Conversation

Notices

1. Embed this notice
   :umu: :umu: (a1ba@suya.place)'s status on Wednesday, 19-Aug-2026 11:23:55 JST :umu: :umu:
   informative enough, thanks
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 19-Aug-2026 11:23:54 JST iced depresso

     in reply to
     @a1ba fast zstd is not. it is good.
     rolz is fast and good.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 19-Aug-2026 11:29:40 JST iced depresso

     in reply to

     • iced depresso
     @a1ba was very surprised by tests with orz (a rust crate) against zstd and orz is very competitive (within a few percent in output side.) considering the algorithmic complexity of orz is so much less.
     
     rolz is basically a chain of linked lists, using a very small rolling hash window to pick a slot. then it looks at prior findings at that slot, and emits a code for like "take the current hash slot, walk the back pointer 4 times, copy this many bytes." (the 'restricted offset lz' part.) orz then just feeds that through huffman. LZ and Huffman are incredibly cheap things with hardware support, too.
     
     zstd requires arithmetic coding which is ass to work with.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 19-Aug-2026 11:35:55 JST iced depresso

     in reply to
     @a1ba its really pareto-optimal, though yan is very competent nontheless.
   • Embed this notice
     :umu: :umu: (a1ba@suya.place)'s status on Wednesday, 19-Aug-2026 11:35:57 JST :umu: :umu:

     in reply to

     • iced depresso
     @icedquinn oh I remember you posted something about orz a while back
   • Embed this notice
     mischievoustomato@0.5dollah.click's status on Wednesday, 19-Aug-2026 12:09:32 JST mischievoustomato

     in reply to

     @a1ba i've actually been doing benchmarks recently and that's the best thing you can call zstd. it's just great.