See where your architecture breaks, and what it costs, before you build it.
- Design
- Simulate
- Find the bottleneck
- Fix
- Compare cost
Draw an architecture, put load on it and watch throughput, p99 latency, errors and monthly cost live. It runs in your browser; the demo below needs no sign-up.
Triple the traffic, watch the database tip over
This is the real simulation, running on this page: 400 requests/s through a load balancer to 4 API servers and one Postgres database. At 3× the database runs at about 130 % load, p99 goes from under 100 ms to about 1 s and about a quarter of the requests fail.
In the app, put a Redis cache in front of the reads: the database drops to about 36 % at 3×, p99 is back under 100 ms and nothing fails, for about $192/month more.
How it works
-
Design
Drag load balancers, services, caches and databases onto a canvas and connect them.
-
Simulate
Press play and set the traffic. Throughput, p99 latency, errors and monthly cost update live.
-
Find the bottleneck
Stackrig names the node that breaks first, not every node that waits on it.
-
Fix
Add a cache, a replica or a larger machine and see at once whether the design holds.
-
Compare cost
Every change shows its monthly price, so you see what the fix costs.
The model against real machines
We run the same designs on real cloud VMs under load and publish how far the model is off. These are the runs of 30 September 2026 on 2-vCPU machines: the mean deviation of the model from the measurement. What Stackrig shows you are model results, not measurements of your system.
-
at most 1 percent
Throughput
Both setups, at a constant rate and with random arrivals.
-
at most 7 percent
Latency, one CPU-bound service
p50, p95 and p99 at a constant rate.
-
23 to 27 percent
The same service, random arrivals
p50 23 %, p95 23 %, p99 27 %.
-
42 to 44 percent
Where it misses: a service in front of Postgres
p99 is off by 42 to 44 % and the median by 100 % and more: the real requests take under a millisecond, so small fixed costs dominate. Near a database's capacity, read Stackrig's tails as optimistic.
From our experiments
-
Self-hosted LLM or API: when does a rented GPU pay off?
1,008 output tokens per second one L4, gpt-oss-20b, 64 requests at once
-
How many users can a $12 server handle? We measured it
800 users on a $12 server
-
How many requests per second can one Postgres handle?
6,000 queries per second one Postgres, 2 vCPUs
What you get
-
Bottlenecks
The root cause is named, not every node that waits on it, with what would help.
Find one -
p99 latency
The tail comes from simulated requests, not from averages.
How it is checked -
Cost
Monthly cost from public price lists, each with its source and date. AWS on Free; GCP, Azure and Hetzner on Pro.
Price a design -
Chaos
Switch off a node or some of its instances, add latency or errors, then recover it.
Break something
Also in the app
-
Challenges
Get a design, a goal and a budget, fix it, and the playground checks your solution live. Three are free.
-
Templates to learn from
Twelve designs, from a URL shortener to Stackrig's own stack, each with one lesson you can reproduce in a minute.
-
Find my stack
Say what you are building and how much traffic you expect. You get three architectures sized for it, each one run in the simulation and priced on AWS.
Answer the questions
Stackrig itself runs on this architecture
- Cloudflare Pages
- a Cloudflare Worker
- Supabase Postgres
- Brevo for e-mail
How the numbers are made
-
A fast model, at any load
Rates, queues, utilization, errors and retries are computed per node and connection, so the cost of a step depends on the size of the design, not on the load: 10 requests/s and 10 million requests/s cost the same.
-
Checked against an exact reference
A discrete-event simulation that follows every single request serves as the reference. In every validation scenario the live model has to match it:
- within 5 %
- for throughput and utilization
- within 15 %
- for p50, p95 and p99
- within 0.3 %
- for the template in the demo above
Where it misses, the deviation is measured and listed, not hidden.
From the blog
-
Self-hosted LLM or API: when does a rented GPU pay off?
gpt-oss-20b on one rented L4, measured: what a million output tokens cost at 1 to 64 requests at once, and how busy the GPU must be to beat five per-token APIs.
-
Could not get lock /var/lib/dpkg/lock-frontend on Ubuntu 24.04
We started a second apt while a first one was installing on a fresh Ubuntu 24.04: it named who held the lock, waited, and went on by itself.
-
Package 'mysql-server' has no installation candidate on Debian 13
We made apt’s “no installation candidate” for mysql-server on a fresh Debian 13, read what apt offers instead, installed it and logged in.
Join the waitlist
The app is open, and Pro is not for sale yet. Leave your email address and we tell you when Pro is ready. We may also ask whether you would like to talk to us about how you design systems today. No newsletter.
You need no waitlist and no account to try Stackrig: the app is free to use in your browser. An account, by email or GitHub, is only for saving designs to the cloud.
The waitlist opens soon.