Charlie Marsh charliemarsh 🔥 popular repo java-ml

ah yes, @charliermarsh best known for his java-ml repo

Bun v1.3.13 ships tomorrow.

Many memory usage & reliability improvements.

It’s not precisely accurate that we stream tarballs to disk. It is extracted in-memory to a temp folder. No tarball is saved.

Previously, Bun didn’t start extracting until the entire tarball finished downloading. compressed and uncompressed tarballs stayed in memory too long.

bun install · peak memory Streaming tarball extraction. 3.44GB 0.2GB 17× less memory Before After Repo with large packages · macOS arm64

In the next version of Bun

bun install streams tarballs to disk

In a large repo, this reduced memory by 17x

So this is effectively a lazy in-memory bytecode cache

The vast majority of what makes test isolation slow is time spent parsing modules. Not evaluation. Parsing. That’s why this works so well.

JavaScriptCore has both an input sourcecode cache and an in-memory bytecode cache. Since JavaScript is lazily parsed, both are needed - and we need to make sure that the cache is hit every time.

To do that, we add another cache that lets us skip the transpiler & disk reads

The trick here is adding one more level on top of JavaScriptCore’s multi-level code caching, and serializing ESM exports/imports.

The main tradeoff here is after compression it’s about 20% larger than VLQ-encoded compressed equivalent. Fortunately, compressing sourcemaps is unnecessary for server-side JavaScript.

Image from Twitter

Decoding now costs close to 0. Lookups cost about 6% more. Encoding gets faster.

_tsc.js (563k mappings) resident after first .stack Mapping.List (main) ~11.3 MB (20 B/mapping) LEB128 stream (commit 1) 2.92 MB (5.4 B/mapping) bit-packed windows (this) 1.29 MB (2.41 B/mapping)

In the next version of Bun

Source maps use up to 8x less memory

Quote Tweet: https://twitter.com/i/status/2044686990051045540

People have been asking for this for years.

It wasn't clear how to make it performant enough to be shippable, until earlier today

Quote Tweet: https://twitter.com/i/status/2020373504189952508

Mar 6: 167 lbs ( -12% )
Apr 14: 156 lbs ( -19% )

close to goal weight

if you’re choosing > 250 attributes for one span

please explain & give an example

me asking can the bits be packed to reduce memory usage or must it be slow & huge

what’s the max length of individual attribute names and values you’ve use in otel?

what’s the max number of attributes you have personally used at work in OpenTelemetry spans?

Bun v1.3.12 ships tonight

her: wait how do you track the calories
me: believe it or not, also claude. I send it a pic of the menu

I’m such a shill (2/2)

me at nopa on a first date

her: how’d you pick this restaurant?
me: I asked claude

[waitress compliments her pants, clothes come up]

her: where do you shop?
me: lost a bunch of weight, had to rebuy everything. sent claude a pic of me and it picked

[few min later, looking at menu] (1/2)

Image from Twitter

love seeing @elysiaJS double weekly npm downloads since December

Quote Tweet: https://twitter.com/i/status/2040257344404410431

claude, add a markdown pretty printer for terminals to Bun & don't make any mistakes

Bun v1.3.12 ships on Monday

os: cgroup-aware availableParallelism / hardwareConcurrency on Linux by Jarred-Sumner · Pull Request #28801 · oven-sh/bun Routes navigator.hardwareConcurrency, os.availableParallelism(), and bun.getThreadCount() through WTF::numberOfProcessorCores(), which is now min(sysconf(_SC_NPROCESSORS_ONLN), sched_getaffinity co...

In the next version of Bun

Threadpool & JIT threads now respect cgroup CPU limits instead of physical cores. This improves resource utilization in Docker & k8s

https://github.com/oven-sh/bun/pull/28801

earthquake

Faster stripANSI by Jarred-Sumner · Pull Request #28767 · oven-sh/bun What does this PR do? Speeds up Bun.stripANSI, Bun.stringWidth, and the shared ANSI parsing helpers used by Bun.sliceAnsi / Bun.wrapAnsi / Node's readline.getStringWidth. Five changes: 1. 4×-un...

PR & benchmarks

https://github.com/oven-sh/bun/pull/28767

┌────────────────────────────────────────┬─────────┬─────────┬─────────┬──────┬────────┐ │ Benchmark │ v1.3.10 │ v1.3.12 │ npm │ Δ │ vs npm │ ├────────────────────────────────────────┼─────────┼─────────┼─────────┼──────┼────────┤ │ stringWidth: UTF-16 hyperlink (440 KB) │ 2.00 ms │ 180 µs │ 743 µs │ 11× │ 4× │ ├────────────────────────────────────────┼─────────┼─────────┼─────────┼──────┼────────┤ │ stripANSI: 1 KB plain text │ 65 ns │ 17 ns │ 32 ns │ 3.9× │ 1.9× │ ├────────────────────────────────────────┼─────────┼─────────┼─────────┼──────┼────────┤ │ stripANSI: OSC 8 hyperlink │ 60 ns │ 45 ns │ 66 ns │ 1.3× │ 1.5× │ ├────────────────────────────────────────┼─────────┼─────────┼─────────┼──────┼────────┤ │ stringWidth: ANSI hyperlink (445 KB) │ 135 µs │ 120 µs │ 1.38 ms │ 1.1× │ 11× │ ├────────────────────────────────────────┼─────────┼─────────┼─────────┼──────┼────────┤ │ stripANSI: bash out

In the next version of Bun

`Bun.stringWidth` gets up to 11x faster
`Bun.stripANSI` gets up to 3.9x faster

guess what’s coming in bun v1.4

Previously, async stacktraces were only supported for JS functions. Now it works for either natively implemented functions or JS functions.

❯ bun-new a.js # New 1 | import fs from "node:fs"; 2 | 3 | const p = "fake"; 4 | 5 | export async function main() { 6 | await fs.promises.lstat(p); ^ ENOENT: no such file or directory, lstat 'fake' path: "fake", syscall: "lstat", errno: -2, code: "ENOENT" at async main (/Users/jarred/a.js:6:21) Bun v1.3.11-debug+5b14b04a8 (macOS arm64) ~ ❯ bun a.js # Previous ENOENT: no such file or directory, lstat 'fake' path: "fake", syscall: "lstat", errno: -2, code: "ENOENT" Bun v1.3.11-canary.1+2e610b140 (macOS arm64)

In the next version of Bun

Async stacktraces are supported on native APIs like node:fs, Bun.write, node:http, node:dns & more.

This makes debugging easier

The curmudgeons hating on it as nothing new are proven wrong by all the cool stuff people are building.

If it was easy before then typography on the web would not be so uniform and nobody would’ve cared enough to try Cheng’s library.