Ad Platform ETL Pipelines That Survive Restatement

Ad platform data is not final when you read it. Meta restates conversions for up to 28 days after the click, Google re-attributes inside its own lookback, and a pipeline that overwrites yesterday's row each morning rewrites its own history without telling anyone. We land every extract as an immutable snapshot keyed by the date it was fetched, so the spend figure that went into a March board pack is still retrievable in September next to what that same day looks like now. The sources we actually connect: GAQL queries against Google Ads through a manager account with its own service credentials, so a staff departure does not take the pipeline with it. Meta through asynchronous insight jobs with polling and backoff, rather than synchronous calls that time out the moment an account grows. TikTok, LinkedIn and Criteo where they matter to the account. In Turkey the marketplace channels are the hard part: Trendyol and Hepsiburada ad reporting arrives as an export rather than an interface, so we write the parser and pin a schema contract that fails loudly the morning a column is renamed, instead of silently loading nulls into a spend column. Where spend stops matching the ledger: Turkish advertisers pay a 7.5 percent digital services tax and 15 percent withholding on payments to non resident platforms, so the booked cost of a campaign is never the number the platform reports. Finance and marketing then spend a year arguing about which figure is real. We carry both through the pipeline as separate columns with the rule written down, and reconcile monthly against the e-fatura record, so the argument happens once. How it runs in production: Python extractors and dbt transformations orchestrated by Dagster, on your infrastructure rather than ours. Where a maintained connector already exists we use it and say so; rewriting a working Airbyte source in order to bill for the effort is not a service. Failures retry with backoff and then land in a dead letter table alongside the request that caused them. Token expiry raises an alert before it expires, not after. Every source carries a heartbeat, because silence is the failure that costs the most and looks like nothing. What stays out of the pipeline: Attribution. A blended model baked into ingestion cannot be revised without a full backfill, so attribution belongs in the warehouse layer where changing your mind costs a query. Also out: any input that arrives as an email attachment on a permanent basis, hand maintained mapping spreadsheets that one person understands, and connectors to platforms carrying a rounding error of your spend, which cost more to keep alive than the decisions they inform. Not the right buyer: If you run two ad accounts and spend under roughly 50 thousand euro a month, a scheduled export into a spreadsheet is the honest answer, and we will say that on the first call rather than after the invoice. We also decline engagements where nobody on your side will hold the credentials or take the rota after handover, because an orphaned pipeline fails quietly and then gets blamed for the quarter. What you receive: The repository, the orchestration configuration, the schema contracts, a runbook naming what breaks and what to do about it, and the reconciliation notebook we used to prove the first monthly close. We run the new pipeline in parallel with your existing reporting for one full month before anything is switched off.