Bogus and AutoBogus

GenAuto.Of<T>() builds test data by itself: plausible values mixed with edge cases, relations between fields, reproducible by seed and shrinkable. Where part of a value has to look real (an address a geocoder accepts, a valid IBAN, a product code in your format), Bogus is the better source for that part. You don't have to choose: hand Bogus the parts it is good at, and GenAuto keeps building the rest.

Either way every Bogus draw goes through CsCheck's PCG, so a failing seed replays the same value, and the value still shrinks.

AutoOptions.Use<T>(gen) replaces the generator for every T inside the value with any Gen<T>. Give it a Bogus Faker<T>:

using CsCheck;
using Nullean.Entities.CsCheck;
using Nullean.Entities.CsCheck.AutoBogus;

record Address(string Street, string City, string PostCode);
record BankDetails(string Iban, string Bic);
record Customer(string Name, int LoyaltyPoints, Address Address, BankDetails Bank);

Gen<Customer> customers = GenAuto.Of<Customer>(configure: o => o
    .Use(GenFakerExtensions.Faker<Address>(f => f.CustomInstantiator(x =>
        new Address(x.Address.StreetAddress(), x.Address.City(), x.Address.ZipCode()))))
    .Use(GenAutoBogus.Of<BankDetails>()));

1. only for GenAutoBogus

Every Customer now has a real-looking address and bank details, and its Name and LoyaltyPoints are still GenAuto's: blank and hostile names, negative and extreme points, the edge cases that find bugs. Bogus alone would never produce them; GenAuto alone would produce an address no validator accepts.

Use it the same way for a rule that GenAuto cannot satisfy by luck. If a constraint needs a realistic email, Probe or AutoTest reports that it produced no valid value and names the rejecting constraints; o.Use(GenFakerExtensions.Faker<EmailAddress>(...)) fixes that for that type only.

The example is tested in tests/Nullean.Entities.CsCheck.AutoBogus.Tests/RicherGenerationTests.cs: addresses always come from Bogus, names still include blank values, and a seed replays.

dotnet add package Nullean.Entities.CsCheck.AutoBogus

using Nullean.Entities.CsCheck.AutoBogus;

Gen<Cart> carts = GenAutoBogus.Of<Cart>(faker =>
{
	// your AutoFaker configuration: binders, overrides, skip rules
});

rules.ForAll(c => c.Owner.Length > 0, carts, iter: 100);

The Gen<T> plugs into anything that takes one: gen: on ForAll and Probe, start: on Preserves, or o.Use<T>(...) as above. The package brings Nullean.Entities.CsCheck and Soenneker.Utils.AutoBogus (a maintained fork of AutoBogus), uses reflection, and is for test projects.

Plausible-only data cannot break most rules: Probe reports them as never violated and the axes as unfalsifiable. That is why the AutoBogus generator is not the default, and why mixing Bogus in per type, as above, usually beats using it for the whole value. The measurements are in docs/autobogus-decision.md.

• Bogus's global lock. Bogus takes a process-wide lock on every draw (Randomizer.Locker, which cannot be turned off). A generator that draws mostly from Bogus stops scaling at about two threads; one that uses Bogus for one small part barely notices.
• configure runs once per thread, on first use, not once per sample: keep it pure.
• IndexFaker and IndexVariable are reset to zero before every draw, so a rule reading them is reproducible but not a useful sequence. f.IndexGlobal, f.UniqueIndex and Faker.GlobalUniqueIndex are process-wide statics shared with every other Faker: avoid them.
• The Bogus dataset version is part of the seed. A Bogus update can change what a seed produces for the parts Bogus draws.

More on what goes through the PCG: Generated data.