[PATCH] Improved Hash.from_xml

from_xml using libxml-ruby SAX Parser

This is interesting. Could you please attach the benchmark xml-feed and benchmark-script. It would be interesting to reproduce.

All the best

In my experience, libxml-ruby leaks really badly when creating xml [1]. Perhaps it doesn't when using SAX, but we'd need to be very sure of that. It also doesn't have an active maintainer at the moment and needs some work [2]. I'm not sure we should be including it as a dependency until there's some progress on these issues.

[1] http://rubyforge.org/tracker/index.php?func=detail&aid=7945&group_id=494&atid=1971 [2] http://rubyforge.org/pipermail/libxml-devel/2007-July/000340.html

I've also had experience with libxml-ruby leaking. That was when I was manipulating Nodes though and I actually put it away because of this on my project. I have not had those issues when just reading XML though.

You're right about the development of libxml-ruby. It's disappointingly inactive at the moment. I followed the thread listed above and it sounds promising but at the moment it sucks.

I wasn't expecting this patch to make it into core at the moment with the current status of libxml-ruby, but I did want to get the idea of using a SAX parser for this going. I've used Hpricot before and for valid xml I've found it very slow compared to libxml-ruby. Having said that, Hpricot doesn't leak from what I've seen.

I will follow up witht the benchmark script in a bit.

A Simple Benchmark Script for libxml-ruby vs Hpricot

Have you tried good ol' REXML SAX?

from irc:

(10:04:31 PM) ezmobius: i dont want to depend on lib-xml at all
(10:04:41 PM) ezmobius: hpricot has binany gems for even windows

octopod: your comment is interesting, I look forward to see the benchmark-script and results.

The previous benchmark with a slightly modified version of the libxml sax code that uses REXML fast stream parsing

On my local machine (15" C2D 2.2GHz MBP)

Rehearsal ----------------------------------------------- sax 4.240000 0.070000 4.310000 ( 4.403817) hpricot 10.370000 0.150000 10.520000 ( 10.748726) rexml based 5.650000 0.070000 5.720000 ( 5.895491)

---

user system total real

sax 4.220000 0.060000 4.280000 ( 4.353719) hpricot 10.160000 0.140000 10.300000 ( 10.472697) rexml based 5.580000 0.050000 5.630000 ( 5.685909)

bah! should have previewed first:

Rehearsal -----------------------------------------------
sax           4.240000   0.070000   4.310000 (  4.403817)
hpricot      10.370000   0.150000  10.520000 ( 10.748726)
rexml based   5.650000   0.070000   5.720000 (  5.895491)
------------------------------------- total: 20.550000sec

                  user     system      total        real
sax           4.220000   0.060000   4.280000 (  4.353719)
hpricot      10.160000   0.140000  10.300000 ( 10.472697)
rexml based   5.580000   0.050000   5.630000 (  5.685909)

Patch containg the rexml version, all credit to has.s..@gmail.com for the hard work :)

That's great Chris. I didn't even think to use REXML, what with all those reports that it was very slow I didn't even stop to think about using it.

A benchmark report with confidence intervals. To conclude one significantly different from another (with 95% certainty) look for non-overlapping conf-intervals.

The benchmark-script is almost identical to chris' and has.sox, credit to them!

Note how various feeds give various results:

All time in seconds.

-- http://feeds.feedburner.com/37signals/beMH (n=1000) --
           mean     95% interval
hpricot    0.14   ( 0.07 .. 0.20 )
libxml     0.14   ( 0.09 .. 0.20 )
rexml      0.07   ( 0.01 .. 0.12 )


-- http://brainspl.at/xml/rss20/feed.xml (n=1000) --
           mean     95% interval
hpricot    0.05   ( 0.01 .. 0.09 )
libxml     0.01   (-0.01 .. 0.04 )
rexml      0.02   (-0.01 .. 0.05 )


-- http://jobs.37signals.com/categories/2/jobs;rss (n=1000) --
           mean     95% interval
hpricot    0.28   ( 0.21 .. 0.36 )
libxml     0.11   ( 0.05 .. 0.18 )
rexml      0.10   ( 0.06 .. 0.14 )


-- http://www.aftenposten.no/eksport/rss-1_0/?seksjon=ece_frontpage (n=1000) --
           mean     95% interval
hpricot    0.05   ( 0.01 .. 0.08 )
libxml     0.01   (-0.01 .. 0.03 )
rexml      0.02   (-0.00 .. 0.05 )


-- http://merb.devjavu.com/projects/merb/report/9?format=rss (n=1000) --
           mean     95% interval
hpricot    0.43   ( 0.37 .. 0.49 )
libxml     0.19   ( 0.11 .. 0.27 )
rexml      0.16   ( 0.09 .. 0.23 )


-- http://odeo.com/profile/OdeoMusic/rss (n=1000) --
           mean     95% interval
hpricot    0.32   ( 0.25 .. 0.39 )
libxml     0.07   ( 0.02 .. 0.12 )
rexml      0.12   ( 0.07 .. 0.17 )

This ticket is really rich in information, that's great. I feel chris and has.sox should decide the final outcome, then we can commit and mark as fixed.

There really is a lot of great information in this ticket. Thanx Chris and Jon for getting involved in this.

Looking a Jons' results I would have to vote for the REXML version. It is consistant, and looks to be fast across most feeds tested.

Cheers Daniel

Thanks for the proper benchmarking Jon, what did you use for that?

Regarding which one to use, I think rexml should win as it's faster and part of ruby stdlib. However before choosing, one thing I was thinking about these benchmarks is whether they reflect the general case of xml file that would be turned into a hash. As I understand it, this is mainly going to be used for turning REST style XML into a params hash. If this is the case we should try with a bunch of randomly generated smaller files. If I'm wrong then ignore this :)

OK, I just looked at the code to see how the benchmarking worked. I thought it might be something I've been meaning to check out for a while http://eigenclass.org/hiki.rb?adaptative+benchmark

(In [350]) Commiting the REXML Hash.from_xml code. Nice work to everyone involved, i love's me some benchmarks. Closes #95