Beware of data generators¶
In time of big data and machine learning when we see the difficulties to find proper data for test automation we’re tempted to turn to the silver bullet “RPA” (Robotic Process Automation). Solutions grow like mushrooms, all with the intention to reduce effort for test data creation and test case maintenance.
We also use test data generators in
baangt, very well dosed, very carefully and never to assess, whether the AUT
(Application Under Test) is suitable for production. We use test data generators based on users data recordings to harden
the application in early stages. We’ll “play” around (automatically), perform random clicks in random orders on buttons,
extend each field to it’s maximum length, fill traditional chinese characters into arabic text fields and similar tests.
That’s a valuable excercise. It helps to harden the application before the masses of end-users will try to beat it down. But we use those tests not in daily or weekly regression tests or when we assess the latest build or increment.
What do we use instead of random generated data?¶
Beautifully, precious, manually designed and recorded test data from the business department (preferably key-users or even power end-users). We extend the dataset whenever a new issue arises in the production if in the post-mortem phase of the issue it turns out, that regression testing could have identified the issue with proper data.
As this test set grows in lines, it grows in value. After a few weeks you’ve covered tests for all your major business logic within the test set. Depending on your product cycles it might mean a hell of maintenance effort (e.g. if your pricing changes every hour or day, you’ll either subclass the assertion methods with your own logic or you’ll be miserable updating 1000s of comparison values in the test data), but this effort might very well lead to keeping customers (as opposed to losing them to competition, if your software is unstable), supporting your whole organization to deliver the highest possible performance (by having reliable software and processes in place).
That’s all nice, but it won’t work for us because …¶
The more you depend on software and the more complex your software and processes are, the less I believe, that it is not possible in your case to run your regression tests with meaningful and reliable test data. Here’s a list of rather lame excuses. If you use any of them, please think again.
||The software and processes are too complex. We’ve trillions of combinations. It would take months to test all that and in the end we anyway wouldn’t know, whether the results are right or wrong.||System landscapes - especially grown ones - can be really complex. But at least try to find a good starting point by combining a few, beautifully crafted test cases and validation rules with ML Data expanders. It’s better than nothing and as your system stability improves and you need less time to fight fires in production, you’ll gain time to invest in improving your test data. You can start an uplifting spiral!|
||We don’t have good quality data in our stages.||
Often you’ll need basic or master data before you can test variations of transactional data. In a production company you’ll need material master data of raw materials, customer master data, vendor master data and much more before you can test your shiny new plant optimization system. But there are more benefits in investing time to create all those prerequisit master data automated.
If the mountain of work is too big to tackle it all at once, slice it down to reasonable chunks. Identify, which data set has the highest value in terms of data quality/reliability of test case output and start with these. A classical example are customer master data. All your test cases are with “John Doe”. First real customer record without a ZIP-Code or strange address format for a certain country, phone number format, etc. breaks your tests. Don’t be like that. Or at least don’t be surprised, if it happens to you. Having a few hundred different customer master records, that you use in your test data base and recreate every now and then doesn’t take much of your time but provides a great deal of increased reliability of your test activities.
||Our internal and external customers are in need for speed. We can hardly finish the features we need to provide on time, let alone test them in E2E-Scenarios. No need for additional overhead. Our developers know, what they do!||Good example of “famous last words”? Regression tests don’t need to run forever. Usually one would have approx.
10-15% in pure (and slow) UI-Tests, 40% of (faster) API-Tests and rely for the rest on earlier test stages (like unit tests).
The less tests you execute, the higher the need for spotless and realistic test data! Yes, of course you can (and should)
combine various test parameters into one E2E-Testcase (e.g. a specific/problematic country code with uncommon
payment terms and a reopened order (instead of a freshly created one)). Better define a reasonable time frame for
regression tests (given the fact, that they should run automatically, maybe a nightly build gives you 2-3 hours?).
Then see, which systems take how long for each test case combination. Based on the time frame, the approximate number
of testcases and the throughput you’ll understand, if and how much you need to run in parallel. With
||Our changes are frequent and fast. We tried to use test automation, but we ended up spending more time with the automation then in development and in the end the reliability of the results was not worth the effort we invested.||Understood. It is in fact true, that UI-Tests tend to need higher maintenance efforts than API-Tests. In
|‘Difficult to obtain data’||We need real data, but it’s hard to come by. We’ve validation on Phone-Numbers, IBAN, BIC and many other fields, so we need to enter real data in test cases. But we don’t have it and we don’t want to use data from production!||
At least you’re not using your customers data for test - that’s great! Indeed for some data it’s pretty difficult to obtain valid test data - most of the time based on the reasoning, that if the mechanism to create valid data was made public it would support criminal activities.
If you encounter such a need you’re in bad luck and your rant is granted. I assume, that your developer and API-Tests mock this interface/data anyway, so you did the maximum possible.
In other cases you can use existing functionality and libraries (e.g.
Build realistic synthetic data for your regression tests and take good care of it. Also, for every bug found in production enhance your test set!