~/series/
Distributed Systems in 30 Days
30 lessons published · new lesson added daily.
1
Day 1
Using Back-of-the-Envelope Math to Estimate Agent Pipeline Performance
2
Day 2
How the CAP Theorem Applies to Multi-Agent LLM Systems
3
Day 3
Why some agent bugs are CAP trade-offs, not hallucinations
4
Day 4
PACELC: the latency trade-off CAP theorem leaves out
5
Day 5
Why a multi-agent review quorum isn't really fault-tolerant
6
Day 6
Why Raft leader election doesn't fit multi-agent LLM pipelines
7
Day 7
Using Raft leader election to stop duplicate agent dispatch
8
Day 8
How replication lag makes a database look like it lost data
9
Day 9
Why 'last write wins' loses data, and how vector clocks help
10
Day 10
Why multi-agent bugs come from consistency, not timing races
11
Day 11
How an Untested Saga Rollback Caused a Double Refund
12
Day 12
Why an Agent's 'Ready to Commit' Can Be Wrong After Truncation
13
Day 13
How a reasoning agent breaks saga idempotency and refunds twice
14
Day 14
Why exactly-once delivery still needs your own idempotency key
15
Day 15
Why retrying a non-idempotent tool call charges a customer twice
16
Day 16
How a cold cache slice after adding a hash-ring node overloaded the DB
17
Day 17
How a single orchestrator agent bottlenecks a fan-out pipeline
18
Day 18
How a bloom filter's false-positive rate climbs as it fills
19
Day 19
Gossip protocols
20
Day 20
Liveness vs Progress Detection
21
Day 21
Agent split-brain
22
Day 22
Idempotency keys for agent tool calls
23
Day 23
Dead-letter queues
24
Day 24
Backpressure in agentic teams
25
Day 25
Pub/sub hidden contracts
26
Day 26
Choosing a shard key
27
Day 27
Replica lag and read-your-writes
28
Day 28
Write-ahead logging for coding agents
29
Day 29
Physical clock skew
30
Day 30
Distributed computing fallacies for agents