Google BigQuery

The diagnostic starts with a question most teams haven't asked out loud: is BigQuery reporting on the business, or running part of it? Retailers loading order, stock and fulfilment data into BigQuery for dashboards carry a different risk profile to ones feeding pricing, segmentation or allocation logic back out of it. We map every object landing in BigQuery against its source system of record before touching a pipeline, because the moment someone builds a decision on stale replicated stock data, that decision carries a cost nobody budgeted for. Where the diagnostic finds genuine source-of-truth confusion, the recommendation is a scoped engagement to fix ownership first and pipeline cadence second.

Ownership gets blurry fast with a warehouse nobody logs into except analysts. BigQuery should hold a copy of stock, order and financial data, not the working copy fulfilment or finance act on in the moment. The failure mode is specific: a warehouse team or finance analyst starts treating a BigQuery table as current because the dashboard refreshes hourly, then stock allocation or a payout reconciliation runs a cycle behind the ERP or ecommerce platform it was built from. Finance loses trust when a month-end report built on BigQuery disagrees with the general ledger by a figure nobody can trace to a single failed load.

The fix is a written ownership line, not a smarter query: ERP owns stock and financial postings, the ecommerce platform owns order state and customer data, BigQuery owns the combined view for reporting. Nothing downstream should write back into an operational system based on a BigQuery table without a logged, explicit handoff.

Point-to-point pipelines into BigQuery age the way any point-to-point integration does: each new source system, a marketplace, a 3PL, a second ERP after an acquisition, means another custom loader built and maintained by whoever had time that quarter. Patchworks, as the integration layer running the connections rather than a bystander forwarding payloads, validates each load, retries the ones that fail, and logs what happened well enough that a failed row count is visible before it reaches a BigQuery table. That is what the integration layer is for: governance over the data, not a passive pipe between systems.

The practical difference shows up in onboarding speed. Adding a new marketplace feed through a properly built connection is a task measured in days; building and testing a bespoke loader from scratch is a project. Cogent² runs this as expert operators using AI-assisted monitoring to flag drift early, not as an AI-first build process standing in for the judgement an experienced integration team brings.

Monitoring BigQuery well means watching load latency and row counts against source volumes, not just whether a job finished. A scheduled load can succeed and still carry half the expected rows, because the API it pulled from paginated differently that day; the job logs green, the table is wrong. Exceptions need their own queue, failed loads, schema drift from a new feed, duplicate rows from a retried batch, each with an owner and a response time, not a dashboard alert nobody is rostered to act on.

The chain, left unmanaged, runs: stock drift, stale dashboard, finance query, manual trace through load logs to find which batch introduced the gap. Caught at the pipeline, this is a short delay. Caught at month-end close, it's a week of someone's time.

Enablement here stays narrow: who owns the field-level mapping document, who gets paged when a load fails, who signs off a schema change before a new feed goes live. A short runbook covering the three or four pipelines that matter most, the ERP financial extract, the ecommerce order feed, any reverse sync into pricing or segmentation, beats a wiki nobody updates. Escalation needs a named owner on both the analytics side and the system feeding it, because an alert with no clear owner sits in a channel until someone notices the dashboard looks wrong.

Hypercare after go-live focuses on load volume under real traffic, not test data: peak order spikes, marketplace sale events, the days when every source system writes faster than usual. That's when a load job sized for normal volume starts timing out or truncating, and nobody notices until a report looks thin. Cogent² holds escalation through the first peak cycles, watching load latency and row counts against known order volumes rather than waiting for finance or merchandising to flag a number that looks off. Beyond hypercare, managed support means someone owns the pipelines when a source API changes shape, a new marketplace connector lands, or a schema update breaks a downstream table silently rather than loudly.