Packed Data Support in Haskell

https://arthi-chaud.github.io/posts/packed/

77•matt_d•7mo ago

Comments

nine_k•7mo ago

> Introducing the ‘packed’ data format, a binary format that allows using data as it is, without the need for a deserialisation step. A notable perk of this format is that traversals on packed trees is proven to be faster than on ‘unpacked’ trees: as the fields of data structures are inlines, there are no pointer jumps, thus making the most of the L1 cache.

That is, a "memory dump -> zero-copy memory read" of a subgraph of Haskell objects, allowing to pass such trees / subgraphs directly over a network. Slightly reminiscent of Cap'n Proto.

90s_dev•7mo ago

We are always reinventing wheels. If we didn't, they'd all still be made of wood.

Zolomon•7mo ago

They mention this in the article.

spockz•7mo ago

It reminds me more of flat buffers though. Does protobuf also have zero allocation (beyond initial ingestion) and no pointer jumps?

cstrahan•7mo ago

No, one example of why being variable sized integers.

See https://protobuf.dev/programming-guides/encoding/

carterschonwald•7mo ago

One thing that sometimes gets tricky in these things is handling Sub term sharing. I wonder how they implemented it.

tlb•7mo ago

> the serialised version of the data is usually bigger than its in-memory representation

I don’t think this is common. Perhaps for arrays of floats serialized as JSON or something. But I can’t think of a case where binary serialization is bigger. Data types like maps are necessarily larger in memory to support fast lookup and mutability.

nine_k•7mo ago

I suppose all self-describing formats, like protobuf, or thrift or, well, JSON are bigger than the efficient machine representation, because they carry the schema in every message, one way or another.

IsTom•7mo ago

If you use a lot of sharing in immutable data it can grow a lot when serializing. A simple pathological example would be a tree that has all left subtrees same as the right ones. It takes O(height) space in memory, but O(2^height) when serialized.

gitroom•7mo ago

honestly i wish more stuff worked this way - fewer hops in memory always makes me happy

lordleft•7mo ago

This was very well written. Excellent article!

NetOpWibby•7mo ago

Is this like MessagePack for Haskell?

We built another object storage

Java FFM zero-copy transport using io_uring

How exchanges turn order books into distributed logs

macOS 26.2 enables fast AI clusters with RDMA over Thunderbolt

AI is bringing old nuclear plants out of retirement

Sick of smart TVs? Here are your best options

Photographer built a medium-format rangefinder, and so can you

Apple has locked my Apple ID, and I have no recourse. A plea for help

GNU Unifont

A 'toaster with a lens': The story behind the first handheld digital camera

Beautiful Abelian Sandpiles

Rats Play DOOM

Show HN: Tiny VM sandbox in C with apps in Rust, C and Zig

OpenAI are quietly adopting skills, now available in ChatGPT and Codex CLI

Computer Animator and Amiga fanatic Dick Van Dyke turns 100

Will West Coast Jazz Get Some Respect?

Formula One Handovers and Handovers From Surgery to Intensive Care (2008) [pdf]

Show HN: I made a spreadsheet where formulas also update backwards

Freeing a Xiaomi humidifier from the cloud

Obscuring P2P Nodes with Dandelion

Go is portable, until it isn't

Ensuring a National Policy Framework for Artificial Intelligence

Poor Johnny still won't encrypt

YouTube's CEO limits his kids' social media use – other tech bosses do the same

Slax: Live Pocket Linux

50 years of proof assistants

Gild Just One Lily

Capsudo: Rethinking sudo with object capabilities

Google removes Sci-Hub domains from U.S. search results due to dated court order

String theory inspires a brilliant, baffling new math proof