Journal · warehouse
Warehouse-first versus tool-first instrumentation
Teams arrive at Socket Nodehub already married to a tool. Amplitude, Mixpanel, a CDP, a home-grown tracker. The marriage is not the problem. The problem is inventing the event contract inside the tool’s UI, then discovering the warehouse cannot reproduce the same metric. In-App Event Tracking Analytics should outlive the subscription.
Tool-first: speed, then regret
Tool-first instrumentation is fast. A PM can add hero_impression_v3 before lunch. Types are loose. Identity is whatever the SDK defaulted to. Six months later, finance asks for a number that matches the ledger, and the tool’s unique-user definition disagrees with the warehouse by a margin nobody can explain in a board meeting.
We do not ban tools. We ask that the tool be a consumer of a written contract: object, action, fire condition, types, null policy. If the tool cannot accept a type, the contract still exists; the mapping is explicit. When the vendor changes, the dictionary does not have to be rewritten from screenshots.
Warehouse-first: slower names, honest types
Warehouse-first means timestamps are timestamps, money is a decimal with currency, and late events have a documented rule. dbt models read state transitions on an Order object instead of inferring revenue from a client property the app could spoof. This is week nine of Instrumentation Studio, taught with BigQuery-shaped examples. Redshift and Snowflake students adapt. That limitation is stated on the syllabus on purpose: we teach the contract, not every dialect.
A practical split
Use the tool for exploratory product questions that tolerate approximate unique users. Use the warehouse for anything that must match finance, inventory, or a regulator. Both can read the same events if the payload was designed once. If you only instrument inside the tool, the warehouse will always be a second-class reconstruction.
House Accounts often start here: four apps, four tool projects, one finance team. The shared object list is the intervention. Vendor screens come second. On-site days in Bangkok, when they happen, are mostly whiteboards, not SDK install theatre.
Duplicates, late arrivals, and retries
Mobile clients retry. If your only dedupe is in the tool’s session logic, the warehouse will double-count when you finally dump raw logs. Include event_id. Prefer server confirmation for money-adjacent actions. Treat client revenue properties as untrusted. These rules feel austere. They are how a dictionary survives a vendor change without a year of archaeology.
If you want the warehouse desk without the full Studio, ask about Warehouse-Ready Event Design. If you want the whole critique culture, take the flagship cohort. Either way, write the contract before you add another impression event.