Builders are used to efficiently construct sequences of bytes from smaller parts. Typically, such a construction is part of the implementation of an encoding, i.e., a function for converting Haskell values to sequences of bytes. Examples of encodings are the generation of the sequence of bytes representing a HTML document to be sent in a HTTP response by a web application or the serialization of a Haskell value using a fixed binary format.

For an efficient implementation of an encoding, it is important that (a) little time is spent on converting the Haskell values to the resulting sequence of bytes and (b) that the representation of the resulting sequence is such that it can be consumed efficiently. Builders support (a) by providing an O(1) concatentation operation and efficient implementations of basic encodings for Chars, Ints, and other standard Haskell values. They support (b) by providing their result as a lazy ByteString, which is internally just a linked list of pointers to chunks of consecutive raw memory. Lazy ByteStrings can be efficiently consumed by functions that write them to a file or send them over a network socket. Note that each chunk boundary incurs expensive extra work (e.g., a system call) that must be amortized over the work spent on consuming the chunk body. Builders therefore take special care to ensure that the average chunk size is large enough. The precise meaning of large enough is application dependent. The current implementation is tuned for an average chunk size between 4kb and 32kb, which should suit most applications.

As a simple example of an encoding implementation, we show how to efficiently convert the following representation of mixed-data tables to an UTF-8 encoded Comma-Separated-Values (CSV) table.

data Cell = StringC String
          | IntC Int
          deriving( Eq, Ord, Show )

type Row   = [Cell]
type Table = [Row]

We use the following imports and abbreviate mappend to simplify reading.

import qualified Data.ByteString.Lazy               as L
import           Data.ByteString.Builder
import           Data.Monoid
import           Data.Foldable                        (foldMap)
import           Data.List                            (intersperse)

infixr 4 <>
(<>) :: Monoid m => m -> m -> m
(<>) = mappend

CSV is a character-based representation of tables. For maximal modularity, we could first render Tables as Strings and then encode this String using some Unicode character encoding. However, this sacrifices performance due to the intermediate String representation being built and thrown away right afterwards. We get rid of this intermediate String representation by fixing the character encoding to UTF-8 and using Builders to convert Tables directly to UTF-8 encoded CSV tables represented as lazy ByteStrings.

encodeUtf8CSV :: Table -> L.ByteString
encodeUtf8CSV = toLazyByteString . renderTable

renderTable :: Table -> Builder
renderTable rs = mconcat [renderRow r <> charUtf8 '\n' | r <- rs]

renderRow :: Row -> Builder
renderRow []     = mempty
renderRow (c:cs) =
    renderCell c <> mconcat [ charUtf8 ',' <> renderCell c' | c' <- cs ]

renderCell :: Cell -> Builder
renderCell (StringC cs) = renderString cs
renderCell (IntC i)     = intDec i

renderString :: String -> Builder
renderString cs = charUtf8 '"' <> foldMap escape cs <> charUtf8 '"'
  where
    escape '\\' = charUtf8 '\\' <> charUtf8 '\\'
    escape '\"' = charUtf8 '\\' <> charUtf8 '\"'
    escape c    = charUtf8 c

Note that the ASCII encoding is a subset of the UTF-8 encoding, which is why we can use the optimized function intDec to encode an Int as a decimal number with UTF-8 encoded digits. Using intDec is more efficient than stringUtf8 . show, as it avoids constructing an intermediate String. Avoiding this intermediate data structure significantly improves performance because encoding Cells is the core operation for rendering CSV-tables. See Data.ByteString.Builder.Prim for further information on how to improve the performance of renderString.

We demonstrate our UTF-8 CSV encoding function on the following table.

strings :: [String]
strings =  ["hello", "\"1\"", "λ-wörld"]

table :: Table
table = [map StringC strings, map IntC [-3..3]]

The expression encodeUtf8CSV table results in the following lazy ByteString.

Chunk "\"hello\",\"\\\"1\\\"\",\"\206\187-w\195\182rld\"\n-3,-2,-1,0,1,2,3\n" Empty

We can clearly see that we are converting to a binary format. The 'λ' and 'ö' characters, which have a Unicode codepoint above 127, are expanded to their corresponding UTF-8 multi-byte representation.

We use the criterion library (http://hackage.haskell.org/package/criterion) to benchmark the efficiency of our encoding function on the following table.

import Criterion.Main     -- add this import to the ones above

maxiTable :: Table
maxiTable = take 1000 $ cycle table

main :: IO ()
main = defaultMain
  [ bench "encodeUtf8CSV maxiTable (original)" $
      whnf (L.length . encodeUtf8CSV) maxiTable
  ]

On a Core2 Duo 2.20GHz on a 32-bit Linux, the above code takes 1ms to generate the 22'500 bytes long lazy ByteString. Looking again at the definitions above, we see that we took care to avoid intermediate data structures, as otherwise we would sacrifice performance. For example, the following (arguably simpler) definition of renderRow is about 20% slower.

renderRow :: Row -> Builder
renderRow  = mconcat . intersperse (charUtf8 ',') . map renderCell

Similarly, using O(n) concatentations like ++ or the equivalent concat operations on strict and lazy ByteStrings should be avoided. The following definition of renderString is also about 20% slower.

renderString :: String -> Builder
renderString cs = charUtf8 $ "\"" ++ concatMap escape cs ++ "\""
  where
    escape '\\' = "\\"
    escape '\"' = "\\\""
    escape c    = return c

Apart from removing intermediate data-structures, encodings can be optimized further by fine-tuning their execution parameters using the functions in Data.ByteString.Builder.Extra and their "inner loops" using the functions in Data.ByteString.Builder.Prim.

Internally, Builders are buffer-filling functions. They are executed by a driver that provides them with an actual buffer to fill. Once called with a buffer, a Builder fills it and returns a signal to the driver telling it that it is either done, has filled the current buffer, or wants to directly insert a reference to a chunk of memory. In the last two cases, the Builder also returns a continuation Builder that the driver can call to fill the next buffer. Here, we provide the two drivers that satisfy almost all use cases. See Data.ByteString.Builder.Extra, for information about fine-tuning them.

Conversion from Char and String into Builders in various encodings.

The ASCII encoding is a 7-bit encoding. The Char7 encoding implemented here works by truncating the Unicode codepoint to 7-bits, prefixing it with a leading 0, and encoding the resulting 8-bits as a single byte. For the codepoints 0-127 this corresponds the ASCII encoding.

The ISO/IEC 8859-1 encoding is an 8-bit encoding often known as Latin-1. The Char8 encoding implemented here works by truncating the Unicode codepoint to 8-bits and encoding them as a single byte. For the codepoints 0-255 this corresponds to the ISO/IEC 8859-1 encoding.

The UTF-8 encoding can encode all Unicode codepoints. We recommend using it always for encoding Chars and Strings unless an application really requires another encoding.