Linear types for linear times!

Builder for strict Text and ByteString, based on linear types. It consistently outperforms lazy Builder from text as well as a strict builder from text-builder, and scales better.

> :set -XOverloadedStrings
> import Data.Text.Builder.Linear
> fromText "foo" <> fromChar '_' <> fromDec (42 :: Int)
"foo_42"

String builders in Haskell serve the same purpose as StringBuilder in Java to prevent quadratic slow down in concatenation.

Classic builders such as Data.Text.Lazy.Builder are lazy and fundamentally are dlist with bells and whistles: instead of actually concatenating substrings we compose actions, which implement concatenation, building a tree of thunks. The tree can be forced partially, left-to-right, producing chunks of strict Text, combined into a lazy one. Neither input, nor output need to be materialized in full, which potentially allows for fusion. Such builders allow linear time complexity, but constant factors are relatively high, because thunks are expensive. To a certain degree this is mitigated by inlining, which massively reduces number of nodes.

Strict builders such as text-builder offer another design: they first inspect their input in full to determine output length, then allocate a buffer of required size and fill it in one go. If everything inlines nicely, the length may be known in compile time, which gives blazingly fast runtime. In more complex cases it still builds a tree of thunks and forces all inputs to be materialized.

This package offers two interfaces. One is a mutable Buffer with linear API, which operates very similar to StringBuilder in Java. It allocates a buffer with extra space at the ends to append new strings. If there is not enough free space to insert new data, it allocates a twice larger buffer and copies itself there. The dispatch happens in runtime, so we do not need to inspect and materialize all inputs beforehand; and inlining is mostly irrelevant. Exponential growth provides for amortized linear time. Such structure can be implemented without linear types, but that would greatly affect user experience by polluting everything with ST monad. Users are encouraged to use Buffer API, and built-in benchmarks refer to it.

The second interface is more traditional newtype Builder = Builder (Buffer ⊸ Buffer) with Monoid instance. This type provides easy migration from other builders, but may suffer from insufficient inlining, allocating a tree of thunks. It is still significantly faster than Data.Text.Lazy.Builder, as witnessed by benchmarks for blaze-builder below.

Let's benchmark builders, which concatenate all Char from minBound to maxBound, producing a large Text:

#!/usr/bin/env cabal
{- cabal:
build-depends: base, tasty-bench, text, text-builder, text-builder-linear
ghc-options: -O2
-}

import qualified Data.Text as T
import qualified Data.Text.Lazy as TL
import qualified Data.Text.Lazy.Builder as TLB
import qualified Text.Builder as TB
import qualified Data.Text.Builder.Linear as TBL
import System.Environment (getArgs)
import Test.Tasty.Bench

mkBench :: Monoid a => String -> (Char -> a) -> (a -> Int) -> Benchmark
mkBench s f g = bench s $ nf (g . foldMap f . enumFromTo minBound) maxBound
{-# INLINE mkBench #-}

main :: IO ()
main = defaultMain
  [ mkBench "text, lazy" TLB.singleton (fromIntegral . TL.length . TLB.toLazyText)
  , mkBench "text, strict" TLB.singleton (T.length . TL.toStrict . TLB.toLazyText)
  , mkBench "text-builder" TB.char (T.length . TB.run)
  , mkBench "text-builder-linear" TBL.fromChar (T.length . TBL.runBuilder)
  ]

Running this program with cabal run Main.hs -- +RTS -T yields following results:

text, lazy:
  4.25 ms ± 107 μs,  11 MB allocated, 912 B  copied
text, strict:
  7.18 ms ± 235 μs,  24 MB allocated,  10 MB copied
text-builder:
  80.1 ms ± 3.0 ms, 218 MB allocated, 107 MB copied
text-builder-linear:
  5.37 ms ± 146 μs,  44 MB allocated,  78 KB copied

The first result seems the best both in time and memory and corresponds to the usual Text builder, where we do not materialize the entire result at all. It builds chunks of lazy Text lazily and consumes them at once by TL.length. Thus there are 11 MB of allocations in nursery, none of which survive generation 0 garbage collector, so nothing is copied.

The second result is again the usual Text builder, but emulates a strict consumer: we materialize a strict Text before computing length. Allocation are doubled, and half of them (corresponding to the strict Text) survive to the heap. Time is also almost twice longer, but still quite good.

The third result is for text-builder and demonstrates how bad things could go with strict builders, aiming to precompute the precise length of the buffer: allocating a thunk per char is tremendously slow and expensive.

The last result corresponds to the current package. We generate a strict Text by growing and reallocating the buffer, thus allocations are quite high. Nevertheless, it is already faster than the usual Text builder with strict consumer and does not strain the garbage collector.

Things get very different if we remove {-# INLINE mkBench #-}:

text, lazy:
  36.9 ms ± 599 μs, 275 MB allocated,  30 KB copied
text, strict:
  44.7 ms ± 1.3 ms, 287 MB allocated,  25 MB copied
text-builder:
  77.6 ms ± 2.2 ms, 218 MB allocated, 107 MB copied
text-builder-linear:
  5.35 ms ± 212 μs,  44 MB allocated,  79 KB copied

Builders from text package degrade rapidly, 6-8x slower and 10-20x more allocations. That's because their constant factors rely crucially on everything getting inlined, which makes their performance fragile and unreliable in large-scale applications. On the bright side of things, our builder remains as fast as before and now is a clear champion.

Measured with GHC 9.6 on aarch64:

If you are not convinced by synthetic data, here are benchmarks for blaze-markup after migration to Data.Text.Builder.Linear:

bigTable
  992  μs ±  80 μs, 49% less than baseline
basic
  4.35 μs ± 376 ns, 47% less than baseline
wideTree
  1.26 ms ±  85 μs, 53% less than baseline
wideTreeEscaping
  217  μs ± 7.8 μs, 58% less than baseline
deepTree
  242  μs ±  23 μs, 48% less than baseline
manyAttributes
  811  μs ±  79 μs, 58% less than baseline
customAttribute
  1.68 ms ± 135 μs, 56% less than baseline

Somewhat surprisingly, text-builder-linear now offers rendering to strict ByteString as well. It is consistently faster than bytestring when a string gets over 32k (which is defaultChunkSize for bytestring builder). For mid-sized strings bytestring is slightly faster in certain disciplines, mostly by virtue of using cbits via FFI, while this package remains 100% native Haskell.

Benchmarks below were measured with GHC 9.6 on aarch64 and include comparison to bytestring-strict-builder: