-
Notifications
You must be signed in to change notification settings - Fork 198
Architecture: happstack
Happstack is a web framework for Haskell. happstack-server provides utilities for responding to HTTP requests. happstack-state provides persistent storage of native Haskell datatypes.
The information below is subject to change for major Happstack releases. It's a very rough guide; the documentation is of course a good place to start too.
ServerPart
is the main monad from which requests are served. Unwrapped in 0.5.0.2, it looks something like:
Request -> (Maybe (Either Response a, SetAppend (Response -> Response)))
Essentially it's a Request -> Response
which can fail or be modified in a bunch of ways.
Below I've outlined some of the things you can do with it.
The top level server function looks something like:
simpleHTTP :: Conf -> ServerPart Response -> IO ()
If you do simpleHTTP (return res)
, then this serves res :: Response
with an HTTP 200 (ok) response code. The easiest way to construct a response is with the ToMessage
typeclass. In Hackage, these are defined in Happstack.Server.ResourceTypes
.
newtype CabalFile = CabalFile BS.Lazy.ByteString
instance ToMessage CabalFile where
toContentType _ = BS.pack "text/plain"
toMessage (CabalFile bs) = bs
newtype SuggestJson = SuggestJson JSON.JSValue
instance ToMessage SuggestJson where
toContentType _ = BS.pack "application/x-suggestions+json"
toMessage (SuggestJson val) = BS.Lazy.pack $ JSON.encode val
The method toResponse
constructs a Response object from this information. Some useful instances of ToMessage
that are built in with Happstack include (), which returns a blank page, and String, which returns a text/plain page.
Happstack provides a way to tamper with responses even before you generate them by providing a stack of functions to filter the response through. It's called FilterMonad
, and ServerPart
is an instance of it. Some useful functions defined in Happstack.Server.SimpleHTTP
, again with simplified typeclasses, are:
setResponseCode :: Int -> ServerPart ()
ok :: a -> ServerPart a
--- and likewise for forbidden, notFound, badRequest, internalServerError
seeOther :: ToSURI uri => uri -> res -> ServerPart res
-- Some instances of ToSURI: String, Network.URI
addHeaderM :: String -> String -> ServerPart ()
setHeaderM :: String -> String -> ServerPart ()
Examples:
randomFind :: ServerPart Response
randomFind = do
num <- liftIO $ randomRIO (1, 3)
case num of
1 -> ok $ toResponse "Important information"
2 -> notFound $ toResponse "The fates have decided your document is missing"
3 -> do
setHeaderM "X-Serve-Info" "Two cubes of sugar"
setResponseCode 418
return $ toResponse "I'm a teapot"
_ -> internalServerError $ toResponse "Impossible"
Happstack utilizes a combinator approach to route URIs that are requested from it. First, it splits URIs into their component parts, so "/page/5/test.html" becomes ["page", "5", "test.html"]. Then it lets you traverse the list using these key functions:
dir :: String -> ServerPart a -> ServerPart a
path :: FromReqURI a => (a -> ServerPart b) -> ServerPart b
nullDir :: ServerPart ()
ServerPart
is an instance of MonadPlus
, so routes are combined together using mplus
. if a certain URI routing returns mzero
, it tries the next one, and this is a very cheap operation. dir
makes sure a ServerPart
is only entered if it can pop the directory in question, so dir "page"
would make the above list become ["5", "test.html"]. path
enters a ServerPart
if there's still a path segment left, regardless of what it is, and passes its value along the way. Otherwise, both of these return mzero. nullDir
makes a branch return mzero if there are still remaining path segments. For example:
simpleHTTP $ msum
[ dir "help" $ return . toResponse $ "help"
, dir "content" $ msum
[ dir "index" $ do
nullDir
return . toResponse $ "index"
, path $ \str -> case str of
('t':_) -> mzero
_ -> return . toResponse $ "other " ++ str
]
, notFound . toResponse $ "404"
]
Here are some example results that cover all of the branches.
/help help
/help/me/ help
/content 404
/content/index/ index
/content/index/help 404
/content/telephone 404
/content/internet other internet
/omg 404
Routing URIs isn't enough, though: it's also important to select certain ways to response to certain HTTP methods, like GET or POST. These are defined in Happstack as constructors of the Method type, and they can be used with:
methodSP :: Method -> ServerPart a -> ServerPart a
-- Executes code only if the method is the Method and
-- there are no remaining path segments (nullDir). This is very
-- commonly used in Happstack code.
methodOnly :: Method -> ServerPart ()
-- Guards against the method, nothing else.
Note that the Hackage approach ignores these combinators in order to merge nearly-redundant routes with the Resource
and ServerTree
objects, but it's useful to know how it works.
After routing your URIs; how do you obtain other information from requests? Happstack provides some functions to take advantage of the fact that it's a Reader
for the Request
object (defined in Happstack.Server.HTTP.Types
).
getHeaderM :: String -> ServerPart (Maybe ByteString)
-- e.g. agent <- fmap (maybe "Unknown" BS.unpack) $ getHeaderM "User-Agent"
The preferred way to get form inputs is with
getDataFn :: RqData a -> ServerPart (Maybe a)
RqData
is a reader monad over the inputs of a request (there is presently no distinction between inputs in a URI and those in a request body, whether multipart or URI-encoded). An example:
mres <- getDataFn $ liftM2 (,) (lookInput "file") (look "summary")
case mres of
Nothing -> badRequest $ toResponse "file and summary inputs have to both be present"
Just (body, _) | maybe True (any (not . isAlpha)) (inputFilename body) ->
badRequest $ toResponse "File name is invalid"
Just (body, summary) -> do
isMinor <- fmap (maybe True $ \_ -> False) $ getDataFn $ look "minor"
return . toResponse $ printf
"You submitted a %d-byte document with the summary %s. This is%s a minor revision.\n"
(BS.length $ inputValue body) (show summary) (if isMinor then "" else " not")
The askRq
method gives you the request object directly, although there's little need for this (except perhaps to get the raw request body).
You can serve a file by creating a lazy bytestring for it and making a ToMessage
type for it. Alternatively, you can use functions in Happstack.Server.HTTP.FileServe
with simplified typeclasses:
serveFile :: (FilePath -> ServerPart String) -> FilePath -> ServerPart Response
asContentType :: Monad m => String -> FilePath -> m String
guessContentTypeM :: Monad m => MimeMap -> FilePath -> m String
You can write serveFile (asContentType "/text/plain") "/path/to/file"
or use the guessing function with mimeTypes :: MimeMap
. Make sure you're generating the path to the file yourself (use System.FilePath.Posix
), rather than letting the user provide it.
All it takes to store a Hackage data structure to be stored is to make a Serialize instance for it, defined in Happstack.Data.Serialize
.
class (Typeable a, Version a) => Serialize a where
getCopy :: Contained (Get a)
putCopy :: a -> Contained Put
This also requires a Typeable and Version instance. Typeable is pretty simple: use the DeriveDataTypeable
extension of GHC. You'll need all of the component parts of a data structure to be Typeable
to use it.
A basic Version instance is:
instance Version DataStructure where mode = Versioned 0 Nothing
This instance allows the struture to be migrate to a newer one by, among other things, incrementing the version number.
With Happstack.State
, you can make a serialize instance for a data structure with Template Haskell magic:
$(deriveSerialize ''DataStructure)
If you have numbers or strings with special meanings (like id numbers) it's a good practice to use a newtype wrapper and make a basic Serialize
instance for it.
If you find you need to declare Serialize
instances for data structures not defined by hackage-server, put them in Distribution.Server.Instances
.
Component
s are data structures which you can query and update with happstack-state. Declaring an instance for a single structure is pretty simple:
instance Component DataStructure where
type Dependencies DataStructure = End
initialValue = emptyDataStruture
The highest level component in hackage-server is in Distribution.Server.State
, and if you want a feature to use querying functions for a component, you should add it there. The main difference between using many combined components and one big data structure is that although updates are atomic within a component, they can't be all combined into one update function if an action affects more than one of them. The upside is that it's so much more flexible for adding new data.
Functions for a component should all be declared in same module. They use the Update
monad, which is a MonadState
, and the Query
monad, which is a MonadReader
. For instance:
incrementByFloat :: String -> Float -> Update DataStructure Bool
incrementByFloat name arg = do
updateStructure (+arg)
data <- get
res <- ... data ...
case res of
Nothing -> return False --failure
Just data' -> put data' >> return True
updateStructure :: (Float -> Float) -> Update DataStructure ()
updateStructure func = modify $ \st -> st { floatThing = func $ floatThing st }
getFloatThing :: Query DataStructure Float
getFloatThing = asks floatThing
-- have to be grouped together
$(mkMethods ''DataStructure ['incrementByFloat
,'getFloatThing
])
keepItUnder :: IO Float
keepItUnder = do
res <- query $ GetFloatThing
when (res > 10) $ update $ IncrementByFloat "thing" (10-res)
return res
Some tips:
- Do not let update arguments be too large (e.g., replacing an entire data structure). In order to replay data transactions, if necessary, happstack-state records the arguments into its database, and you can get a gigabyte-sized database very easily.
- As a consequence of the above, Update/Query functions that are listed in mkMethods
have to have serializeable arguments. It doesn't prevent making helper
functions like
updateStructure
, only using them in mkMethods. - Make sure update operations maintain the consistency of any internal
indices and minimize server race conditions by integrating checking into
the Update operation itself, possibly with additional invariant-checking
afterwards.
keepItUnder
fails to do this (imagine how calls could interleave). - Try not to do any Template Haskell in feature modules themselves, partly because of TH declaration ordering rules, and also to keep features focused on how to expose the data structures rather than their gritty internals.
- It's a good idea to supply operations to query the entire component (defined
as
ask
) and replace the entire component (defined asput
). This is useful for backup.