Failure analysis
Plan recovery around an interrupted write
Filesystem recovery and application recovery protect different things on an unattended edge computer.
Monthly archive
2025 archive
Essays and guides, organised by their coverage period.
Retrospective essays are grouped by the month they cover. They were written and added on November 2, 2026; that is the publication date shown on each article.
December
1 article→
November
1 articleDeployment design
A running process is not a ready model service
Connect startup, readiness and restart policy to the application’s actual state.
October
1 articleArchitecture notes
Give a small inference server a queue policy
Accepted requests, waiting requests and rejected requests need explicit limits before throughput is useful.
September
1 articleDiagnostic notebook
Read the operator trace before changing inference settings
A runtime profile can turn a vague CPU complaint into a testable performance question.
August
1 articleMemory analysis
Use zswap to investigate pressure, not erase a memory limit
Compressed swap can change a workload’s failure behaviour, so evaluate responsiveness alongside memory use.
July
1 articleStorage analysis
Cold start is a storage workload of its own
Measure the time to a usable application before using an SSD’s headline transfer rate as an explanation.
June
1 articleThermal design
Design the cooling check around the enclosure
Use the assembled system and a sustained task to decide whether a thermal design has enough margin.
May
1 articleDeveloper workflow
Pin the inference container and keep an update path
A reproducible image needs a deliberate replacement process to stay useful.
April
1 articleConversion notes
An ONNX export needs an input contract
Preserve preprocessing and expected outputs before treating model conversion as complete.
March
1 articleMemory planning
Build a memory ledger before increasing context
Weights, generation state and the rest of Linux need separate entries in a small computer’s memory budget.
February
1 articlePipeline design
Draw the camera pipeline before changing the accelerator
A stage-by-stage latency budget makes the original AI HAT+ easier to assess for a specific vision job.
January
1 articlePlatform briefing
A Jetson power mode belongs in the build record
Treat the Super update as a configuration change whose benefits depend on the rest of the system.