System Design Simple

How big apps really work, explained simply with diagrams. System design, AI coding tools (Claude, Codex) and web security. New short video + article every week.

[System Design Simple @SysDesignSimple]

In my small simulation, three new apps plugged into one AI client and its checked code never changed. Each app hands over a menu of its tools, and the descriptions on that menu decide what gets picked.

#AISystems #SystemDesign #AI

[System Design Simple @SysDesignSimple]

Two fans click the same seat at the same moment. Who wins?

Check-then-act leaves a gap in which two transactions both see a row as free. Close it with one conditional UPDATE and branch on the row count: PostgreSQL re-checks the WHERE on…

https://t.co/rjDWMNS92b

#SystemDesign

[System Design Simple @SysDesignSimple]

CDN push vs pull: why Netflix doesn't buffer in Kathmandu

A CDN keeps copies on edge servers near viewers, so most requests skip the long trip to the origin. A pull edge fetches a chunk on its first request, so the first viewer pays the…

https://t.co/WHYF9TBdDt

#SystemDesign

[System Design Simple @SysDesignSimple]

Streaming AI replies: why text arrives before completion

A language model writes its reply one token at a time, so an app can either wait for the whole reply or show each piece as it arrives. Streaming moves when the first words appear…

https://t.co/2PcPeiPODf

#SystemDesign

[System Design Simple @SysDesignSimple]

AI can't see a clock. So how did it tell me the time?

A model cannot read a clock: it only sees the tokens in its request. The app offers a tool, the model writes a structured request to call it, the app runs it and sends the result…

https://t.co/K3KxNW38ED

#SystemDesign

[System Design Simple @SysDesignSimple]

Load balancing algorithms: which server gets your request?

A load balancer picks a server per request: take turns (round robin), fewest unfinished requests (least connections), or a hash of a key such as the user ID. With mixed request…

https://t.co/rndbBmQjCD

#SystemDesign

[System Design Simple @SysDesignSimple]

99.9% uptime: 8.76 hours of downtime per year

99.9% availability, measured by time, allows 0.1% of a year down: 8,760 hours × 0.001 = 8.76 hours, about 8 h 46 min. Each extra nine divides the allowance by ten: 99% is 3.65 days, 99.99% is…

https://t.co/grMfGXqtZl

#SystemDesign

[System Design Simple @SysDesignSimple]

Series vs parallel availability: two servers in a row

When a request must pass through two servers in a row, both must be up, so availabilities multiply: 0.99 × 0.99 ≈ 98%, about twice the failures of one server. Two copies side by…

https://t.co/L4mHq4UdiY

#SystemDesign

[System Design Simple @SysDesignSimple]

My photo search demo held 20,000 fake photo vectors and checked under 4 percent of them per search, yet still found most of the best ten. An HNSW index jumps far on a sparse top layer, then walks close.

#AISystems #SystemDesign #AI

[System Design Simple @SysDesignSimple]

In our simulation, jobs that needed at most 2 ms of work averaged 31 seconds, because the shop called ten services one by one and one of them was slow. Posting each order once to a topic brought the fast jobs down to 1.87 ms.

#PubSub #SystemDesign #Messaging

[System Design Simple @SysDesignSimple]

50K fans, 5K seats. Design it.

Performance is the latency of one request at low load; scalability is how throughput and latency behave as load and resources grow. Queueing delay rises like ρ/(1−ρ) in the M/M/1 model, so run near 70%…

https://t.co/fFy9f9eePH

#SystemDesign

[System Design Simple @SysDesignSimple]

Service discovery: how 500 services find each other

Callers find live instances through a registry: instances heartbeat in, silent ones expire after a TTL, callers cache the list. Dead instances still get traffic for about TTL +…

https://t.co/6TGIyznETz

#SystemDesign

[System Design Simple @SysDesignSimple]

How can an AI answer with no network? On-device AI explained

On-device AI can answer with no network because the model executes on your machine: its weights, tokenizer and config are files on the disk, and a runtime does the math on the…

https://t.co/IUOg5Ldf0P

#SystemDesign

[System Design Simple @SysDesignSimple]

Publish/subscribe: why 2 ms jobs waited 31 seconds behind one slow service

When a publisher calls every subscriber in turn, its time per event is the sum of all the calls, so one slow subscriber sets the pace for everyone. In our…

https://t.co/2C2RhDJaeZ

#SystemDesign

[System Design Simple @SysDesignSimple]

AI citation checks: the link opened, but the quote was missing

A link that opens proves a page exists at that address; it says nothing about the quoted words. In our lab all 60 cited local files existed and were readable while 30 planted…

https://t.co/IGSqNheCGq

#SystemDesign

[System Design Simple @SysDesignSimple]

I searched car and keyword search found only 2 of 14 road-car sentences; the other 12 say vehicle, engine or hatchback. In our toy lab, an embedding built by counting shared sentences put 8 car sentences in its top 10, 6 of them never saying …

#AISystems #SystemDesign #AI

[System Design Simple @SysDesignSimple]

Our demo's 3 a.m. login rule raised 78.9 false alarms per 1,000 logins, and only 1.9% of its alerts were real. Giving each user a habit card of their own hours, country and devices cut that to 3.8 per 1,000 and caught 15 of 20 plant…

#Security #SystemDesign #CyberSecurity

[System Design Simple @SysDesignSimple]

1M rides a day. What do you do in the first 5 minutes?

Write functional requirements as user actions and non-functional ones as SLOs: an SLI (good ÷ valid events), a target and a window. 99.9% over 30 days leaves a 43.2 minute error…

https://t.co/ZkNjofG6nE

#SystemDesign

[System Design Simple @SysDesignSimple]

Claude Code cloud sessions: how the setup script, cache and network access really work

Each Claude Code cloud session runs in a fresh Ubuntu 24.04 VM. A cloud environment decides its network access, variables and a setup script. The…

https://t.co/CsbDWjDGQ3

#SystemDesign

[System Design Simple @SysDesignSimple]

Multi-tenancy: pool vs silo, tenant context, and noisy neighbor limits

Multi-tenancy serves several customers in one environment, with dedicated (silo) resources, shared (pool) resources, or a mix per layer. In pooled layers, a tenant ID…

https://t.co/JtsvJj7hii

#SystemDesign