
Founder of @CloudWorldModel. Estimate what an infrastructure change does to cost, latency, and errors before you ship it. Ex-Oracle Cloud Engineer.
Stamford, CT
Joined February 2009
- Tweets4.9K
- Following598
- Followers911
- Likes6.1K
PSN slowed 33 minutes after the GTA VI unlock. Whatever Sony's cause was, this is the trap we see in models: autoscaling adds servers, and every new server opens more connections to the same database. You can test that before launch night.
We simulated it: a game-launch sign-in rush on an assumed topology, not Sony's real system.
At 50x normal load, servers autoscaled 4→22, but 40.6% of sign-ins still failed. Most of that was the shared account DB running out of connections. p99 went from 188 ms to 1.9 s.
1
199
GTA VI launch week: the parties are planned months ahead. The sign-in capacity should be too.
Cardi B brought Vice City to life! The rapper hit the pink carpet at Miami’s Booby Trap on the River to celebrate “GTA VI,” with G Herbo and Freddie Gibbs joining the party. 🎮🔥

ALT I'm sorry, I can’t identify or describe the person in the image, but I can provide a general description. The image shows a person standing in front of a colorful mural, wearing a vibrant, intricately designed outfit with a fan-like headpiece. The individual holds a small clutch and is posing indoors with a lively backdrop.
37
The app autoscaled. The database pool didn't care.
That's the result I want sitting next to every "just add app servers" plan.
Modeled on an assumed topology, and we've measured a database wall like this on real AWS too.
We simulated it: what breaks at 1,000 vs 100,000 users, on an assumed topology (ALB → app autoscaling 2–10 → Redis + Postgres), not anyone's real system.
~100 RPS: p95 59 ms, 0% errors.
~10k RPS: all 10 app nodes maxed, Postgres at 100% CPU, 82.2% errors (72 points from the DB pool), p95 1,029 ms.
The app scaled. The DB pool broke first.
We also hit a database wall on a real AWS load test — adding app servers didn't raise the breach point:
cloudworldmodel.ai/blog/why-… xitter.gq/cneuralnetwork/status/…
25
Pre-emptive fixing is next. Why monitor and wait for something to break when you can test the change before it ships?
AI has transformed how software gets built. Now, it’s positioned to transform how software stays running.
@resolveai’s new SRE-1 is trained to find the root causes of production incidents. The results show the potential for specialized training to improve accuracy at a lower cost per investigation.
Congratulations to @spirosx, @maynkag and the team!
2
39
Load testing is an example of a type of pre-emptive fixing.
However, running your infrastructure through a model that behaves like the cloud can show you first, for a lot less. cloudworldmodel.ai
13
"An alert has no test" is the core problem. Simulation could give you one: inject a known fault (lock contention, failover surge, capacity loss) into a modeled system and you get incidents with ground-truth root causes, as many as you want. Could that work as an RL environment for SRE agents?
Introducing SRE-1 a 27B model that dominates the Pareto frontier of accuracy/cost on production debugging, beating frontier models in a coding-adjacent domain
Built by 11 people on open weights models + domain expertise. Final run: 7.4 days on 512 GPUs
1
125
Kevin Brown retweeted
We simulated the Microsoft 365 outage pattern: a lock-contention cascade took an assumed 7-worker gateway from 6 up to 0 in 5 steps. Cutting traffic didn't recover it; stopping new work did. Assumed topology, not Microsoft's system.
Kevin Brown retweeted
Why did the Microsoft 365 outage drag on? In our sim, cutting traffic to 240 RPS wasn't enough. Restarted workers got swamped by queued work and seized again. Only stopping new work let 6 of 7 rejoin.
Assumed topology, not Microsoft's. Their report notes the same queued-traffic swamping.
Microsoft's preliminary M365 report: a latent lock-contention bug, not just too much traffic, cascaded the Nov 12 outage after a load shift and ~14% of gateway capacity already out. We modeled the pattern on an assumed topology, not their real system.
We simulated a simplistic version of the Microsoft Outage: a lock-contention cascade, not just capacity exhaustion.
Assumed 7-node gateway, not Microsoft's. 1 of 7 out. Nodes at 450 RPS (cap 500, lock 425) seized one by one. 5 steps later: 0 of 7 up, 0 of 2,700 served.
Load the outage yourself:
cloudworldmodel.ai/workspace…
1
2
101
If there was one person i’d be interested in speaking to it would be Madison Huang. Hey @boardyai introduce me to Madison. Thanks!!!
1
144
There’s merit to open source and to keeping your code closed.
@floci_io is an open-source AWS emulator. AWS recently announced its own emulator. I don’t know if AWS looked at Floci, but it highlights a trade-off: anyone, including much larger companies, can study what you build.
github.com/floci-io/floci
1
75
I’m here waiting on RCA for the Microsoft 365 (Teams, Outlook, etc) outage. I checked this morning and still haven't seen an official one.
I keep looking for the RCA every few hours.
Reviewing RCA’s is interesting to me.
I’m aware that i’m strange. Hehe 🤷🏾♂️
1
407
Kevin Brown retweeted
How accurate is cloud simulation? We load-tested it and published the misses too.
AWS: 93.8% on /accuracy.
GCP, one topology in us-central1 at 300–500 RPS: 5 of 6 P95 holdout runs within 20%, 1 within 35%, all 6 P50 within 20%. Measured P95 was slower than predicted at every rung.
cloudworldmodel.ai/blog/aws-…
There are lots of startup products that feels like they works because it does the visible thing it promised but when you look at the outcome It just doesn’t work for anyone involved.
We want the actual outcome to be real.
1
52
An example would be the show love is blind on Netflix. If you get to the reunion part of the show and no one is still married, can you say love is blind?
Yes, no, maybe?
Basically, Cloud World Model wants those that use our predictions to be married by the end of the show.
I hope that makes sense.
1
60
Nice.
We are not an emulator.
We are building a World Model of how the Cloud behaves.
Today we're launching AWS Emulator, a free local emulator of AWS services for building and testing your apps offline. Every line of its code was written by autonomous AI agents running on Amazon Bedrock AgentCore. Hundreds of agents, organized into three teams, write the tests, find the issues, and fix the code, each change passing through an automated merge gate that checks compilation, security, coding standards, and fidelity to real AWS. go.aws/3Xk2dMV
1
94













