[Haskell-cafe] Efficient use of ByteString and type classes in template system

[Donald Bruce Stewart]

johan.tibell:
> Hi Haskell Caf?!
> 
> I'm writing a perl/python like string templating system which I plan
> to release soon:
> 
> darcs get http://darcs.johantibell.com/template
> 
> The goal is to provide simple string templating; no inline code, etc..
> An alternative to printf and ++.

Ok. You might also want to briefly look at the other templating system I
know of in Haskell, this small module by Stefan Wehr,

    http://www.cse.unsw.edu.au/~dons/code/icfp05/tests/unit-tests/VariableExpansion.hs

Just a quick thing he did for the ICFP contest, but does indicate one
way to do it (i.e. via pretty printing).

> 
> Example usage:
> 
> >import qualified Data.ByteString as B
> >import Text.Template
> >
> >helloTemplate = "Hello, $name! Would you like some ${fruit}s?"
> >helloContext = [("name", "Johan"), ("fruit", "banana")]
> >
> >test1 = B.putStrLn $ substitute (B.pack helloTemplate) helloContext
> 
> I want to make it perform well, especially when creating a template
> once and then rendering it multiple times. "Compiling" the template is
> a separate step from rendering in this use case:
> 
> >compiledTemplate = template $ B.pack helloTemplate
> >
> >test2 = B.putStrLn $ render compiledTemplate helloContext
> 
> A template is represented by a list of template fragments, each
> fragment is either a ByteString literal or a variable which is looked
> up in the "context" when rendered.
> 
> >data Frag = Lit ByteString | Var ByteString
> >newtype Template = Template [Frag]
> 
> This leads me to my first question. Would a lazy ByteString be better
> or worse here? The templates are of limited length. I would say the
> length is usually between one paragraph and a whole HTML page. The
> Template data type already acts a bit like a lazy ByteString since it
> consists of several chunks (although the chunck size is not adjusted
> to the CPU cache size like with the lazy ByteString).

Probably lazy bytestrings are better here, since you get O(n/k) append
cost, rather than O(n).  If most strings are small, it mightn't be
noticeable.

> Currently the context in which a template is rendered is represented
> by a type class.
> 
> >class Context c where
> >    lookup :: ByteString -> c -> Maybe ByteString
> >
> >instance Context (Map String String) where
> >    lookup k c = liftM B.pack (Map.lookup (B.unpack k) c)
> >
> >instance Context (Map ByteString ByteString) where
> >    lookup = Map.lookup
> >
> >-- More instance, for [(String, String)], etc.
> 
> I added this as a convenience for the user, mainly to work around the
> problem of not having ByteString literals. A typical usage would have
> the keys in the context being literals and the values some variables:

note sure if it is relevant, but:

    pack "Foo"

will be converted via rewrite rules to a bytestring literal at compile
time. So there's no overhead for having String literals.

> 
> >someContext = Map.fromList [("name", name), ("fruit", fruit)]
> 
> I'm not sure if this was a good decision, With this I'm halfway to the
> (in)famous Stringable class and it seems like many smarter people than

Yes, seems a little worrying.

> me have avoided introducing such a class. How will this affect
> performace? Take for example the rendering function:
> 
> >render :: Context c => Template -> c -> ByteString
> >render (Template frags) ctx = B.concat $ map (renderFrag ctx) frags
> >
> >renderFrag :: Context c => c -> Frag -> ByteString
> >renderFrag ctx (Lit s) = s
> >renderFrag ctx (Var x) = case Text.Template.lookup x ctx of
> >                           Just v  -> v
> >                           Nothing -> error $ "Key not found: " ++ 
> >                           (B.unpack x)
> 
> How will the type dictionary 'c' hurt performance here? Would
> specializing the function directly in render help?

Hmm. Hard to say: look at the Core code and we will know.

Really though, you'll need some stress test cases to be able to make
resonable conclusions about performance.

> 
> >render (Template frags) ctx = B.concat $ map (renderFrag f) frags
> >    where f = flip Text.Template.lookup ctx
> >
> >renderFrag f (Var x) = case f x of
> 
> I can see the implementation taking one of the following routes:
> - Go full Stringable, including for the Template
> - Revert to Context = Map ByteString ByteString which was the original
> implementation.
> - Some middle road, without MPTC, for example:
> >class Context c where
> >    lookup :: ByteString -> c ByteString ByteString -> Maybe ByteString
> This would allow the user to supply some more efficient data type for
> lookup but not change the string type. Having a type class would allow
> me to provide things like the possibility to create a Context from a
> record where each record accessor function would server as key.
> Something like:
> 
> >data Person { personName :: String, personAge :: Int }
> would get converted (using Data?) to:
> >personContext = [("personName", show $ personName aPerson),
> >                 ("personAge", show $ personAge aPerson)]
> but not actually using a Map but the record itself.
> 
> I guess my more general question is: how do I reason about the
> performance of my code or any code like this? Are there any other
> performance improvements that could be made?
> 
> Also, I would be grateful if someone could provide some feedback on
> the implementation, anything goes!
> 
> I still have some known TODOs:
> 
> - Import error messages for invalid uses of "$".
> - Improve the regex usage overall.
> - Add some more functions; the plan is to add those function which
> could be expressed in efficiently with the current interface. An
> example is things like renderAndWrite, when writing doing a B.concat
> first is unnecessary.

I'd suggest: keep it simple and fast. Then work out what extra stuff you
need.

-- Don