Personal tools

HXT/Conversion of Haskell data from/to XML

From HaskellWiki

Revision as of 14:43, 21 April 2008 by UweSchmidt (Talk | contribs)

Jump to: navigation, search

1 Serializing and deserializing Haskell data to/from XML

With so called pickler functions and arrows, it becomes rather easy and straight forward to convert native Haskell values to XML and vice versa. The module Text.XML.HXT.Arrow.Pickle and submodules contain a set of picklers (conversion functions) for simple data types and pickler combinators for complex types.


2 The idea: XML pickler

For conversion of native Haskell data from and to external representations, there are two functions necessary, one for generating the external representation and one for reading/parsing the representation. The read/show pair often form such a pair of functions.

A so called pickler is a value with two such conversion functions, but because it's necessary to apply a whole sequence of conversion functions at once, there is a state holding the external data, that has to be updated during encoding and decoding. So the simplest form of a pickler converting between a value of type t and a sequence of Chars looks like this.

type St    = [Char]
data PU a  = PU { appPickle   :: (a, St) -> St
		, appUnPickle :: St -> (a, St)

Andrew Kennedy has described in a programming pearl paper [1], how to define primitive picklers and a set of pickler combinators to de-/serialize from/to (Byte-)Strings.

The HXT picklers are an adaptation of these pickler combinators. The difference to Kennedys approach is, that the external representation is not a list of Chars but a list of XmlTrees. The basic picklers for the primitve types (Int, Bool,...) will convert simple values into XML text nodes. New are the picklers for creating XML element and attribute nodes.

The HXT pickler type is defined as follows

data St		= St { attributes :: [XmlTree]
		     , contents   :: [XmlTree]
data PU a	= PU { appPickle   :: (a, St) -> St
		     , appUnPickle :: St -> (Maybe a, St)
		     , theSchema   :: Schema

In XML there are two places for storing informations, the attributes and the element contents. Furthermore the pickler contains a third component for type information. This enables the derivation of a DTD from a set of picklers. In the following examples we do not need this component.

We will see, that with the predefined picklers and pickler combinators we don't have to look very much into these internals. Let's start with an example.

3 Example: Processing baseball league data

3.1 The XML data structure

From the set of HXT/Practical example we'll take the data structure from HXT/Practical/Simple2 dealing with baseball league data. First let's have an idea about the structure of the XML data. The structure is not defined by a DTD or schema, so wee have to guess some issues. Here is a part of the example XML file:

<SEASON YEAR="1998">
  <LEAGUE NAME="National League">
    <DIVISION NAME="East">
      <TEAM CITY="Atlanta" NAME="Braves">
        <PLAYER GIVEN_NAME="Marty" SURNAME="Malloy"
            POSITION="Second Base" GAMES="11"
            GAMES_STARTED="8" AT_BATS="28" RUNS="3"
            HITS="5" DOUBLES="1" TRIPLES="0"
            HOME_RUNS="1" RBI="1" STEALS="0"
        <PLAYER GIVEN_NAME="Ozzie" SURNAME="Guillen"
            POSITION="Shortstop" GAMES="83"
            GAMES_STARTED="59" AT_BATS="264" RUNS="35"
            HITS="73" DOUBLES="15" TRIPLES="1"
            HOME_RUNS="1" RBI="22" STEALS="1"
        <PLAYER GIVEN_NAME="Danny" ... HIT_BY_PITCH="0">
        <PLAYER GIVEN_NAME="Gerald" ...>
      <TEAM CITY="Florida" NAME="Marlins">
      <TEAM CITY="Montreal" NAME="Expos">
      <TEAM CITY="New York" NAME="Mets">
      <TEAM CITY="Philadelphia" NAME="Phillies">
  <LEAGUE NAME="American League">
    <DIVISION NAME="East">
    <DIVISION NAME="Central">

3.2 The Haskell data model

Let's first analyze the underlying data model and then define an appropriate set of Haskell data type for the internal representation.

  • The root type is a Season, consisting of a year an a set of Leagues
  • The Leagues are all identified by a String and consist of a set of Divisions, so it's a Map.
  • The Divisions are also identified by a String and consist of a list of Teams, so it's again a Map
  • A Team has three components, a teamName, a city, and a list of Players
  • A Player has a lot of attributes, for simplicity of the example in the internal modell we will not take all fields into account. Just six fields are included, the firstName, the lastName, the position, atBats, hits and era. All others will be ignored.

So the Haskell data model looks like this:

import Data.Map (Map, fromList, toList)
data Season = Season
    { sYear    :: Int
    , sLeagues :: Leagues
	      deriving (Show, Eq)
type Leagues   = Map String Divisions
type Divisions = Map String [Team]
data Team = Team
    { teamName :: String
    , city     :: String
    , players  :: [Player]
	    deriving (Show, Eq)
data Player = Player
    { firstName :: String
    , lastName  :: String
    , position  :: String
    , atBats    :: Maybe Int
    , hits      :: Maybe Int
    , era       :: Maybe Float
	      deriving (Show, Eq)

3.3 The predefined picklers

In HXT there is a class XmlPickler defining a single function xpickle for overloading the xpickle function name.

class XmlPickler a where
    xpickle :: PU a

For the simple data types there is an instance for XmlPickler, which uses the primitive pickler xpPrim for conversion from and to XML text nodes. This primitive pickler is available for all types supporting read and show.

instance XmlPickler Int where
    xpickle = xpPrim
instance XmlPickler Integer where
    xpickle = xpPrim

For composite data there are predefined pickler combinators for tuples, lists and Maybe types.

instance (XmlPickler a, XmlPickler b) => XmlPickler (a,b) where
    xpickle = xpPair xpickle xpickle
-- similar instances for (,,), (,,,), ...
instance XmlPickler a => XmlPickler [a] where
    xpickle = xpList xpickle
instance XmlPickler a => XmlPickler (Maybe a) where
    xpickle = xpOption xpickle
  • xpPair take two picklers and builds up a pickler for a tuple type. There are also pickler combinators for triples, 4- and 5- tuples.
  • xpList takes a pickler for an element type and gives a list pickler
  • xpOption takes a pickler and returns a pickler for optional values.

Furthermore we need pickler for generating/reading element and attribute nodes

  • xpElem generates/parses an XML element node
  • xpAttr generates/parses an attribute node

Most of the other structured data is pickled/unpickled by converting the data to/from tuples, lists and options. This is done by a wrapper pickler xpWrap.

3.4 Constructing the example picklers

For every Haskell type we will define a pickler.

For the own data types we will declare instances of the XmlPickler class.

instance XmlPickler Season where
    xpickle = xpSeason
instance XmlPickler Team where
    xpickle = xpTeam
instance XmlPickler Player where
    xpickle = xpPlayer

Then the picklers are developed top down starting with xpSeason.

xpSeason	:: PU Season
    = xpElem "SEASON" $
      xpWrap ( uncurry Season
	     , \ s -> (sYear s, sLeagues s)) $
      xpPair (xpAttr "YEAR" xpickle) xpLeagues

A Season value is mapped onto an element SEASON with xpElem. This constructs/reads the XML SEASON element. The two components of Season are wrapped into a pair with xpWrap. xpWrap needs a pair of functions for a 1-1 mapping between Season and (Int, Leagues). The first component of the pair, the year is mapped onto an attribute YEAR. The attribute value is handled with the predefined pickler for Int. The second one, the Leagues are handled by xpLeagues.

xpLeagues	:: PU Leagues
    = xpWrap ( fromList
	     , toList ) $
      xpList $
      xpElem "LEAGUE" $
      xpPair (xpAttr "NAME" xpText) xpDivisions

xpLeagues has to deal with a Map value. This can't done directly, but the Map value is converted to/from a list of pairs with xpWrap and (fromList, toList). Then the xpList is applied for the list of pairs. Each pair will be represented by an LEAGUE element, the name is mapped to an attribute NAME, the divisions are handled by xpDivisions.

xpDivisions	:: PU Divisions
    = xpWrap ( fromList
	     , toList
	     ) $
      xpList $
      xpElem "DIVISION" $
      xpPair (xpAttr "NAME" xpText) xpickle

The divisions are pickled by the same pattern as the leagues.

xpTeam	:: PU Team
    = xpElem "TEAM" $
      xpWrap ( uncurry3 Team
	     , \ t -> (teamName t, city t, players t)) $
      xpTriple (xpAttr "NAME" xpText) (xpAttr "CITY" xpText) (xpList xpickle)

With the teams we have to wrap the three components into a 3-tuple with xpWrap and then pickle a triple of two attributes and a list of players.

xpPlayer	:: PU Player
    = xpElem "PLAYER" $
      xpWrap ( \ ((f,l,p),(a,h,e)) -> Player f l p a h e
	     , \ t -> ((firstName t, lastName t, position t),(atBats t, hits t, era t))) $
      xpPair (xpTriple (xpAttr "GIVEN_NAME" xpText)
	               (xpAttr "SURNAME"    xpText)
	               (xpAttr "POSITION"   xpText))
             (xpTriple (xpOption (xpAttr "AT_BATS" xpickle))
	               (xpOption (xpAttr "HITS"    xpickle))
	               (xpOption (xpAttr "ERA"     xpPrim )))

The Player pickler looks a bit clumsy, because of the six fields. A Player is mapped to an element PLAYER. But because of the many components, we wrap a Player value in a pair of triples to use the predefined picklers xpPair and xpTriple.

When needing picklers for more than five components in various places, it is straight forward to derive e.g. an 'xp10Tuple`` from the HXT sources of xpTriple and others.

New in this case is the use of xpOption for mapping Maybe values onto optional attributes.

The other attributes used in the input, are ignored during unpickling the XML, but this is the only place where the pickler is tolerant with wrong XML.

3.5 A simple application

import Text.XML.HXT.Arrow
-- ...
main	:: IO ()
    = do
      runX ( xunpickleDocument xpSeason [ (a_validate,v_0)
					, (a_trace, v_1)
					, (a_remove_whitespace,v_1)
					, (a_preserve_comment, v_0)
					] "simple2.xml"
	     xpickleDocument xpSeason [ (a_indent, v_1)
				      ] "new-simple2.xml"
      return ()
-- the dummy for processing the unpickled data
processSeason	:: IOSArrow Season Season
    = arrIO ( \ x -> do {print x ; return x})

This application reads in the complete data used in HXT/Practical/Simple2 from file simple2.xml and unpickles it into a Season value. This value is processed (dummy: print out) by processSeason and pickled again into new-simple2.xml

The unpickled value, when formated a bit, looks like this

      { sYear = 1998
      , sLeagues = fromList
	[ ( "American League"
	  , fromList
	    [ ( "Central"
	      , [ Team { teamName = "White Sox"
		       , city = "Chicago"
		       , players = []}
		, ...
	    , ( "East"
	      , [ Team { teamName = "Orioles"
		       , city = "Baltimore"
		       , players = []}
		, ...
	    , ( "West"
	      , [ Team { teamName = "Angels"
		       , city = "Anaheim"
		       , players = []}
		, ...
	, ( "National League"
	  , fromList
	    [ ( "Central"
	      , [ Team { teamName = "Cubs"
		       , city = "Chicago"
		       , players = []}
		, ...
	    , ( "East"
	      , [ Team { teamName = "Braves"
		       , city = "Atlanta"
		       , players =
			 [ Player { firstName = "Marty"
				  , lastName = "Malloy"
				  , position = "Second Base"
				  , atBats = Just 28
				  , hits = Just 5
				  , era = Nothing}
			 , Player { firstName = "Ozzie"
				  , lastName = "Guillen"
				  , position = "Shortstop"
				  , atBats = Just 264
				  , hits = Just 73
				  , era = Nothing}
			 , ...
		, ...
	    , ( "West"
	      , [ Team { teamName = "Diamondbacks"
		       , city = "Arizona"
		       , players = []}
		, ...

4 2. Example: A toy programming language

In this second example we will develop the picklers the other way round. We start with a given data model and derive an XML document structure.

The complete source of this example is included in the HXT distribution.

4.1 The abstract syntax for the programming language

type Program	= Stmt
type StmtList	= [Stmt]
data Stmt
    = Assign  Ident  Expr
    | Stmts   StmtList 
    | If      Expr  Stmt (Maybe Stmt)
    | While   Expr  Stmt
      deriving (Eq, Show)
type Ident	= String
data Expr
    = IntConst	Int
    | BoolConst Bool
    | Var       Ident
    | UnExpr	UnOp  Expr
    | BinExpr	Op    Expr  Expr
      deriving (Eq, Show)
data Op
    = Add | Sub | Mul | Div | Mod | Eq | Neq
      deriving (Eq, Ord, Enum, Show)
data UnOp
    = UPlus | UMinus | Neg
      deriving (Eq, Ord, Read, Show)

A program is a statement, four variants of statement are defined, assignments, sequences, branches and loops. The expressions have five variants, constants, identifiers, unary and binary expressions. The operators are realized as enumeration types.

For developing the picklers, there are two new aspects. This example contains sum data types and it's a recursive structure.

4.2 The pickler definitions

xpProgram :: PU Program
xpProgram = xpElem "program" $
	    xpAddFixedAttr "xmlns" "program42" $
instance XmlPickler UnOp where
    xpickle = xpPrim
instance XmlPickler Op where
    xpickle = xpWrap (toEnum, fromEnum) xpPrim
instance XmlPickler Expr where
    xpickle = xpAlt tag ps
	tag (IntConst _    ) = 0
	tag (BoolConst _   ) = 1
	tag (Var _         ) = 2
	tag (UnExpr _ _    ) = 3
	tag (BinExpr _ _ _ ) = 4
	ps = [ xpWrap ( IntConst
		      , \ (IntConst i ) -> i ) ( xpElem "int"  $
					         xpAttr "value" $
					         xpickle )
	     , xpWrap ( BoolConst
		      , \ (BoolConst b) -> b)  ( xpElem "bool" $
						 xpAttr "value" $
						 xpWrap (toEnum, fromEnum) xpickle )
	     , xpWrap ( Var
		      , \ (Var n)       -> n)  ( xpElem "var"  $
						 xpAttr "name"  $
						 xpText )
	     , xpWrap ( uncurry UnExpr
		      , \ (UnExpr op e) -> (op, e))
                                               ( xpElem "unex" $
						 xpPair (xpAttr "op" xpickle) xpickle )
	     , xpWrap ( uncurry3 $ BinExpr
		      , \ (BinExpr op e1 e2) -> (op, e1, e2))
                                               ( xpElem "binex" $
						 xpTriple (xpAttr "op" xpickle) xpickle xpickle )
instance XmlPickler Stmt where
    xpickle = xpAlt tag ps
	tag ( Assign _ _ ) = 0
	tag ( Stmts _ )    = 1
	tag ( If _ _ _ )   = 2
	tag ( While _ _ )  = 3
	ps = [ xpWrap ( uncurry Assign
		      , \ (Assign n v) -> (n, v))
                                               ( xpElem "assign" $
						 xpPair (xpAttr "name" xpText) xpickle )
	     , xpWrap ( Stmts
		      , \ (Stmts sl) -> sl)    ( xpElem "block" $
						 xpList xpickle )
	     , xpWrap ( uncurry3 If
		      , \ (If c t e) -> (c, t, e))
                                               ( xpElem "if" $
						 xpTriple xpickle xpickle xpickle )
	     , xpWrap ( uncurry While
		      , \ (While c b) -> (c, b))
                                               ( xpElem "while" $
						 xpPair xpickle xpickle )

The root pickler is xpProgram which wraps the main statement in a program element. The program element is decorated with a fixed attribute, defining a name space declaration, just for demonstrating the use of the xpAddFixedAttr.

For the operators two variants are shown. The UnOp is converted with read/show (xpPrim), The Op is in XML represented by a number (xpWrap (toEnum, fromEnum)).

The Expr and Stmt picklers are a bit more interesting. We have to select a pickler for every constructor of the data type. This is done by mapping each variant to a number and then index a list of picklers with this number. For all variants the values are converted with xpWrap into simple values or tuples, and then these values are mapped to XML elements. The simple fields are encoded in attributes, the complex (and recursive) are encoded as child elements.

The complete pickler definitions consist of about 60 lines of code.

4.3 A simple program as Haskell value

p2 :: Program
p2 = Stmts		
     [ Assign x (IntConst 6)
     , Assign y (IntConst 7)
     , Assign p (IntConst 0)
     , While
       ( BinExpr Neq (Var x) (IntConst 0) )
       ( If ( BinExpr Neq ( BinExpr Mod (Var x) (IntConst 2) ) (IntConst 0) )
	    ( Stmts
	      [ Assign x ( BinExpr Sub (Var x) (IntConst 1) )
	      , Assign p ( BinExpr Add (Var p) (Var y) )
	    ( Just ( Stmts
		     [ Assign x ( BinExpr Div (Var x) (IntConst 2) )
		     , Assign y ( BinExpr Mul (Var y) (IntConst 2) )
    x = "x"
    y = "y"
    p = "p"

An example program with rather all variants of statements and expressions.

4.4 The serialized program as XML

<program xmlns="program42">
    <assign name="x">
      <int value="6"/>
    <assign name="y">
      <int value="7"/>
    <assign name="p">
      <int value="0"/>
      <binex op="6">
        <var name="x"/>
        <int value="0"/>
        <binex op="6">
          <binex op="4">
            <var name="x"/>
            <int value="2"/>
          <int value="0"/>
          <assign name="x">
            <binex op="1">
              <var name="x"/>
              <int value="1"/>
          <assign name="p">
            <binex op="0">
              <var name="p"/>
              <var name="y"/>
          <assign name="x">
            <binex op="3">
              <var name="x"/>
              <int value="2"/>
          <assign name="y">
            <binex op="2">
              <var name="y"/>
              <int value="2"/>

This document is generated by executing the following piece of code

storeProgram :: IO ()
  = do
    runX ( constA p2
	   xpickleDocument xpProgram
             [ (a_indent, v_1)
             ] "pickle.xml"
    return ()

It's loaded from a file with

loadProgram :: IO Program
  = do
    [p2] <- runX ( xunpickleDocument xpProgram
                     [ (a_remove_whitespace, v_1)
                     , (a_validate, v_0)
                     ] "pickle.xml"
    return p2

The (a_remove_whitespace, v_1) option is necessary because the XML document is indented when written.

5 A few words of advice

These picklers are a powerful tool for de-/serializing from/to XML. Only a few lines of code are needed for serializing as well as for deserializing. But they are absolutely intolerant when dealing with none valid XML. They are intended to read machine generated XML, ideally generated by the same pickler. When unpickling hand written or by foreign tools generated XML, please validate the XML before reading, preferably with RelaxNG or XML Schema, because of the more powerful type system than those with DTDs.

When designing picklers, one must be careful to put enough markup into the XML structure, to read the XML back without the need for a lookahead and without any ambiguities. The simplest case of a not working pickler is a pair of primitve picklers e.g. for some text. In this case the text is written out and concatenated into a single string, when parsing the XML, there will only be a single text and the pickler will fail because of a missing value for the second component. So at least every primitive pickler must be combined with an xpElem or xpAttr.

It's possible to define various picklers per data type, and picklers can be used one way, just for serializing into XML/HTML. So this approach can also be used to easily generate parts of a HTML document. Examples can be found in the Holumbus search engine project [[2]] and the Haskell api search engine Hayoo! [[3]]. There the HTML code for the search results is generated with picklers.

Please do not try to convert a whole large database into a single XML file with this approach. This will run into memory problems when reading the data, because of the DOM approach used in HXT. In the HXT distribution, there is a test case in the examples dir performance, where the pickling and unpickling is done with XML documents containing 2 million elements. This is the limit for a 1G Intel box (tested with ghc 6.8).

There are two strategies to overcome these limitations. The first is a SAX like approach, reading in simple tags and text elements and not building a tree structure, but writing the data instantly into a database. For this approach the Tagsoup package can be useful. The disadvantage is the programming effort for collecting and converting the data.

The second and recommended way is, to split the whole bunch of data into smaller pieces, unpickle these and link the resulting documents together by the use of 'hrefs.

6 Reading/writing between XML and Haskell data types without XML picklers

This is an example for reading and writing XML without the use of picklers. It was developed before the picklers were added to HXT. The code shows that it's much more effort to implement a conversion than with the technic described above.

6.1 Serializing to Xml

We can create an HXT tree from a single-layer data class as follows:

import IO
import Char
import Text.XML.HXT.Arrow
import Data.Generics
-- our data class we'll convert into xml
data Config = 
   Config { username :: String,
            logNumDays :: Int,
            oleDbString :: String }
   deriving (Show, Typeable,Data)
-- helper function adapted from
-- (gshow replaced by gshow')
introspectData :: Data a => a -> [(String, String)]
introspectData a = zip fields (gmapQ gshow' a)
    where fields = constrFields $ toConstr a
gshow' :: Data a => a -> String
gshow' t = fromMaybe (showConstr(toConstr t)) (cast t)
-- function to create xml string from single-layer Haskell data type
xmlSerialize object = "<" ++ show(toConstr object) ++ ">" ++ 
   foldr (\(a,b) x  -> x ++ "<" ++ a ++ ">" ++ b ++ "</" ++ a ++ ">") "" ( introspectData object )
   ++ "</" ++ show(toConstr object) ++ ">"
-- function to create HXT tree arrow from single-layer Haskell data type:
createHxtArrow object = runLA( constA ( xmlSerialize object ) >>> xread)
-- create a config object to serialize:
createConfig = Config { username = "test", logNumDays = 3, oleDbString = "qsdf" }
-- test function, using our Config data type
testConversion = createHxtArrow( createConfig ) ()

-- hughperkins

6.2 Deserializing from Xml

Here's a solution to deserialize a simple haskell data type containing Strings and Ints.

It's not really pretty, but it works.

Basically, we just convert the incoming xml into gread-compatible format, then use gread :-D

Currently it works for a simple single-layer Haskell data type containing Ints and Strings. You can add new child data types by adding to the case statement in xmlToGShowFormat.

If someone has a more elegant solution, please let me know ( )

module ParseXml
import IO
import Char
import List
import Maybe
import Data.Generics hiding (Unit)
import Text.XML.HXT.Arrow hiding (when)
data Config = Config{ name :: String, age :: Int } 
--data Config = Config{ age :: Int } 
   deriving( Data, Show, Typeable, Ord, Eq, Read )
createConfig = Config "qsdfqsdf" 3
--createConfig = Config 3
gshow' :: Data a => a -> String
gshow' t = fromMaybe (showConstr(toConstr t)) (cast t)
-- helper function from
introspectData :: Data a => a -> [(String, String)]
introspectData a = zip fields (gmapQ gshow' a)
    where fields = constrFields $ toConstr a
-- function to create xml string from single-layer Haskell data type
xmlSerialize object = "<" ++ show(toConstr object) ++ ">" ++ 
   foldr (\(a,b) x  -> x ++ "<" ++ a ++ ">" ++ b ++ "</" ++ a ++ ">") "" ( introspectData object )
   ++ "</" ++ show(toConstr object) ++ ">"
-- parse xml to HXT tree, and obtain the value of node "fieldname"
-- returns a string
getValue xml fieldname | length(resultlist) > 0 = Just (head resultlist)
                                | otherwise = Nothing
    where resultlist = (runLA ( constA xml >>> xread >>> deep ( hasName fieldname ) >>> getChildren >>> getText ))[]
-- parse templateobject to get list of field names
-- apply these to xml to get list of values
-- return (fieldnames list, value list)
xmlToGShowFormat :: Data a => String -> a -> String
xmlToGShowFormat xml templateobject = 
   where mainconstructorname = (showConstr $ toConstr templateobject)
         fields = constrFields $ toConstr templateobject
         values = map ( \fieldname -> getValue xml fieldname ) fields
         datatypes = gmapQ (dataTypeOf) templateobject
         constrs = gmapQ (toConstr) templateobject
         datatypereps = gmapQ (dataTypeRep . dataTypeOf) templateobject
         fieldtogshowformat (value,datatyperep) = case datatyperep of
            IntRep -> "(" ++ fromJust value ++ ")"
            _ -> show(fromJust value)
         formattedfieldlist = map (fieldtogshowformat) (zip values datatypereps)
         go = "(" ++ mainconstructorname ++ " " ++ (concat $ intersperse " " formattedfieldlist ) ++ ")"
xmlDeserialize xml templateobject = fst $ head $ gread( xmlToGShowFormat xml templateobject)
dotest = xmlDeserialize (xmlSerialize createConfig) createConfig :: Config
dotest' = xmlDeserialize ("<Config><age>12</age><name>test name!</name></Config>") createConfig :: Config