Connor Brewster (@cgbrewster) profile picture

Christian - Rust enthusiast - Software engineer @ https://xitter.gq/t.co/kN13dcELYg

Oklahoma, USA
Joined July 2016
Photo by Connor Brewster (@cgbrewster)
This has been one of my favorite engineering efforts at Replit, we upgraded from containers to microVMs and rebuilt the entire control plane all without users noticing. The new infrastructure is far more secure and has unlocked powerful futures like Docker support!
3
17
354
Connor Brewster retweeted
Replit Agent switched from Temporal to Restate for three main reasons - Lower latency and cost efficiency: Restate executes durable steps inline, adding just a few milliseconds of overhead per step. This allows Replit to make thousands of operations durable without the overhead of scheduling each one as a separate activity. - Unique building blocks for stateful, concurrent agents: Virtual Objects and Durable Coroutines simplify coordinating long-lived agent sessions, parallel subagents, user steering, approvals, and cancellation, and make the architecture easier to extend. - Restate BYOC: Fully managed deployments in Replit's own cloud account keep sensitive data private and support both its multi-tenant cloud stack and dedicated enterprise offerings. With Restate, Replit can use fast, durable execution inside the agent loop, giving them a simpler, more resilient architecture. Replit now runs tens of Restate cells across regions and availability zones, and handled a major promotional traffic spike of nearly 25 000 durable actions per second per cell a few weeks after migration. Read the full story, including how Replit built its agent architecture on Restate 🧵
2
7
1
27
139,017
Connor Brewster retweeted
I just wrapped up my internship @modal. This summer I worked on container filesystems, speeding up agentic workloads by up to 4x and directory traversals by 12x. I wrote about this systems work behind it here: amitsai.com/writing/erofs/?v… Huge thanks to @amitpr, @jonobelotti_IO, the runtime team, and everyone at Modal for an amazing summer!
12
25
6
269
33,765
Connor Brewster retweeted
New post: strong vs weak sealing in AtomicLog + LogDrive. The interesting bit is where you put the fencing responsibility: in the storage write path, or after the write but before acknowledgement. That choice changes both the storage semantics you need and where/how you pay the coordination cost. jack-vanlightly.com/blog/202…
1
13
120
5,881
Connor Brewster retweeted
So many S3 performance misunderstandings. Let me clear this up for you all.
There is a lot of hype around putting the WAL on object storage. However, for low-latency systems, e.g. OLTP databases, object storage is a suboptimal primitive for the WAL: (1) PUTs are expensive for small WAL writes [bytes/$]. (2) Object stores are optimized for throughput, not latency, with write latencies often around 10-50 ms. Because every transaction must wait for its WAL record to become durable, every transaction inherits that latency. This points to a missing storage primitive. In our (@maxikuschewski @melhindiCS @viktor_leis) latest accepted research paper at VLDB, we looked at how to close that gap.
87
235
36
3,937
965,208
Determinate Nix 3.22.2 now out! 🎉 We've bumped upstream Nix to 2.35.2 and provided 5-12% faster evaluation with mimalloc, HTTP/3 binary cache fetching, faster `nix copy` in some scenarios, GC roots for `nix eval`, and @Replit auth for FlakeHub. This is the one you really want, folks ❄️ 🚀 🔗 determinate.systems/blog/cha…
1
5
25
884
Connor Brewster retweeted
There’s been some interesting discussion recently around WALs on S3, and separately around abstraction as reduction. I’ve done a write up on the LogDrive OSDI paper: how it extends the shared log protocol Virtual Consensus by decomposing the Loglet into separate sequencing and durability abstractions. This is research that came out from Confluent's Kafka-on-S3 work and next-gen K2 engine (which powers Freight Clusters). The interesting part is that reducing durability to a smaller abstraction makes it much more composable, making it easier to adapt the log to changing storage services/prices with only minimal code changes. jack-vanlightly.com/blog/202…
32
2
196
12,770
Connor Brewster retweeted
This is one @restatedev cell during the time when @Replit launched the free mode for their agent. Close to 25k durable actions/sec in a single cell. And there are quite a few cells. Makes this probably one of the highest volume durable execution use cases out there.
Something we @restatedev are incredibly proud of is powering Replit's agents with durable execution. Their scale and the sophistication of how their agent architecture uses durable execution is absolutely amazing. Truly an honor to be a part of that.
4
9
2
46
9,002
Connor Brewster retweeted
Proud to have partnered with Replit on this project. Replit Free Mode is powered by @restatedev ❤️
Replit Free Mode, powered by @OpenAI GPT-5.6 Luna. Let’s make intelligence accessible to everyone.
8
15
1
318
15,680
Connor Brewster retweeted
Something we @restatedev are incredibly proud of is powering Replit's agents with durable execution. Their scale and the sophistication of how their agent architecture uses durable execution is absolutely amazing. Truly an honor to be a part of that.
Replit Free Mode, powered by @OpenAI GPT-5.6 Luna. Let’s make intelligence accessible to everyone.
7
11
2
77
17,476
Connor Brewster retweeted
Replit Free Mode, powered by @OpenAI GPT-5.6 Luna. Let’s make intelligence accessible to everyone.
323
469
338
4,691
2,420,285
Connor Brewster retweeted
Segmented compaction is what happens when you combine the best parts of Cassandra's time windowed compaction and RocksDB's column families. It's dangerously awesome, and Jason wrote about how SlateDB does it: slatedb.io/blog/segment-orie…
1
9
56
3,452
Connor Brewster retweeted
Restate 1.7 is out. It is the biggest release in over a year! The star of this release is 𝗳𝗹𝗼𝘄 𝗰𝗼𝗻𝘁𝗿𝗼𝗹, powered by 𝘃𝗶𝗿𝘁𝘂𝗮𝗹 𝗾𝘂𝗲𝘂𝗲𝘀. Restate can now steer load across your the durable functions, agents, and event-driven services with flexible rules. Together with Restate's existing capabilities for durable execution, stateful functions (virtual objects), and reliable RPC/messaging, it forms a durable foundation for your backends and agents that goes much beyond traditional durable execution. And all of it in a single dependency-less binary. Read more here: restate.dev/blog/announcing-…
8
19
4
75
160,686
Connor Brewster retweeted
We are excited to launch 𝗥𝗲𝘀𝘁𝗮𝘁𝗲 𝗕𝗬𝗢𝗖 (Bring-your-own-Cloud) today. It is fully managed @restatedev, with all the features of Restate Cloud, but in your own account, in a dedicated VPC. Data never leaves your infra. Restate BYOC is already live in production with multiple customers, powering big workloads for AI agents and workflow backends. It has run loads of 100k durable actions/sec. And today we’re finally publicly announcing it. Learn more here, and let us know what you think! restate.dev/blog/announcing-…
17
20
6
79
202,662
Connor Brewster retweeted
Hundreds of thousands of you unleashed your creativity on Replit Agent this weekend -- completely free. Our infrastructure handled ~4× the usual load, with tens of thousands of agents running in parallel. A real test of our scale, and a powerful reminder of how passionate this community is. Thank you for building with us 🧡
53
46
23
306
56,532
Replit's deep dive into their snapshotting infrastructure, posted Thursday, is worth a read. It highlights a key shift: once agents operate on real application state, experimentation becomes the norm. Agents branch state, explore in parallel, and backtrack frequently. Assumptions that break traditional databases and storage systems. What struck me is how closely @Replit's architecture mirrors what we built with Fluid Storage at Tiger Data. Different implementations, same conclusion: databases need to be forkable, and that capability has to live in the storage layer. I wrote up a comparison on what this convergence means for agentic experimentation 👇
11
11
91
15,595
Connor Brewster retweeted
i wrote a blog post on how we built a REPL-based self-testing subagent to go from 20min -> 200 min of autonomous runtime in Agent 3 :) the main thing we wanted to solve was the problem of LLMs generating apps that looked like they worked but were really just high-fidelity mocks
3
8
6
67
31,933
Announcing UnwrapGuard: Stop production outages before they ship. Scan your Rust codebase for dangerous unwrap() calls and guarantee runtime stability. unwrapguard.replit.app
2
2
11
5,113
Built with Replit Agent design mode xitter.gq/amasad/status/19908594…
Announcing Design, a new Replit experience focused on beautiful UIs. The first non-slop AI design experience, powered by Gemini 3.0
200