Architect's Reference
Lookup tables for the decisions you make more than once. Articles are written once; these are revised whenever the underlying numbers move — and every page carries the date it was last verified, plus the sources it was verified against.
6 references · 34 questions answered · 63 terms defined · 3 tools · Last reviewed
Searches 106 pages, questions, terms and tools
Performance & Latency
1 entryDistributed Systems
2 entries
Referenceref.02The Distributed Systems Failure Catalog
Fourteen ways distributed systems fail — trigger, telemetry signature, mitigation, and the fix that makes it worse.
Reviewed 16 min
Referenceref.03Consistency Models Reference
Linearizability to eventual, serializable to read committed, CAP to PACELC — and what each real database actually gives you by default.
Reviewed 12 minQuality Attributes
1 entryData Stores
1 entryAgent Engineering
1 entry34 questions this desk answers
If you came here mid-decision, start here rather than with a document title. Each one links to the answer itself, not the page it sits on.
The Distributed Systems Failure Catalog
- What is the difference between a thundering herd and a cache stampede?
- Why does exponential backoff not stop a retry storm?
- How do you recover from a metastable failure?
- Why are even-numbered clusters a problem for split brain?
- What is a gray failure and why do health checks miss it?
- What is coordinated omission in latency measurement?
Architecture Fitness Functions Catalog
- What is an architecture fitness function?
- How is a fitness function different from a normal test?
- How do you add fitness functions to an existing codebase without failing every build?
- Which fitness functions should a team add first?
- What are fitness functions for LLM and agent systems?
- Why is repository-wide test coverage a poor fitness function?
Agent Architecture Patterns
- What is the difference between an AI workflow and an agent?
- When should you use a multi-agent architecture?
- Why does the evaluator-optimiser pattern often fail?
- What is the difference between orchestrator-worker and parallelisation?
- What controls does a tool-use loop need in production?
- How do you reduce the cost of an LLM agent system?
Next door
Frequently asked
How current are these numbers?
Every page carries the date it was last verified and a changelog of what changed on that date, so you can see a figure's age before you rely on it. Pages are revised when the hardware, the practice or a vendor default moves — not on a publishing schedule.
Can I cite this in a design review or ADR?
Yes, and each page lists the sources every figure was checked against, with the date it was checked — cite those alongside the page so the argument survives without you. Treat the numbers as order-of-magnitude inputs that frame the trade-off, not as a prediction of what your system will do.
Can I feed these pages to an AI assistant?
Yes. Every reference page serves its full text as plain markdown at its own URL with /llms.txt appended, so you can paste the whole table into a coding agent or an architecture chat instead of asking it to recall numbers it will get wrong by three orders of magnitude.
Machine-readable: append llms.txt to any reference URL for the full text as markdown — for example /reference/latency-numbers-2026/llms.txt. Paste one into a coding agent rather than asking it to recall the numbers from memory.



