Lazy scanl is not actually lazy.

[kindaro]

See the source code:

bytestring/Data/ByteString/Lazy.hs

Line 611 in 416b719

scanl f z = snd . foldl k (z,singleton z)

— Since it is written in terms of foldl, it cannot be any less strict, and foldl forces all the spine. There is a way to write scanl in terms of foldr though! See here for example.

I checked the strictness property and I found that the current definition forces all the spine and that an alternative definition may be given in terms of foldr that is strictly lazier.

This is the code I am using to check strictness:

explosiveTail = (`L.append` error "Tail of this byte string is undefined!")

property = \ xs ->
        L.take (L.length xs + 1) (L.scanl (+) 0 (explosiveTail (xs `L.append` L.singleton 1))) == (L.take (L.length xs + 1) . L.pack . fmap (L.foldr (+) 0) . L.inits) xs

Am I missing something or should this be fixed? If it should be fixed, I can do it on the cheap as part of #364 since I am already writing a bunch of strictness checks for all other lazy scans there.

[kindaro]

I provisionally fixed it in #364.

[Bodigrim]

Ah, interesting catch!