Series

Start with the failure, follow the evidence across boundaries, and do not stop at a green build when the cause is still unknown.

Identifiers, clocks, migrations, constraints, and queries all have to respect one awkward fact: the data will probably outlive the code that wrote it.

What happens after an HTTP 200, during a retry, or when half the work succeeded? The boundary is only useful once those failures have names.

Find what the old system actually does, preserve the behavior that matters, and create a seam small enough to change without betting everything on the rewrite.

Start with the change that hurts. The useful role, invariant, or dependency usually becomes obvious before its pattern name does.

Objects should make the next change less surprising through usable construction, honest roles, small collaborations, and contracts callers can rely on.

Language, work, ownership, and system relationships reveal boundaries that a directory tree cannot.

Clean diagrams eventually meet stored data, concurrent writes, and framework lifecycles. That is where mapping, reconstitution, and ownership stop being cosmetic decisions.

Commands ask for work; events record accepted facts. Queues, outboxes, and long-running processes make the difference expensive to ignore.

Choose the test boundary from the defect you need to catch, not from the coverage number you want to report.

When is a module no longer enough to own a change? Follow the deployment, coordination, scaling, and ownership costs before paying for a network boundary.

Measure the bottleneck first, because requests, queries, caches, workers, and long-lived processes do not fail in the same way.

The reason behind a change should survive the trip from commit and review to changelog, release decision, and shipped artifact.

A successful deploy command says almost nothing about compatibility, rollback, or what production is doing now.

A lockfile can repeat a choice. It cannot tell you who was allowed to change the artifact, which inputs executed, or where the credentials went.

Technical decisions still have to fit commercial reality. Risk, capacity, ownership, pushback, and the conditions for changing course belong in the decision too.

Use AI to reach evidence faster without handing it comprehension, authority, or the decision.

Backups, replicas, and runbooks matter only when they help restore the first useful service under the failures you actually expect.

A retrospective of a private logistics rating engine: the original PHP design, model changes, performance investigations and the Go port, with the code and measurements behind the decisions. Dates set the reading order, not when the work happened.