https://kernwise.com/writing/ Technical notes from Kernwise on building and fixing systems: databases, data pipelines and applied machine learning. en Thu, 15 Oct 2026 00:00:00 GMT https://kernwise.com/writing/a-model-score-is-not-a-probability/ https://kernwise.com/writing/a-model-score-is-not-a-probability/ Thu, 15 Oct 2026 00:00:00 GMT Why a classifier with a usable AUC can still mislead the decision layer that reads its scores, and how expected calibration error and isotonic regression measure and close the gap. Machine learning Calibration Decisioning https://kernwise.com/writing/records-export-safe-to-run-twice/ https://kernwise.com/writing/records-export-safe-to-run-twice/ Thu, 15 Oct 2026 00:00:00 GMT Design notes on exporting records and scanned documents into a search index, covering stable keys, a ledger per stage, retries that know which failures are temporary, and partial reruns chosen by configuration. Data pipelines Search Azure https://kernwise.com/writing/before-you-migrate-to-postgres-17/ https://kernwise.com/writing/before-you-migrate-to-postgres-17/ Sat, 26 Sep 2026 00:00:00 GMT Six checks to run before a major-version upgrade, while rolling back is still cheap. Postgres Latency https://kernwise.com/writing/pool-exhaustion-is-not-database-latency/ https://kernwise.com/writing/pool-exhaustion-is-not-database-latency/ Sat, 26 Sep 2026 00:00:00 GMT A short explanation of why your dashboards lie about which layer is slow, with the four checks that distinguish them. Latency Platform