[ANN] Walters - a faster HTML escaping lib for JRuby

Ruby JRuby
1

Walters (https://github.com/wmeissner/walters) is a jruby extension
optimised for fast html/xml/uri/url/href escaping. It originally
started as a Houdini wrapper to test XNI, but then I ported the
Houdini C code to java so JRuby would have a pure java implementation
(which makes it a lot easier to deploy in a WAR). It was named for
Norman Murray Walters, an Australian contemporary of Houdini’s.

Benchmarks.

The performance delta isn’t as impressive as Houdini/EscapeUtils is
for MRI, because JRuby has hotspot to do the heavy lifting, but you
still end up with at least a 3x speedup.

Escaping 1000 bytes of text requiring escaping 1000000 times under
jruby-1.7.4:

Rack::Utils.escape_html 89.980000 0.230000 90.210000 ( 90.272000)
Haml::Helpers.html_escape 50.420000 0.170000 50.590000 ( 51.147000)
ERB::Util.html_escape 44.650000 0.130000 44.780000 ( 45.518000)
CGI.escapeHTML 36.230000 0.090000 36.320000 ( 36.358000)
String#gsub 35.490000 0.090000 35.580000 ( 35.587000)
Walters.escape_html 10.090000 0.030000 10.120000 ( 10.126000)

That equates to not quite 100MB/sec on a 2.26ghz core2duo. If your
web app is trying to push more than that through a 5 year old cpu, you
probably have other problems :slight_smile:

You can also monkey patch any of the above methods (see the README) to
use Walters.escape_html, so you can get automagical speedups just with
a few lines in a rails initializer.

2

Oh nice…I saw a few folks complaining about the Houdini ext and not
getting something for JRuby. Nice going, Wayne :slight_smile:

  • Charlie (mobile)
3

Oh, do you still have the XNI version? Might be a good showcase.

  • Charlie (mobile)
4

The XNI version is still there (its what you get when you install the
gem on MRI), but it is currently fairly slow (but still faster than
the pure-ruby alternatives). When I did the native implementations of
XNI for JRuby and MRI, I went for a fairly simple implementation
versus the highly optimised implementation JRuby’s FFI has. The MRI
backend also releases the GIL on every ruby->C call (to emulate the
JRuby behaviour), which has a bit of overhead.

Things like Houdini (i.e. where “MOAR SP33D!” is the design goal), are
pretty much the worst-case scenario for XNI - by design it has to do
things like copy strings instead of passing a pointer to the string
backing store (since it can’t on JRuby), and release the GIL on MRI.
That shouldn’t worry most C extensions, which should be to access
functionality, not for performance (for that, you really want a
dedicated java/C-ext for the specific VM).