xeno
A fast event-based XML parser in pure Haskell
https://github.com/ocramz/xeno
LTS Haskell 22.33: | 0.6 |
Stackage Nightly 2024-09-07: | 0.6 |
Latest on Hackage: | 0.6 |
xeno-0.6@sha256:804c6130217f8f73720aeefc8bd1f4fe8a7009f8deb81cc38c18ee445d9c8f98,3264
Module documentation for 0.6
xeno
A fast event-based XML parser.
Features
- SAX-style/fold parser which triggers events for open/close tags, attributes, text, etc.
- Low memory use (see memory benchmarks below).
- Very fast (see speed benchmarks below).
- It cheats like Hexml does (doesn’t expand entities, or most of the XML standard).
- Written in pure Haskell.
- CDATA is supported as of version 0.2.
Please see the bottom of this file for guidelines on contributing to this library.
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 const 0 input -- Count attributes.
Right 2
> fold (\m _ -> m + 1) (\m _ _ -> m) const const const const 0 input -- Count elements.
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.
Contributors
See CONTRIBUTORS.md
Contribution guidelines
All contributions and bug fixes are welcome and will be credited appropriately, as long as they are aligned with the goals of this library: speed and memory efficiency. In practical terms, patches and additional features should not introduce significant performance regressions.
Changes
0.6
* GHC 9.4 compatibility :
* get rid of MonadFail (ST s) instance use
* mtl-2.3 compatibility
0.4.3
* Ensure we don't grow with a negative size in DOM parser (#48)
* Flatten code nesting in process function (#45)
* Introduce a whitespace-around-equals CPP flag (#44)
* Use modify' instead of modify in fold (#42)
0.4.2
* all benchmarks marked as such in the Cabal file
0.4
* A number of optimizations and some changes in ergonomics. Thanks to Dmitry Krylov (dmalkr) and Michal Gajda (mgajda) !
* breaking API changes :
* The parameters to function 'Xeno.SAX.process' are now wrapped in a Process type
* Speed optimizations :
* function 'Xeno.DOM.predictGrowSize'
* Xeno.DOM.Robust
* Benchmark improvements :
* Added benchmarks for ByteStringZeroTerminated
* Added benchmarks for big files (bench/SpeedBigFiles.hs)
* Benchmarks run non-threaded
0.3.5.2
* Fix dependency lower bounds (GHC 8.0.1 is the earliest version currently supported)
0.3.5
* Improve error handling (#24 #26, mgajda)
0.3.4
* Fixed #14 and add test for #15
* Fixed typos in the examples (unhammer)
0.3.2
* Fixed DOM parsing from bystrings with non-zero offset (#11, qrilka)
0.3
* Fixed name parsing (for attributes and tags) so it conforms with the XML spec (qrilka)
* Fixed parsing failure when root tag is preceded by white space (though without checking for white space characters specifically) (qrilka)
* Added contribution guidelines (ocramz)
0.2
* Added CDATA support (Rembane)
0.1
* First Hackage release