is pretty bare-bones. (Actually the main thing I missed in it was having to parse the request query and POST data by hand. At least it does provide a function to percent-decode URL components.) It has a philosophy of pre-parsing all headers into structured data types as a way of avoiding programming errors. It's an interesting idea; I have mixed feelings about it, but I think it's a valid idea to build a framework on (after all, if you're going the trouble of making a new framework, you might as well try some new ideas rather than creating yet another run-of-the-mill web framework).
You start the server by calling
run-server with a callback (as mentioned above). Whenever a new request comes, the callback will be called with a request object and the request body as a bytevector. The callback must return (at least1) two values: a response object and a response body. Guile allows some shortcuts to be taken: Instead of a response object, you can pass an association list of response headers and the framework will automatically make a response object out of it. The response body may be either a string (which will be automatically encoded to the proper encoding, usually UTF-8), a bytevector, or a procedure; in the latter case, the procedure will be invoked with a port as an argument, and whatever you print to that port will be sent as the response body.
Guile comes with support for SXML, an S-expression based tree representation of XML. This means you can write things like:
(sxml->xml `(div (@ (class "foo")) "Hello, world"))
and it will emit
<div class="foo">Hello, world</div>. The nice thing is that strings appearing in the tree will be automatically escaped approprietely, so you don't have to worry about escaping (or forgetting to escape) data that may contain special characters, such as
That very feature was at first what led me not to want to use SXML, appealing though it was, to render Blognir pages. The reason is that post contents in Blognir come raw from a post file; I didn't want to parse the file HTML contents into SXML just to dump it again as HTML in the output2, and I saw no way to insert a raw string in the middle of an SXML tree bypassing the escaping in the output. So I began this adventure by printing chunks of HTML by hand. At some points I needed to escape strings to insert them in the HTML, so I wrote a small wrapper function to call
sxml->xml on a single string and return the escaped string (by default
sxml->xml prints to a port rather than returning a string).
When I got to the post comments form, where I have to do a lot of escaping (because all field values have to be escaped), I decided to use
sxml->xml for once, for the whole form, rather than escaping the individual strings. I found it so nice to use that I decided to look up the source code for
sxml->xml to see if there wasn't a way to insert raw data in the SXML tree without escaping, so I could use it for the whole page, not just the form. And sure enough, I found out that if you put a procedure in the tree,
sxml->xml will call that procedure and whatever it prints is emitted raw in the result. This feature does not seem to be documented anywhere. (In fact, the SXML overview Info page says (This section needs to be written; volunteers welcome.). Maybe that's up to me!) By that point I had already written most of the rest of the page by printing HTML chunks, and I did not go back and change everything to use SXML, but I would like to do so. I did use SXML afterwards for generating the RSS feeds though, with much rejoicing.
Parameters are used for dynamically scoped values. They are used like this:
;; 23 is the initial value of the parameter. (define current-value (make-parameter 23)) (define (print-value) (display (current-value)) (newline)) (print-value) ;; prints 23 (parameterize ([current-value 42]) (print-value)) ;; prints 42 (print-value) ;; prints 23 again
My original plan was to create a bunch of parameters for holding information about the current request (
current-post-data and the like), so I wouldn't have to pass them as arguments to every helper request handling function; I would just bind the parameters at the main
handle-request function, and all functions called from
handle-request would be able to see the parameterized values.
The problem with my plan is that instead of returning the response body as a string from
handle-request, I was returning a procedure for the web framework to call. By the time the procedure was called,
handle-request had already finished, and the
parameterize form was not in effect anymore. Therefore the procedure saw the parameters with their initial value rather than the value they had when the procedure was created. Oops!
Because closures don't close over their dynamic scope (that's kinda the whole point of dynamic scope), parameters ended up not being very useful for me in this case. I just passed everything as, ahem, parameters (the conventional kind) to the subfunctions.
Despite its crude design, the original Blognir is pretty fast; it takes around 1.7ms to generate the front page in my home machine. I got Parenthetical Blognir at around 3.3ms for now. I'm sure there are still optimizations that can be done, and I may still try some things out, but right now I don't have any pressing need to make things faster than that.
I did learn a few things about optimizing Guile programs in the process, though. I used the
ab utility (package apache2-utils on Debian) to measure response times, and Guile's statistical profiler to see where the bottlenecks were. I did not keep notes on how much impact each change I did had on performance (and in many cases I changed multiple things at the same time, so I don't know the exact impact of each change), but I can summarize some of the things I learned.
This is standard advice and applies to every programming language, but do use the profiler to find out performance bottlenecks. They often come from unexpected places, and it is far better to see the numbers than trying to guess where the bottlenecks might be.
Some things are slow because they take up CPU themselves, and some things are slow because they generate garbage, so your program spends more time in the garbage collector. The procedure currently taking the most time in Parenthetical Blognir is
%after-gc-thunk, and I still have not analysed it more carefully to determine where this is coming from.
Guile has two versions of
format (the Guile counterpart to
printf): one from the
(ice-9 format) module, written in Scheme, which supports various format specifiers; and a bare-bones one, written in C, available under the name
simple-format, which only supports the
~s format specifiers with no qualifiers.3
simple-format is much faster than the
(ice-9 format), so if you only need the bare
~s formats and your program makes heavy use of them, I recommend using
My program was spending quite a bit of time on
read-string to read post contents. It only reads the contents of a post to dump it into the response, and
read-string will waste some time decoding the input from UTF-8 to Guile's internal string type, dynamically growing a string to the right size to fit the post contents, just to immediately decode it back to UTF-8 and discard the string as garbage. I found it more efficient to write a little function which takes two ports (the open post file and the open response port) and just copy the raw bytes in fixed-size chunks from one port to the other:
(use-modules (rnrs bytevectors) (ice-9 binary-ports)) (define (pump-port in-port out-port) (let ([buffer (make-bytevector 65536)]) (let loop () (let ([count (get-bytevector-n! in-port buffer 0 65536)]) (when (not (eof-object? count)) (put-bytevector out-port buffer 0 count) (loop))))))
For extra garbage reduction, you can reuse the buffer across calls (but be careful if the code may be called from multiple threads).
In general, I liked the experience of rewriting the blog in Guile. It was the first time I did interactive Scheme development with Emacs (previously I had only used the REPL directly), and it was pretty cool. Some things could be better, but I see this more as an opportunity to improve things (whether by contributing to existing projects, by writing libraries to make some things easier, or just as things to take into account if/when I decide to have a go again at trying to make my own Lisp), rather than reason for complaining.
There are still a few features missing from the new blog system for feature parity with the current one, but it already handles all the important stuff (posts, comments, list of recent comments, filtering by tag, RSS feeds). I hope to be able to replace the current system with the new one Real Soon Now™.
If you return more than two values, the extra values will be passed back to the callback as arguments on the next call. You can use it to keep server state. If you want to use this feature, you can also specify in the call to
run-server the initial state arguments to be passed in the first call to the callback.
And my posts are not valid XML anyway; I don't close my
<p> tags when writing running text, for instance.
There is also a
format binding in the standard environment. It may point to either
simple-format or the
format, depending on whether
(ice-9 format) has been loaded or not.
Copyright © 2010-2018 Vítor De Araújo
O conteúdo deste blog, a menos que de outra forma especificado, pode ser utilizado segundo os termos da licença Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International.
Powered by Blognir.