Choose the smallest real source#
Not sample data — real, but small. Sample data proves the connection works and nothing about whether it works on your material, which is where the surprises live: an unusual character set, a column that is empty half the time, a date written the American way.
Connect it#
- Create the credential in the source system, scoped as narrowly as it will go. Read-only, if reading is all that is needed.
- Store it where your team already keeps secrets. A token pasted into a chat is a token you will be rotating later.
- Name the connection after what it is, including the environment. A connection called "test" outlives every test.
- Run the import once, by hand, before scheduling anything.
{
"name": "orders-staging",
"source": "https://data.example.com/exports/orders",
"auth": { "type": "bearer" },
"schedule": null
}Fill this in: replace the example above with the real shape docs expects, and say which fields are required.
Check what actually came through#
Do not trust a green tick. Check three things by hand.
| Check | How | What a failure looks like |
|---|---|---|
| Count | Compare the number of records against the source | Off by a round number, usually a page limit |
| Edges | Open the oldest and the newest record | Timestamps shifted by a fixed number of hours: a time zone |
| Ugly rows | Find a record with accents, emoji or an empty field | Question marks, truncation, or a silent skip |
Then widen it#
Once the small source is right, add the rest one at a time, repeating the same three checks after each. Connecting six at once means debugging six at once.
Next steps#
- Your first five minutes — if you skipped ahead.
- Core concepts — what the product calls each moving part.
- Troubleshooting — when an import refuses.