rockwotj

9fs is better than NFSv3 but NFSv4 is a different beast and more powerful

Honestly the biggest limitation is its quite a big protocol and can be complicated, but I don’t see any major advantages you could do yourself protocolwise over a custom FUSE thing except for just simplification. In my case I am hoping the kernel client can save me a lot of work of building a good client

Wouldn't 9fs/plan9fs be more suitable for this use case from my limited understanding though?

Can you elaborate on why? NFS is really a great protocol with a lot of tricks to reduce round trips

author here! Glad you liked the post :)

Motivation for combining Chorus and SlateDB for NFS

I work on an AI agent (tasklet) and we give every agent a linux machine. Having durable storage that is cheap, fast and multi-tenant is really important for our product. NFS is a great protocol (if complicated), and object storage is just the cheapest. But making it fast and reliable is key.

other use cases

Any use case for SlateDB that you are willing to pay more for less latency but keep disaggregated storage without another system.

GCP specific

Actually AWS and Azure zonal storage also support append operations, so I think the approach could be extended to all three major clouds. I don’t have a need for that ATM

pricing

Probably worth a whole separate blog TBH. It would be cheaper than Kafka but more expensive than just using the built in WAL for SlateDB or OSWALD

But this is model behavior and just a public issue tracker which claude code has just without code? I don’t see how it’s any different than https://github.com/anthropics/claude-code for these issues.

I do appreciate that codex is open source generally, but I don’t think it matters for this class of issue as the model is closed still

Because it’s harder. It’s a stateful protocol and has a lot more features

Nothing wrong with that, but you should remove the “at rest in S3” footnote from the write latency on the frontpage of the website, because that is not what is measured

Yeah in this case the footnote to the write latency specifically says “at rest in S3”, which is what caused me to go look at the source. To be clear I have no problem with the ZeroFS of only flushing on fsync.

I am very excited for object storage first systems like this to leverage low latency zonal storage for write ahead logs to keep the disaggregated storage but greatly reduce write latency. That ends up being more expensive, but is likely a good tradeoff in lots of cases I have seen

The sub-millisecond writes with data in S3 is false and impossible. If you look at the benchmark the fsync is not timed, so this is just the latency of either the network or in kernel file operations depending on the mount settings

Oh of course all deepseek and glm are. Multiple people have seen GLM self report that it is claude, which makes it super obvious.

I think the surprising thing is I expect flash to be a pure distillation and strictly worse quality but clearly it’s more nuanced than that.

I have also found deepseek flash beat pro in some of my own internal evals for tasklet.ai it’s really surprising and I don’t understand it

will finally fracture.

If true, it will only be replaced with something else. With what is anyone’s guess.

I actually just started doing this by having Fable roleplay as Jeff Dean and to use Codex as Sanjay driving the implementation and have them go back and forth. Works really well and it’s cool to see AI pair program

This seems really great! I am a long time Bazel fan, but it’s not perfect, and I have always thought there was something simpler than bazel that would give a lot of these kinds of benefits. Having something that works for a hybrid rust/typescript codebase would be great.

I encourage you for you dev command to work with frontend setups like vite so it’s unified with backends and still supports HMR

IIRC (it’s been awhile) this was an optimization that was inspired by someone noticing all of Google spent a lot of time in vDSO using https://research.google/pubs/google-wide-profiling-a-continu... Google generally doesn’t have any (well that I have ever heard of) hard realtime requirements like the high frequency trading or HPC systems do.

Cool stuff thanks for sharing!

Related reading is absl’s linear approximation of gettime from cycle counters, which I thought is a neat trick: https://github.com/abseil/abseil-cpp/blob/351086314d46e73d43...

Production-grade 17 operation types. 40+ semantic validation rules. Auth, rate limiting, PostgreSQL storage.

Man where in the LLM training data did the “production ready “ come from? That whole list screams AI. Humans want social proof, not a list.

https://www.datadoghq.com/blog/datadog-time-series-foundatio...

https://moment-timeseries-foundation-model.github.io/

https://arxiv.org/abs/2403.07815

A friend at work used one to predict when our CEO would post in Slack, which is verry entertaining to see if correct.

Probably the better link: https://research.google/blog/a-decoder-only-foundation-model...

which isn't saying that much

I mean I also think this move doesn’t make sense, but I always find these type of comments interesting. Do people think they could do better in Mark’s shoes?

Or it’s compiled in the cloud?

One would think it must otherwise there are all these issues with compiling, signing tc if they don’t have xcode installed etc. I would guess it’s some webview wrapper with a layer to expose desktop app functionality

Fundamentally, the "guessing when its your turn thing" needs to be baked into the model. I think the full duplex mode that Moshi pioneered is probably where the puck is going to end up: https://arxiv.org/abs/2410.00037

A category of SaaS companies are cooked in the age of AI - why not just have the perfect purpose built thing instead of fighting bloated enterprise software with a billion customizations and knobs that makes everything complicated?

I am sure they made a bid. The blog makes it sounds like he talked to multiple labs.

Technology does not determine the success of a company. I’ve seen amazing tech fail, and things strapped together with ducktape and bubblegum be a wild success.

https://rockwotj.com/blog/async-wasm-in-seastar/ - been meaning to write about this for a long time. Thanks for the nudge :)

I have done this integrating wasmtime into a a C++ seastar.io application. I’ll have to write a post about it

Claude is king for agentic workflows right now because it’s amazing at tool calling and following instructions well (among other things)

I thought everyone was just using open telemetry traces for this? This is just a classic observability problem that isn’t unique with agents. More important yes, but not unique functionally.

networking costs are so high in AWS I doubt this makes sense