Architecture & economics

Move the decisions, not the tables

A warehouse that had fallen behind: nightly loads that finished later every quarter, a reporting server nobody dared touch, and a platform that could answer last year’s questions and none of this year’s. The move that worked was sequenced by the decisions the reports served, not by the tables, and it ran two platforms for longer than anyone wanted.

• →Behind does not mean old; it means unable to answer this year’s questions in time
• →Sequence the move by the decisions a report serves, and move its consumers with it

One number, two histories

Mergers, departmental silos, platform migrations: the surface story differs every time. The invariant is that two groups have incompatible definitions of the same metric, and the pipelines are the easy part.

• →Definition reconciliation is the project; the pipeline work is support
• →Agree on the metric before writing the model, or relitigate it forever

Rebuilding ingestion people trust

Pipelines decay quietly and the rebuild question is never "is this broken" but "can we still justify trusting it." Choosing the integration approach, handling sources that do not agree with each other, and proving the result before anyone is asked to rely on it.

• →Pick the integration approach from the source constraints, not from preference
• →Parallel run until every material variance is explained, not merely small

What the platform costs

Warehouse spend is the one number a data leader owns end to end and the one most cannot explain. Making cost legible, pushing the decision to the people creating it, and knowing which inefficiencies are worth leaving alone.

• →Unattributed cost cannot be optimized: attribution comes before efficiency
• →Cheap and slow is a decision, not a default

What a faster refresh actually buys

A request for real-time data is a request for a new operating promise. How to find out whether the earlier answer changes an action, what that promise costs, and when a slower schedule is the right one.

The tools I use

The current working set, kept alongside the 2025 list rather than edited over it. The gap between the two is the interesting part: the core barely moved, a compliance tool became a runtime dependency, and part of the list stopped being data tooling at all.