Snowflake

The diagnostic starts upstream of Snowflake, not inside it. We trace every feed landing there: Shopify order exports, NetSuite GL postings, WMS stock movements, carrier scans, marketplace settlement files. Most of these were built one at a time, by different teams, on different cadences. A warehouse team builds a nightly batch, finance wants hourly, and nobody checks whether the field names mean the same thing twice.

The pattern repeats across retailers: dashboards that look confident and are quietly wrong. Stock counts a day stale during a promotion. Revenue figures that depend on which refund batch ran first. We map each feed back to its source system, flag where the mapping was guessed rather than agreed, and price the fix before anyone touches a pipeline.

Snowflake is a mirror, not a decision maker. The moment a warehouse team starts trusting a Snowflake view instead of the WMS for pick availability, ownership has quietly moved and nobody signed off on it.

Before any pipeline work, we fix three questions: which system owns stock level, which owns order status, which owns the customer record. Snowflake ingests all three but is allowed to originate none of them. Get this wrong and finance ends up reconciling against a number that was already stale by the time the warehouse team saw it.

The second decision is cadence. Batch once a day is fine for trend reporting and wrong for stock-out alerts. Mixing cadences inside one dashboard, hourly stock beside daily orders, is how a buyer reorders against a figure that was never live.

Most of the feeds into Snowflake start life as point-to-point builds: a developer script pulling NetSuite postings, a scheduled export from Shopify, a CSV someone uploads every morning because nobody built the connector properly the first time. It works until the person who wrote the script leaves, or a second marketplace needs the same feed and gets built again from scratch.

Patchworks replaces that with a connector model built for this: Snowflake authenticates through OAuth 2 authorisation code, field mappings are declared once per object, and adding a new source, a new carrier, a new marketplace, becomes a configuration change rather than a new script. That is what makes onboarding a second warehouse or a new sales channel a matter of weeks rather than months.

Cogent² runs this as the operator, not the vendor. We design the mapping, set retry and polling behaviour against Snowflake's async statement API, and build monitoring that uses AI to spot a schema drift or a stalled load early, with an engineer, not a model, deciding what happens next.

Monitoring here is less about pipeline health and more about staleness nobody flagged. A load can succeed, return a 200, and still be wrong: row counts down a third because an upstream export ran against a filtered view, or a schema change on the WMS side silently drops a column Snowflake expected.

We watch row-count deltas against the prior run, schema drift on incoming tables, and gaps between expected and actual load windows. The exception queue catches a feed that completed but delivered too little, which a basic pass or fail alert never will. The chain is short: stale feed, confident dashboard, wrong reorder.

The cost often lands with finance, a close built on a table that quietly lost three days of rows, found only once it is already in a board pack.

Enablement here is narrow but non-negotiable: one person owns schema change requests, one owns the exception queue when a load fails silently, and both are named in the runbook, not implied. SOPs cover what to check before trusting a dashboard during peak: load timestamp, row count against the prior cycle, and which feeds are still batch. Skip this and the team re-learns the same lesson every peak, usually from a buyer who already reordered against a stale count.

Hypercare after go-live is about watching the feeds under real volume, not the warehouse itself. Peak multiplies both event volume and the cost of a missed load: a settlement file that lands late during a marketplace sale event means finance closes against partial revenue and has to restate.

We stay on the exception queue through the first peak cycle, tuning polling intervals and retry windows as volume grows rather than leaving thresholds set for a quiet November. After that, escalation runs through a named owner on our side and a named owner on the client's, not a shared inbox nobody checks on a Sunday.