One server, 100 users

Twitter clone · Level 10 intuition points

Step 1 of 3: 1Hold the load
Step 2 of 3: 2Survive the incident
Step 3 of 3: 3Set complete

Your Twitter clone just launched. About 100 people are trying it. You have one small server and a database.

This level is the warm-up. You will look at the parts, make a guess, run a load test and read the charts. Nothing is hard yet. Then one small thing goes wrong, and you fix it.

Set 1 of 2

Your first run

Look at the canvas. A request travels from the users (the client) to the app server to the database, and back.
Press Run load test. You make a quick guess first: how many requests per second, and which box fills up first.
Watch the boxes and the charts while the test plays. A box turns amber, then red, when its CPU fills up.
Open the Results tab. It says passed or failed, and the Why section shows the numbers behind it.

Test suites

not run yet

The design is validok so farnot run yet

One app server and a database. Nothing else to check yet.

Costs $30 a month or lessok so far, now $12 a monthnot run yet

Your design costs $12 a month now.

Holds 100 usersp95 under 500 ms, errors under 1%not run yet

Not tested yet. Press Run load test.

Survives: shared link at 130 slocked until the normal test passesnot run yet

Unlocks after your design passes the normal test.

Traffic

Users at the peak: 100
Each user sends: 1 request every 5 seconds
Peak requests per second: you work this out

Requirements

p95 latency (slowest 5% of responses): under 500 ms
Errors: under 1%
Budget: $30 a month

Words you will see

Client: Whoever uses your app: a phone, a laptop, a browser. Here they are the Users box. Each one sends a request every 5 seconds.
Server: A computer that runs your code. It takes a request, does the work and sends back an answer. Doing work uses its CPU.
Database: Where the data lives (the tweets and the users). For every request the server asks it a question and waits.
Requests per second (RPS): How many requests arrive every second. 100 users, one request each every 5 seconds, is 20 RPS.
CPU %: How busy a box is. 0% is idle and 100% is full. A full box cannot start new work, so requests wait in a line.
p95 latency: The time that 95 out of 100 requests beat. It shows what the slow users feel. A box that is almost full makes it jump. The limit here is 500 ms.
Error rate: The share of requests that failed. A request that waits more than 10 seconds gives up. The limit here is 1%.

What one request costs

The app server needs 6 ms of CPU for every request, and the database needs 6 ms to answer its question.
One CPU core has 1,000 ms of work to give every second. If the work that arrives in a second is more than the cores can give, the rest waits.
When the database shares the server's box, both jobs use the same cores: 12 ms a request. On its own box, each box only does its own part.

Prices per month: 1 vCPU $12, 2 vCPU $24, 4 vCPU $48. A vCPU is one CPU core you rent. A database box costs the same as a server of its size.

Your design

The database shares the app server's box. That costs nothing extra. Give it a box of its own with the Own database box button above the canvas.

Numbers are provisional estimates. The AI Mock covers Level 2 for now.

Your design

100×1$12DB

Users

100 users

at the peak

Users grow from 0 to 100 in 120 seconds, then stay for 60 seconds. The server starts the test ready, as a running app would be. The replay takes about 10 seconds.

Cost$12/ $30 a month

Load test

Run the load test to see the numbers and charts here.

Latency (response time)