Aoife Donnelly

I work backwards from the failure, because both of these tasks fail in ways the happy path does not predict. On test data, the requirement that sounds simplest is generating data matching a schema, and it is the one with the most edge cases. Foreign keys have to be satisfied, so rows cannot be generated independently. Unique constraints collide before the table fills. Check constraints encode rules that have to be read and respected. And a value that satisfies the schema can still break a parser downstream, which is the failure that appears far from the cause. Realistic data and valid data are different goals. A column declared as a string can hold a number, a date in an unexpected format, an empty string and a very long value, and those are the four cases that break exports. Generating only well-formed values tests nothing. The volume question has an answer that surprises people: a small number of rows repeated frequently exposes join and aggregation bugs better than a large table does. On queries, the recurring cost is the one that runs once per row rather than once per statement. A query built inside a loop over results turns a page fetch into hundreds of round trips, and the fix is nearly always to fetch the related rows in one statement and join in memory. Generated data has one more property worth wanting, and it is reproducibility: a seed, so a failure can be reproduced exactly rather than approximated. I finish each one with the failure it was written to catch, since a generator without a stated target is a generator whose output nobody has checked.