Data Generators and Fakers

Often, you don’t want to generate totally random data; it suffices that some aspects of it are random. This naturally raises the question: Where can one get non-random, natural data from, and how can one integrate this into Fandango?

Augmenting Grammars with Data

The straightforward solution would be to simply extend our grammar with more natural data. In order to obtain more natural first and last names in our ongoing names/age example, for instance, we could simply extend the persons.fan rule

<first_name> ::= <name>

to

These are the given names on Pablo Picasso’s birth certificate.

<first_name> ::= <name> | "Alice" | "Bob" | "Eve" | "Pablo Diego José Francisco de Paula Juan Nepomuceno Cipriano de la Santísima Trinidad"

and extend the rule

<last_name> ::= <name>

to, say,

<last_name> ::= <name> | "Doe" | "Smith" | "Ruiz Picasso"

then we can have Fandango create names such as

Pablo Diego José Francisco de Paula Juan Nepomuceno Cipriano de la Santísima Trinidad Smith,63706191375955899769

Eve Doe,650508058
Eve Smith,0186517241640449
Alice Ebhiepegnkjvc,0
Alice Smith,0762530232503
Pablo Diego José Francisco de Paula Juan Nepomuceno Cipriano de la Santísima Trinidad Smith,08642495378759936
Eve Smith,9
Pablo Diego José Francisco de Paula Juan Nepomuceno Cipriano de la Santísima Trinidad Smith,7275986301386
Alice Smith,37
Bob Doe,896931412

Note that we still get a few “random” names; this comes as specified by our rules. By default, Fandango picks each alternative with equal likelihood, so there is a 20% chance for the first name and a 25% chance for the last name to be completely random.

Note

Future Fandango versions will have means to control these likelihoods.

Using Fakers

Frequently, there already are data sources available that you’d like to reuse – and converting each of their elements into a grammar alternative is inconvenient. That is why Fandango allows you to specify a data source as part of the grammar - as a Python function that supplies the respective value. Let us illustrate this with an example.

The Python faker module is a great source of “natural” data, providing “fake” data for names, addresses, credit card numbers, and more.

Here’s an example of how to use it:

from faker import Faker
fake = Faker()
for i in range(10):
  print(fake.first_name())

James
Joseph
Jonathan
Donald
Joanna
Jennifer
Mercedes
Donald
Pamela
Noah

Have a look at the faker documentation to see all the fake data it can produce. The methods first_name() and last_name() are what we need. The idea is to extend the <first_name> and <last_name> rules such that they can draw on the faker functions. To do so, in Fandango, you can simply extend the grammar as follows:

A generator applies to all alternatives, not just the last one.

<first_name> ::= <name> := fake.first_name()

:= is the assignment operator in several programming languages; in Python, it can be used to assign values within expressions.

The generator := EXPR assigns the value produced by the expression EXPR (in our case, fake.first_name()) to the symbol on the left-hand side of the rule (in our case, <first_name>).

Important

Whatever value the generator returns, it must be parseable by at least one of the alternatives in the rule. Our example works because <first_name> matches the format of fake.first_name().

Tip

If your generator returns a string, a “match-all” rule such as

<generated_string> ::= <char>* := generator()

will fit all possible string values returned by generator().

We can do the same for the last name, too; and then this is the full Fandango spec persons-faker.fan:

from faker import Faker
fake = Faker()

include('persons.fan')

<first_name> ::= <name> := fake.first_name()
<last_name> ::= <name> := fake.last_name()

Note

The Fandango include() function includes the Fandango definitions of the given file. This way, we need not repeat the definitions from persons.fan and only focus on the differences.

Note

Python code (from Python files) that you use in a generator (or in a constraint, for that matter) needs to be imported. Use the Python import features to do that.

Important

include(FILE) is for Fandango files, import MODULE is for Python modules.

This is what the output of the above spec looks like:

Marcus Robinson,150253358580
Jill Silva,7820908425213

Jonathan Castaneda,785662435526

Monica Allison,769215083322816962

Christopher Davis,9561429
Ralph Griffin,5468079424361740

Randall Sherman,4912365575693293

Beth Duarte,2738717231668326995
Alisha Lee,907573

Jesus James,25630364

You see that all first and last names now stem from the Faker library.

Number Generators

In the above output, the “age” fields are still very random, though. With generators, we can achieve much more natural distributions.

After importing the Python random module:

import random

we can make use of dozens of random number functions to use as generators. For instance, random.randint(A, B) return a random integer \(n\) such that \(A \le n \le B\) holds. To obtain a range of ages between 25 and 35, we can thus write:

<age> ::= <digit>+ := str(random.randint(25, 35));

Important

All Fandango generators must return strings or byte strings.

• Use str(N) to convert a number N into a string

• Use bytes([N]) to convert numbers N into bytes.

The resulting Fandango spec file produces the desired range of ages:

David Pace,29

Daniel Williams,30

Tracy Marshall,25
Donald Kemp,27

Phillip Hernandez,30

Daniel Clark,33
David Shah,26

Jeffrey Gonzalez,28
Russell Carter,30

Chelsea Johnson,28

We can also create a Gaussian (normal) distribution this way:

<age> ::= <digit>+ := str(int(random.gauss(35)));

random.gauss() returns floating point numbers. However, the final value must fit the given symbol rules (in our case, <digit>+), so we convert the age into an integer (int()).

These are the ages we get this way:

William Griffin,35
Eddie Bell,33

Jennifer Mills,35
Charlene Moran,35

Edward Ford,35

Abigail Herrera,34
Katie Lopez,35

Heather Wright,36

John Pena,37
Stephanie Bryant,33

In Statistical Distributions, we will introduce more ways to obtain specific distributions.

Generators and Random Productions

In testing, you want to have a good balance between common and uncommon inputs:

• Common inputs are important because they represent the typical usage, and you don’t want your program to fail there;

• Uncommon inputs are important because they uncover bugs you may not find during alpha or beta testing, and thus avoid latent bugs (and vulnerabilities!) slipping into production.

We can easily achieve such a mix by adding rules such as

<first_name> ::= <name> | <natural_name>
<natural_name> ::= <name> := fake.first_name()

With this, both random names (<name>) and natural names (<natural_name>) will have a chance of 50% to be produced:

James Carter,61

Eti Owens,3975501167620815
Kathy Coleman,9414524

Evphoesn Mata,018136
Epnjjlhwvr Lopez,03613096922

Imqriewvlwtqhwormy Mcdonald,474857368599
Kevin Willis,93914270682620150722

Carly Schaefer,263221133105

Brenda Hall,759283
Barbara Olson,5390403234030

Combining Generators and Constraints

When using a generator, why does one still have to specify the format of the data, say <name>? This is so for two reasons:

To allow access to the elements of the generator output, Fandango parses the output according to the symbol rules.

1. It allows the Fandango spec to be used for parsing existing data, and consequently, mutating it;

2. It allows additional constraints to be applied on the generator result and its elements.

In our example, the latter can be used to further narrow down the set of names. If we want all last names to start with an S, for instance, we can invoke Fandango as

Grammar symbols <...> support all Python string methods.

$ fandango fuzz -f persons-faker.fan -c '<last_name>.startswith("S")' -n 10

and we get

Elizabeth Santana,945864365894594947

Veronica Stewart,92

Amy Scott,5070659

Jamie Shaffer,44484303244

Andres Shea,37566

Benjamin Steele,97251

Hunter Sanders,208323062
Jeff Smith,13714825266815552

Michael Sanchez,8269

Donald Smith,87364238910081992825

When to use Generators, and when Constraints

One might assume that instead of a generator, one could also use a constraint to achieve the same effect. So, couldn’t one simply add a constraint that says

<first_name> == fake.first_name()

In case this should work, this is only through some internal Fandango optimization that directly assigns computed values to symbols.

Unfortunately, this does not work.

The reason is that the faker returns a different value every time it is invoked, making it hard for Fandango to solve the constraint. Remember that Fandango solves constraints by applying mutations to a population, getting closer to the target with each iteration. If the target keeps on changing, the algorithm will lose guidance and will not progress towards the solution.

Likewise, in contrast to our example in Combining Generators and Constraints, one may think about using a constraint to set a limit to a number, say:

$ fandango fuzz -f persons-faker.fan -c '<last_name>.startswith("S")' -c 'int(<age>) >= 25 and int(<age>) <= 35' -n 10

This would work:

Kayla Sparks,31

Janet Salas,31

Kayla Salas,31

Kayla Sparks,35
Kayla Scott,31

David Sanchez,31
Susan Sparks,31

Janet Sparks,31

John Smith,31

Kayla Salas,35

But while the values will fit the constraint, they will not be randomly distributed. This is because Fandango treats and generates them as strings (= sequences of digits), ignoring thur semantics as numerical values. To obtain well-distributed numbers from the beginning, use a generator.

1. If a value to be produced is random, it should be added via a generator.

2. If a value to be produced is constant, it can go into a generator or a constraint.

3. If a value to be produced must be part of a valid input, it should go into a constraint. (Constraints are checked during parsing and production.)