The deploy is the cache
2026.197Presigned URLs cap at seven days -- a protocol ceiling. A static site mints URLs at build and serves them until the next deploy. The clock starts at mint.
Presigned URLs cap at seven days -- a protocol ceiling. A static site mints URLs at build and serves them until the next deploy. The clock starts at mint.
The spec says what should happen. Code says what does. Only one of them actually runs.
Every team eventually builds a retry mechanism. They all build it during an incident. None of them write the spec first.
Method names encode intent, not contract. The gap is structural, and the fix is not better names.
Hiring isn't broken. It was never designed. Two sides fill different forms with different fields, and the join fails before the conversation starts.
The deploy broke. The checks were green. The mechanism measures what it can see - and what it can't see is everything that breaks.
connection refused is not a diagnosis. It's a location marker. The error is telling you where the failure surfaced, not where it started.
The config said workers: 4. The machines had 32 cores. The number had been there since 2017. Nobody chose it. Nobody knew why.
An attention mask drawn in ASCII, and the architectural trade it names.
What Tesla's 3-6-9 has to say about scaling systems -- and the invariants that hold while the rest moves.
The abstractions rotate every few years. The failure modes don't.
Prompt engineering as product specification, confidence as contract -- what changes when document processing becomes understanding, not extraction.
How a cascade of exact, prefix, and fuzzy strategies resolves broken OCR output into valid addresses -- with confidence scoring so downstream systems know what to trust.
A regulation arrives. The work looks like encryption at first. It is not.
The three-state machine that keeps a failing dependency from taking down everything else -- with sony/gobreaker, a custom atomic implementation, and retry composition.
Cache-aside, write-through, write-behind -- and the three problems that make caching hard: stampedes, coordinated expiry, and invalidation.
Streaming replication, automated failover, and the operational lessons that aren't in the PostgreSQL docs -- split-brain, silent lag, and why DR drills matter more than DR plans.
The board put us ahead of every customer ticket - until it stayed green while the data underneath it went wrong.
The index was there. The query was right. Postgres stopped using it anyway.
Service discovery, health checks, and dynamic config across a growing fleet of microservices -- and the moment KV store got too popular.