Articles in this category. Show all categories
Administrators of a city government's personnel system must each touch only their part of it: everything, one agency, or one unit. I made the scope data — a grant is user, role and scope — checked by one rule, requested and approved by different people, and expiring on its own.
An organisation chart that changes by regulation is still master data for relational records. I keep it in SQL as a table that points at itself, with versions per agency and effective dates, read it with one query and recursive CTEs, and leave drawing it to the interface.
A personnel database that never enforced its rules had to become clean data for a new system while the old one kept running. Ordered, repeatable phases behind a read-only door, explicit maps, honest placeholders and a report instead of a crash made that possible.
Templates come from real stacks, and copies quietly carry their data, keys, open databases and root-owned paths. My database, object storage and application generators leave the old data, keys and access behind, and read each application's facts from the application itself.
Every project gets its own OpenTelemetry Collector, Elasticsearch and Kibana from one command, showing data on dashboards from the first start, by owning the data directory, starting services in the order they need, and keeping the official mappings on one node.
Opening a broker to browsers and devices on the internet is less about encryption than identity. Each entrance now accepts one kind of credential, every token is bound to its own name and addressed to this broker, and TLS ends in nginx.
Every new project needs a gateway, sign-in, and a first administrator before its backend exists. One generator command now builds that stack, working on first boot, by handling the values that fail silently when set by hand and a rerun that would keep the old database.
SMTP and IMAP credentials fail late and far from their cause: a reset email that never arrives. A small check proves both directions first, with a tokenised test message, a read-only inbox check, and broken config kept distinct from a broken server.
A 4 GB laptop GPU once sent every heavy job to a borrowed Colab runtime. Deciding per piece of work instead — by a measured number, by reliability, by a change of method — split one project across three places and deleted 9975 lines of machinery on the way home.
A borrowed Colab GPU ends every session by time limit or daily quota, often in the middle of hours of image generation. Picking up where it stopped meant treating each new runtime as a machine that knows nothing — and may not even be the same machine as yesterday's.
Get an email when a new article is published.