xeno

A fast event-based XML parser.

Blog post.

Features

• It’s a SAX-style/fold parser which triggers events for open/close tags, attributes, text, etc.
• It uses low memory (see memory benchmarks below).
• It’s very fast (see speed benchmarks below).
• It cheats like Hexml does (doesn’t expand entities or CDATA, or most of the XML standard).
• It’s written in pure Haskell.

Performance goals

The hexml Haskell library uses an XML parser written in C, so that is the baseline we’re trying to beat or match roughly.

The Xeno.SAX module is faster than Hexml for simply walking the document. Hexml actually does more work, allocating a DOM. Xeno.DOM is slighly slower or faster than Hexml depending on the document, although it is 2x slower on a 211KB document.

Memory benchmarks for Xeno:

Case                Bytes  GCs  Check
4kb/xeno/sax        2,376    0  OK
31kb/xeno/sax       1,824    0  OK
211kb/xeno/sax     56,832    0  OK
4kb/xeno/dom       11,360    0  OK
31kb/xeno/dom      10,352    0  OK
211kb/xeno/dom  1,082,816    0  OK

I memory benchmarked Hexml, but most of its allocation happens in C, which GHC doesn’t track. So the data wasn’t useful to compare.

Speed benchmarks:

benchmarking 4KB/hexml/dom
time                 6.317 μs   (6.279 μs .. 6.354 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 6.333 μs   (6.307 μs .. 6.362 μs)
std dev              97.15 ns   (77.15 ns .. 125.3 ns)
variance introduced by outliers: 13% (moderately inflated)

benchmarking 4KB/xeno/sax
time                 5.152 μs   (5.131 μs .. 5.179 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 5.139 μs   (5.128 μs .. 5.161 μs)
std dev              58.02 ns   (41.25 ns .. 85.41 ns)

benchmarking 4KB/xeno/dom
time                 10.93 μs   (10.83 μs .. 11.14 μs)
                     0.994 R²   (0.983 R² .. 0.999 R²)
mean                 11.35 μs   (11.12 μs .. 11.91 μs)
std dev              1.188 μs   (458.7 ns .. 2.148 μs)
variance introduced by outliers: 87% (severely inflated)

benchmarking 31KB/hexml/dom
time                 9.405 μs   (9.348 μs .. 9.480 μs)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 9.745 μs   (9.599 μs .. 10.06 μs)
std dev              745.3 ns   (598.6 ns .. 902.4 ns)
variance introduced by outliers: 78% (severely inflated)

benchmarking 31KB/xeno/sax
time                 2.736 μs   (2.723 μs .. 2.753 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 2.757 μs   (2.742 μs .. 2.791 μs)
std dev              76.93 ns   (43.62 ns .. 136.1 ns)
variance introduced by outliers: 35% (moderately inflated)

benchmarking 31KB/xeno/dom
time                 5.767 μs   (5.735 μs .. 5.814 μs)
                     0.999 R²   (0.999 R² .. 1.000 R²)
mean                 5.759 μs   (5.728 μs .. 5.810 μs)
std dev              127.3 ns   (79.02 ns .. 177.2 ns)
variance introduced by outliers: 24% (moderately inflated)

benchmarking 211KB/hexml/dom
time                 260.3 μs   (259.8 μs .. 260.8 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 259.9 μs   (259.7 μs .. 260.3 μs)
std dev              959.9 ns   (821.8 ns .. 1.178 μs)

benchmarking 211KB/xeno/sax
time                 249.2 μs   (248.5 μs .. 250.1 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 251.5 μs   (250.6 μs .. 253.0 μs)
std dev              3.944 μs   (3.032 μs .. 5.345 μs)

benchmarking 211KB/xeno/dom
time                 543.1 μs   (539.4 μs .. 547.0 μs)
                     0.999 R²   (0.999 R² .. 1.000 R²)
mean                 550.0 μs   (545.3 μs .. 553.6 μs)
std dev              14.39 μs   (12.45 μs .. 17.12 μs)
variance introduced by outliers: 17% (moderately inflated)

DOM Example

Easy as running the parse function:

> parse "<p key='val' x=\"foo\" k=\"\"><a><hr/>hi</a><b>sup</b>hi</p>"
Right
  (Node
     "p"
     [("key", "val"), ("x", "foo"), ("k", "")]
     [ Element (Node "a" [] [Element (Node "hr" [] []), Text "hi"])
     , Element (Node "b" [] [Text "sup"])
     , Text "hi"
     ])

SAX Example

Quickly dumping XML:

> let input = "Text<tag prop='value'>Hello, World!</tag><x><y prop=\"x\">Content!</y></x>Trailing."
> dump input
"Text"
<tag prop="value">
  "Hello, World!"
</tag>
<x>
  <y prop="x">
    "Content!"
  </y>
</x>
"Trailing."

Folding over XML:

> fold const (\m _ _ -> m + 1) const const const 0 input -- Count elements.
Right 2

> fold (\m _ -> m + 1) (\m _ _ -> m) const const const 0 input -- Count attributes.
Right 3

Most general XML processor:

process
  :: Monad m
  => (ByteString -> m ())               -- ^ Open tag.
  -> (ByteString -> ByteString -> m ()) -- ^ Tag attribute.
  -> (ByteString -> m ())               -- ^ End open tag.
  -> (ByteString -> m ())               -- ^ Text.
  -> (ByteString -> m ())               -- ^ Close tag.
  -> ByteString                         -- ^ Input string.
  -> m ()

You can use any monad you want. IO, State, etc. For example, fold is implemented like this:

fold openF attrF endOpenF textF closeF s str =
  execState
    (process
       (\name -> modify (\s' -> openF s' name))
       (\key value -> modify (\s' -> attrF s' key value))
       (\name -> modify (\s' -> endOpenF s' name))
       (\text -> modify (\s' -> textF s' text))
       (\name -> modify (\s' -> closeF s' name))
       str)
    s

The process is marked as INLINE, which means use-sites of it will inline, and your particular monad’s type will be potentially erased for great performance.