Lately, I’ve been building some web applications using version 2.0.4 of the Play Framework. Overall, I find I like the framework. In my opinion, it compares favorably with technologies like Django and Ruby on Rails, especially for those of us who find that compile-time type safety provides useful benefits.
However, one area where Play lags far behind the competition is in CSS asset compilation.
So, I decided to run some tests…
By default, Play supports the LESS dynamic stylesheet language, which supports variables, mixins, nested rules, and other useful features; LESS stylesheets compile down to standard CSS stylesheets. Rails uses Sass to provide a similar capability. Indeed, there are more similarities than differences between the Sass and LESS languages.
Rails and Play both allow you to edit a CSS source (LESS or Sass) while the framework is running in development mode. When you reload your web site, from a browser, the framework automatically recompiles the LESS or Sass source, producing the CSS files that must be served to the browser.
In Play, LESS compilation is slow. Extremely slow. It’s also buggy, frequently producing JVM core dumps on my Linux machine. More on that, below.
I haven’t measured Rhino’s performance in running the LESS compiler outside of Play, so I don’t know whether Rhino is slower or faster than, say, Node.js. (It would be cool if someone would do that.)
Play’s LESS compilation strategy doesn’t do any form of dependency management. The Managed assets section of http://www.playframework.org/documentation/2.0.4/Assets reads:
By default play compiles all managed assets that are kept in the app/assets older. The compilation process will clean and recompile all managed assets regardless of the change. This is the safest strategy since tracking dependencies can be very tricky with front end technologies.
That documentation also states:
Note if you are dealing with a lot of managed assets this strategy can be very slow.
For this reason there is a way to recompile only the change file and its supposed dependencies. You can turn on this experimental feature by adding the following to your settings:
incrementalAssetsCompilation := true
incrementalAssetsCompilation key (an SBT setting) isn’t actually
available, though, despite being documented. According to
http://stackoverflow.com/questions/12368679, the feature didn’t make it
into Play 2.0.3 (or 2.0.4).
For the intrepid reader, the code that does the LESS-to-CSS translation is in the following source file:
There are several possible causes for the slowness, including:
- Is the LESS compiler slower than the Sass compiler?
- Is running LESS via Rhino slower than running the
lessccommand (via Node.js) on the same files?
- How much is Twitter Bootstrap contributing to the problem?
I tested the first two. As for Twitter Bootstrap, I’m certain that it is
contributing. I’m keeping the Bootstrap source LESS files in my
app/assets/stylesheets tree, largely because it’s easier to modify
_variables.less that way. Twitter Bootstrap is big. Twitter
Bootstrap’s LESS files have 5,125 lines of LESS (including about 770 lines of
white space and comments). My LESS files add another 991 lines (about 240 of
which are white space and comments).
Using a precompiled Bootstrap would probably make a significant difference in LESS recompilation, but I’m not interested in doing that until I’ve exhausted other possibilities, because of the convenience of working with the Bootstrap sources.
I ran the following tests, to try to get an idea of what to do next. Note that a graph follows the test descriptions.
Test 1: Less recompilation via Play
This test is simple: Run a script that touches one of my
.less input files,
then attempt to access a non-authenticated page in the application. Play
detects the modified LESS source file and invokes its internal Rhino-based
recompilation of all LESS files.
Here’s the quick-and-dirty shell script I used:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
The test measures wall clock time, because that’s the measurement that matters to me, as a developer: How long I have to wait for the page to reload, after I’ve modified a LESS source file.
I ran this test with 50 iterations.
Aside: This test was surprisingly difficult to complete, because the Java 7 VM
on my Linux system kept dumping core. Sometimes, I’d only get two requests to
succeed, before the JVM crapped out, presenting me with one of those
hs_err_pidXXXX.log files. Some of the errors are listed below.
- Mean: 19.63 seconds
- Median: 19.48 seconds
- Standard Deviation: 1.45
You read that correctly: The mean and median times to recompile my LESS files are both just shy of 20 seconds. And those figures are fairly consistent, as the standard deviation and the graph illustrate.
No wonder I’ve been avoiding making minor CSS edits.
Here are some of the JVM errors, when Play barfed during the tests:
I didn’t see core dumps like these with the other tests, and I saw plenty of them with the embedded Rhino-based LESS conversion. Clearly, there are problems with Play’s use of Rhino, beyond just being severely slow.
Test 2: Less recompilation via
In this test, I ran the following script 50 times:
1 2 3 4 5 6
The code runs
lessc twice, for each input file, generating both a regular
and a compressed version of the CSS output, since that’s what Play does.
- Mean: 5.34 seconds
- Median: 5.1 seconds
- Standard Deviation: 0.91
That’s more reasonable.
Test 3: Sass, via Play
For this test, I installed the Play-Sass plugin, which is implemented as
an SBT plugin. This plugin implements automatic recompilation of Sass
assets using the
sass command that’s installed with the Ruby
Running this test required a bit more up-front work, including:
- pulling down John W. Long’s sass-twitter-bootrap repository
- removing the Bootstrap LESS files and replacing them with Long’s Bootstrap Sass equivalents
- installing the
sassgem in a separate rvm gemset
- spending about an hour converting my LESS files to Sass, and
- verifying that the application still worked after I was finished with all the painstaking conversions.
Having completed the above, I ran the first script again. Like the original LESS test, the Sass test “warmed up” the server, touched one of my Sass files, and then issued a query for a page, forcing Play to recompile the Sass files.
- Mean: 8.34 seconds
- Median: 8.27 seconds
- Standard Deviation: 0.39
lessc, but still far better than Play’s built-in LESS compiler.
Graph of Results
I should write a LESS compilation SBT plugin that compiles the LESS files via the
lessccommand. (This is how the Play-Sass plugin works.) If I were to do so, and assuming I could easily disable Play’s default LESS compilation, the plugin should be significantly faster, based on these test results.
UPDATE (22 May, 2013) Jonathan Parsons has put together a
play-lesscplugin that uses the
lessccommand to compile LESS files. It can dramatically decrease LESS compilation times in a Play project. See https://github.com/jmparsons/play-lessc.
switching from LESS to Sass also provides a significant increase in CSS compilation speed. It can be a pain to convert all the stylesheets, but once it’s done, it’s done.
Using a precompiled version of Bootstrap might provide the best improvement, since it’s a huge chunk of the LESS code that gets compiled each time. It comes at a big cost in convenience for me, though, so it’s not my first choice.
I should also point out that this issue hasn’t discouraged me from using Play. The framework is still evolving, and this stuff will get better. Plus, courtesy of these tests, I now have some feasible work-arounds.
UPDATE: In a separate thread, James Roper pointed out that Play’s internal LESS compiler can be disabled, by specifying the following build setting:
lessEntryPoints := Nil
Since I want to use some kind of CSS preprocessor, this isn’t an option for me.