Upbuild
Upbuild / AI Data Exports
AI data exports in, BigQuery and Sheets out
An export takes a data stream, runs it for every account you point it at, on a clock you set, and writes the rows into BigQuery or Google Sheets. You change it by sentence, and every change is a version you can restore.
Every platform on one subscription
The anatomy
One export is one data stream, a list of accounts, one or more pulls and one destination. The stream says what to ask for. The accounts say who to ask it of. Each pull carries its own time, its own date window and its own write mode, so a daily pull can merge the last seven days into the same table a monthly pull rewrites in full.
Every day at 06:00, every Monday, day 3 of the month, every N days, or only when you press Run now. Each pull writes by merge, append or replace, whichever suits that rhythm.
Pick them by hand, or write a rule: names starting with, containing or matching a pattern, minus the ones you exclude. The rule is checked nightly and before every run.
It pulls a data stream for as many accounts as you choose, and delivers the rows to the destination you choose.
Change it
Every export has a panel called "Ask about this export", and its opening line is the whole contract: it knows the stream, the accounts, the pulls, the destination and the mails, and every change it makes is a version you can restore. Say "pull the last 30 days on the monthly run" and the monthly pull moves. No YAML file, no ticket.
Pause the clock, resume it, run one pull now, retry only the accounts that did not land, restore an older version, duplicate the export, or share it with a colleague as viewer or editor.
Every save is a numbered version with its author and a note, sitting in a picker beside the tabs. Choosing an older one does not change the export until you press Restore.
Ask me to change this schedule. I know its stream, its accounts, its pulls, its destination and its mails, and every change I make is a version you can restore.
Extract, transform, load
Extract: Upbuild walks each platform's API page by page, per account, resuming from the last page it read. Transform: rows are flattened, typed and given four columns of Upbuild's own. Load: the rows go into a staging table, then one MERGE, INSERT or REPLACE statement, then the staging table is dropped. That is a marketing data pipeline, and the only thing you wrote was a sentence.
account_id, account_name, connection and fetched_at, so forty accounts can share one table and you can always tell whose row you are looking at. run_id and version are optional extras.
Every number lands as FLOAT64, because a metric that was 12 yesterday and 12.5 today should not break a column. Dates and timestamps are recognised, nested values travel as JSON text.
Drawn from the rows the stream's own test returned. The green columns are Upbuild's, so many accounts can share one table.
| Column | Type | Source |
|---|---|---|
| account_idmerge key | STRING | UPBUILD |
| account_nameas the platform names it | STRING | UPBUILD |
| fetched_atwhen this pull ran | TIMESTAMP | UPBUILD |
| date_startthe day the row covers | DATE | STREAM |
| campaign_idmerge key | STRING | STREAM |
| cost_microswhole today, fractional tomorrow | FLOAT64 | STREAM |
Setting one up
The new export wizard asks for the same five things every time: Data stream, Accounts, When, Destination, Outcome. Or skip it entirely and ask the panel, which fills the same five in, one question at a time, saving each answer as you give it.
Choose a stream that already exists, built for a dashboard or saved from an answer in chat. Then pick the accounts under your managers, or write the rule that finds them. You only ever see the accounts your own login can reach.
Add a pull for each rhythm you need, each with its time, its dates and its write mode. Point it at a BigQuery project, dataset and table, or at a Google Sheets tab. Name the key columns if it merges.
The last step shows the columns that will land before a single row moves. Anything unfinished stays a draft, and a draft has no clock, so a half-built export never writes to your table by accident.
Straight answers
Written for the people who will own the table afterwards.
Failure is per account, not per run. An account the platform brakes goes to waiting and is asked again a quarter of an hour later, up to eight times: two hours of patience before anything is called a failure. A write fault is retried five times. The run finishes done, partial or failed, the mail names the accounts that did not land in the platform's own words, and Retry runs only those. A run that fails outright writes nothing.
One export runs its stream for every account you point it at, spread over 30 minutes with never more than five calls in flight, so it does not exhaust a platform's hourly allowance and starve everything else. The work runs in twelve-minute ticks and resumes from the last page read, so nothing depends on a request staying open. One live organisation reaches 859 accounts and runs 101 self-refreshing dashboards on the same connections.
Rows are extracted from every platform's API page by page, flattened, typed and stamped with account and fetch columns, then loaded through a staging table and one MERGE, INSERT or REPLACE statement, in a single transaction where it matters. Merging on keys means a restated day heals instead of duplicating. What it is not: a modelling layer. Upbuild lands typed source tables and leaves the modelling to your own SQL.
Three ways, all honest. The schedule is configured in English and versioned, so changing a window is a sentence rather than a redeploy. Every platform sits on one subscription instead of a line per source. And the same streams already feed dashboards and chat, so the warehouse pipeline and the reporting on top of it are not two purchases. If all you need is raw Google Ads to BigQuery, a plain connector is a reasonable choice.
Your project, your rules
The destination signs in one of three ways: with the Google login of the person who connected it, with a service account key you paste, or with Upbuild's writer identity, granted on the one dataset you choose. Upbuild creates the table if it is missing, never creates a dataset, and never deletes a table it did not make.
Get in touch