Which types of test data matter?
Choose according to the test. Valid data checks an accepted path; invalid data checks rejection; boundary data checks transitions. For a field accepting amounts from 1 to 100, try 0, 1, 100 and 101, then missing and malformed values. This classification describes the purpose. Synthetic generation, masked source records and curated fixtures describe how the data is obtained; the classifications are not interchangeable.
Why can production-like data miss a bug?
Common records may never exercise a rare failure. A payment test needs explicit states: approved, declined, already reversed and reversed twice. Define expected balances and event counts before creating records. A second reversal must not create a second refund if the business contract requires idempotency. Referential integrity checks only whether relationships are valid; it does not establish that the business outcome is correct.
How much data do you need?
Use the smallest dataset that proves the behavior for functional tests. For performance tests, also define volume, value distributions, concurrency and query patterns. A million identical rows can give misleading results if production has skewed keys or varied document sizes. Record the setup and cleanup procedure, then choose a library, generator or provisioning platform based on those requirements.
Illustrative scenarios and editorial acceptance criteria; not measured product benchmarks.