Engineering Principles

earned, not adopted

Each principle has an origin story and at least one place it was applied again. Open one.

explanation
When two code paths can produce the same derived value, they eventually disagree, because retries, reordering and partial failures guarantee it. Consistency isn't achieved by synchronizing computations; it's achieved by making all but one of them impossible.
origin
Born from debugging a customer-facing metric that two services computed independently at Dynamox.
where i applied it again
Reapplied in workspace synchronization design: one owner per entity type, everyone else converges.
explanation
Standards don't fail because they're wrong; they fail because ignoring them is less work. A design system must out-compete local solutions on effort, so importing the component has to beat writing one.
origin
Building AQTech's first Design System as an intern with no authority to mandate anything.
where i applied it again
Same logic drove observability adoption at Dynamox: triage ownership designed to be lighter than ignoring errors.
explanation
A design whose correctness follows from structure (idempotency, single ownership, reversibility) needs no vigilance. A clever design that's correct under assumptions decays as the assumptions do.
origin
Contrasting the reconciliation-job approach (clever, leaky) with the single-owner redesign (boring, airtight).
where i applied it again
Production data corrections: three-phase reversible scripts instead of careful one-off UPDATEs.
explanation
A library saves weeks now and costs indefinitely, through upgrades, gaps in core behavior, and someone else's roadmap. The closer a component is to the heart of your domain, the stronger the case for owning it.
origin
Evaluating tree-selector libraries at AQTech and finding every candidate failed in core behavior.
where i applied it again
Same calculus shaped the analytics service: own the query path, rent the storage engine.
explanation
Synchronization means maintaining agreement between copies forever. Recomputation means deriving truth from source on demand. The second is slower per operation and dramatically cheaper per year.
origin
The staleness-signal design: services signal 'this may have changed' instead of shipping computed values around.
where i applied it again
Guides every caching and projection decision I make since.
explanation
A mandated standard is followed while someone watches. A standard that wins on merit, proven on real screens and cheaper to follow than to skip, survives its author leaving. Mine did.
origin
The AQTech Design System's standards remaining in use after I left the company.
where i applied it again
How I introduce every practice since: evidence first, adoption second, mandate never.
explanation
Review compresses years of someone else's judgment into comments on your actual work. Reading review feedback seriously, and later giving it seriously, is the highest-density learning loop I know.
origin
Freelance work under an experienced engineer who treated every PR as a teaching moment.
where i applied it again
As the frontend reference at AQTech, review became how I taught the Design System.
explanation
Implementation docs answer 'what is this'. Decision docs answer 'why is it this and not the alternative', the only question that matters six months later. Context, constraints, alternatives, trade-offs.
origin
Writing the OLTP vs OLAP evaluation as a document and watching it end debates before they started.
where i applied it again
This entire knowledge base is the principle applied to a career.
explanation
"Every read path subtracts X" is a placement smell, not a requirement. If a state can be applied once at the write boundary, every existing consumer becomes correct without changing, and every future consumer becomes correct without knowing the rule exists. N compensations is the same bug written N times.
origin
A platform-wide "hibernated asset" state specified as an exclusion rule for eight different read paths (counters, app payload, reports, print, adherence) when it could be applied once by removing the asset from route scope.
where i applied it again
The same question closed a per-request warehouse join: the hierarchy was being resolved at every read when the pipeline could materialize it once.
explanation
Divergence checks compare two derivations and pass when they agree, and NULL agrees with NULL. Integrity needs at least one independent completeness invariant, asserted against a different source, or a whole class of missing data stays invisible while the dashboard stays green.
origin
A pipeline consistency check that compared raw against curated using the same event field, reporting perfect agreement while 79% of rows were missing a derived column and 10% of records were being dropped entirely by SQL NULL semantics.
where i applied it again
Now a standing question on any derived column: what asserts that it is populated, and does that assertion read from somewhere else?