Generating Characterization Tests for Legacy Code

Slides via Jonas Follesø (@follesoe)

Huge methods (~3000+ lines)

Dav

e &

Kar

in h

ttp:

//w

ww

.flic

kr.c

om/p

hoto

s/dn

k_uk

/352

5103

502/

~50 slow integration tests

How the development team felt...

What they needed

What you should read

Fras

er S

peirs

htt

p://

ww

w.fl

ickr

.com

/pho

tos/

fras

ersp

eirs

/339

5595

360/

Legacy code is code without tests.  Code without tests is bad code. 

-Michael C. Feathers

A characterization test is test that characterizes the actual behavior of a piece of code.

It acts as a change detector, protecting legacy code from unintended changes

public double Calc(double inv, double rt, int y){ double ret = 0; for (int i = 1; i <= y; i++) { ret = inv * Math.Pow(1.0 + rt / 100.0, i); } return ret;}

[TestMethod]public void Calc_characterization(){ var calc = new CalcUtil(); double result = calc.Calc(10000, 10, 10);

Assert.AreEqual(42.0, result); }

Assert.AreEqual failed. Expected:<42>.

Actual:<25937.424601>.

[TestMethod]public void Calc_characterization(){ var calc = new CalcUtil(); double result = calc.Calc(10000, 10, 10);

Assert.AreEqual(25937.424601, result); }

Test run completed. Results 1/1 passed.

public double CalculateCompoundInterest(double investment, double interest, int

years){ double projectedValue = 0.0; for (int year = 1; year <= years; year++) { projectedValue = investment *

Math.Pow(1.0 + interest / 100.0, year); }

return projectedValue;}

Fras

er S

peirs

htt

p://

ww

w.fl

ickr

.com

/pho

tos/

fras

ersp

eirs

/339

5599

536/

…A pinch point is a natural encapsulation boundary. When you find a pinch point, you’ve found a narrow funnel for all the effects of a large piece of code…

-Michael C. Feathers

fast characterization tests

The Golden Master Approach

Before making any change to production code, do the following:

1) Create X number of random inputs, always using the same random seed, so you can always generate the same set over and over again. You will probably want LOTS of inputs for good coverage. 2) Bombard the class or system under test with these random inputs.

3) Capture the outputs for each individual random input.

A picture's worth a 1000 tests.

Unit testing asserts can be difficult to use.

Approval tests simplify this by taking a snapshot of the results, and confirming that they have not changed.