Old School
Jan. 28th, 2026 09:41 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
billroperWe have a lot of unit tests that are running fine in the JDeveloper environment that do *not* run in the build environment. This has been a source of a lot of frustration for my co-worker who has been trying to solve the problem so I set out to dig into it today.
The first thing I did was get him to confirm that the build wasn't doing anything acutely stupid, since there are multiple separate build steps that actually rely on each other in a loop. This is an annoying problem (and not correctable in any reasonable way), but you can check through the loop by hand and he did so and determined that my first guess was wrong. Oh, well.
Ok. If you can't use the debugger because the failures are in the Gradle build, you can always go old school. It was time to start modifying the code and putting in debugging statements to drop data to the console. And so I did.
I got the stack trace for the mysterious exception that was shutting down a large number of the failing unit tests and went to stare at the source. Eventually I realized that *someone* (and it may well have been me) had written a particularly stupid little class loader that will load one class in one environment and a different class in another environment -- say, for instance, when you are running your unit test in the build environment. And when the alternate class it is trying to load isn't in the build environment either *and* is the completely wrong class to be loading, well, you've hit the jackpot.
I quickly wrote the missing class to implement the interface, causing every method to throw a NotImplementedException. This was the right choice in this case, because the interface described how to talk to the database and none of the database code is present in this particular part of the build, so you can't get it to run the unit test. And now my Class Not Found exception turned into a NotImplementedException -- and I was already dropping the stack trace out so that I could see that I was attempting to load the access record.
Huh. Now the model that I was working with in this set of unit tests was one that I had initialized by hand and it normally works fine -- until you try loading a feature that requires an access record, which it then tries to load from the database, which is not present. I wrote a class to initialize the access record and tucked it into the model initialization routine and almost *everything* started running correctly.
Happy dance! Except there were still a few (like four) unit tests that were failing for some unknown reason. One of these was working with a tiny grid that I'd initialized in my tiny model. Other things were working correctly with the grid, so why wasn't this one?
I finally wrote a routine to dump the tiny grid to the console and discovered that none of the values that I'd tried to change in the grid were actually getting stored -- with the result that the unit test didn't find the expected values and promptly failed. I managed to find where the error codes were stored and they were all the same code, so I found all of the places that could give me that error code and stuck a debugging statement in, which eventually led me to the point of failure.
I was getting an exception when I tried to parse the input value of "160.1". The exception claimed that the character in the zero position was failing, but when I looked at the exception, I realized that I had cleverly just passed back a zero instead of looking to see which character in the string was the problem. Well, *that's* easily fixed. It's the character in the *third* position. It's the decimal point. Why do I believe that a decimal point is an illegal character? Could it have something to do with localization and assembling the set of valid characters based on the locale's decimal separator?
Yes. Yes, it could.
Gradle not only runs the unit tests in multiple threads, but it *reuses* threads. The LocaleTest switched a thread to the French locale, where the decimal separator is a comma, and didn't switch it back. It does now.
And suddenly, the rest of the unit tests started working. Well, except for one test where the correct files aren't present in the build environment. I commented that particular test back out as a problem for another day.
And then I pulled out all of the spurious debugging code that I'd put in. I kept the grid printing code around though.
You never can tell when you're going to need it. :)
Or to quote one of my songs:
"I want to be old. I want to be wise.
I want to be just a little bit smarter than all those younger guys."
Amazing what you can learn from experience.
The first thing I did was get him to confirm that the build wasn't doing anything acutely stupid, since there are multiple separate build steps that actually rely on each other in a loop. This is an annoying problem (and not correctable in any reasonable way), but you can check through the loop by hand and he did so and determined that my first guess was wrong. Oh, well.
Ok. If you can't use the debugger because the failures are in the Gradle build, you can always go old school. It was time to start modifying the code and putting in debugging statements to drop data to the console. And so I did.
I got the stack trace for the mysterious exception that was shutting down a large number of the failing unit tests and went to stare at the source. Eventually I realized that *someone* (and it may well have been me) had written a particularly stupid little class loader that will load one class in one environment and a different class in another environment -- say, for instance, when you are running your unit test in the build environment. And when the alternate class it is trying to load isn't in the build environment either *and* is the completely wrong class to be loading, well, you've hit the jackpot.
I quickly wrote the missing class to implement the interface, causing every method to throw a NotImplementedException. This was the right choice in this case, because the interface described how to talk to the database and none of the database code is present in this particular part of the build, so you can't get it to run the unit test. And now my Class Not Found exception turned into a NotImplementedException -- and I was already dropping the stack trace out so that I could see that I was attempting to load the access record.
Huh. Now the model that I was working with in this set of unit tests was one that I had initialized by hand and it normally works fine -- until you try loading a feature that requires an access record, which it then tries to load from the database, which is not present. I wrote a class to initialize the access record and tucked it into the model initialization routine and almost *everything* started running correctly.
Happy dance! Except there were still a few (like four) unit tests that were failing for some unknown reason. One of these was working with a tiny grid that I'd initialized in my tiny model. Other things were working correctly with the grid, so why wasn't this one?
I finally wrote a routine to dump the tiny grid to the console and discovered that none of the values that I'd tried to change in the grid were actually getting stored -- with the result that the unit test didn't find the expected values and promptly failed. I managed to find where the error codes were stored and they were all the same code, so I found all of the places that could give me that error code and stuck a debugging statement in, which eventually led me to the point of failure.
I was getting an exception when I tried to parse the input value of "160.1". The exception claimed that the character in the zero position was failing, but when I looked at the exception, I realized that I had cleverly just passed back a zero instead of looking to see which character in the string was the problem. Well, *that's* easily fixed. It's the character in the *third* position. It's the decimal point. Why do I believe that a decimal point is an illegal character? Could it have something to do with localization and assembling the set of valid characters based on the locale's decimal separator?
Yes. Yes, it could.
Gradle not only runs the unit tests in multiple threads, but it *reuses* threads. The LocaleTest switched a thread to the French locale, where the decimal separator is a comma, and didn't switch it back. It does now.
And suddenly, the rest of the unit tests started working. Well, except for one test where the correct files aren't present in the build environment. I commented that particular test back out as a problem for another day.
And then I pulled out all of the spurious debugging code that I'd put in. I kept the grid printing code around though.
You never can tell when you're going to need it. :)
Or to quote one of my songs:
"I want to be old. I want to be wise.
I want to be just a little bit smarter than all those younger guys."
Amazing what you can learn from experience.