Dilemma over authenticity of gcov generated code coverage percentage where unit tests are not technically correct

Question

When I joined my company as a new comer and I was exploring the unit test suite of the product code. It is using gtest framework. But when I checked all the tests, they were testing the whole functionality by calling real functions and asserting expected output. Below is one such test case as an example:

TEST(nle_26, UriExt1)
{
    int threadid = 1;
    std::shared_ptr<LSEng> e = std::make_shared<aseng:: LSEng >(threadid, "./daemon.conf");
    std::shared_ptr<LSAttrib> attr = e->initDefaultLSAttrib();
    e->setLSAttrib( attr );
    std::shared_ptr<DBOwner> ndb = e->initDatabase(datafile,e->getLogger());
    e->loadASData(ndb);
    e->setVerbose();

    std::shared_ptr<NewMessage> m = std::make_shared<NewMessage>(e->getLogger());
    ASSERT_TRUE(m != nullptr);
    ASSERT_TRUE(e != nullptr);
    m->readFromFile("../../msgs/nle1-26-s1");
    e->scanMsg(m, &scan_callBack_26, NULL);
    std::map<std::string, std::vector<std::string>> Parts = e->verboseInfo.eventParts;
    std::vector<std::string> uris = Parts["prt.uri"];
    ASSERT_EQ(uris.size(), 2 );
    ASSERT_EQ(uris[0] , "mailto:www.us_megalotoliveclaim@hotmail.com");
    ASSERT_EQ(uris[1] , "hotmail.com");
}

I found all the tests in the unit test directory having the same pattern like:

1. Creating and initialising actual object
2. Calling actual function
3. Starting actual daemon
4. Loading actual database of size around 45MB
5. Sending actual mail for parsing to daemon by calling actual scanMsg function, etc.

So, all the tests appear more of as functional tests, rather than unit tests.

But, the critical part is, on their official intranet site, they have projected the code coverage percentage of this product as 73%, computed using gcov.

Now, code profiling tools like gcov computes coverage on the following params:

1. How often each line of code executes
2. What lines of code are actually executed
3. How much computing time each section of code uses.

As, these tests are running actual daemon, loading real database and calling actual functions to scan the message, of course, above 3 params will play some role in it, so I doubt it will be completely zero.

But my bothering questions are:

1. Black box testing also does functional testing just as this, so what's the difference between above and functional test?. In blackbox, testers unaware of the inside code, writes test cases to test the functionalities specific to requirements. How above such kind of tests are different than that? So does gcov generated coverage on this test suite, can be trusted or misleading?

2. Apparently, gcov code coverage data is based on test suite with all technically incorrect unit tests, does it mean the actual code coverage may be even zero?

3. In unit test, we mock function calls using google mock-like framework rather than calling actual calls Purpose of unit test is to test the code itself, by smallest unit wise. But above tests, seemingly more like functional tests, can gcov generate reliable code coverage data based on it??

This is haunting me for last two days. So thought to serve on the table for experts.

Awaiting wonderful insights :)

Thanks.

Answer 1

I agree with @ThomasOwens's anwer as far as you should comply to the companies test strategy.

I disagree with his answer as far as different test types serve different purposes, have different properties and consequences. IMHO it is important to know the differences to be able to implement them for maximum efficiency (with the development process in mind, not some technical performance...). You cannot do this if you do not use the terminology of test types correctly.

What you describe looks to my like a misconception: They call it unit tests because they are implemented using a unit test framework. But as you observed, this are not unit tests, but application tests (or module test at best).

Real unit tests verify the behavior of a unit in isolation. Test that require resources not directly accessible in RAM are not unit test by definition (since they depend on more then the tested code).

It is possible to write unit tests without using mocks, but that would require to implement all business logic in independent units and do the "wiring" through delegation objects that are "to simple to fail". This is hard to do consequently.

Unit Tests should implement the FIRST RTFM rules.
The acronyms resolve to:
https://dzone.com/articles/writing-your-first-unit-tests:

• Fast
• Iindependent
• Repeatable
• Self validating
• Timely

and
https://www.youtube.com/watch?v=ybDhy7b2i2o

• Readable
• Trustwothy
• Fast (yes, this is crucial... ;o))
• Maintainable

None of them applies to the test you describe.

But

That does not mean that the test your company created are bad or useless or anything like that. It just means that they are not unit tests, not more ... not less...

Answer 2

What you are describing seems like the difference between solitary and sociable unit tests. I would have some questions about loading a 45MB database and sending actual mail - the database is quite large and sending email often requires a lot of other pieces. Conceptually, though, it seems like the unit tests are sociable unit tests.

The concerns that I have are mostly around the utility of the tests. It can take a long time to load and reload a 45MB database many times during a test cycle to ensure the proper state before a test. The database and the mail server are things that I'd consider mocking out instead of using sociable tests, in order to improve performance and repeatability of the tests.

The gcov tool appears to be doing exactly what was designed to do, which is measure the lines of code executed by the test. As long as lines of code in tools not developed by the organization aren't showing up, it seems reasonable.

Depending on the context, the distinctions between unit and integration and system tests or between white box and black box tests or other categorizations may not matter at all. In an organization that tests each unit in isolation, perhaps differentiating between solitary and sociable unit tests or treating all unit tests as solitary while integration and system tests as sociable make a difference, but it may not matter to another organization.

The better approach is to look at the overall testing strategy, considering who writes tests, when and how they get executed, and if they are accomplishing their goals of improving product quality and reducing risks. If the tests are achieving their goals and the tools are providing the insights that the team needs to find gaps and close them, the specific terminology doesn't matter so much, especially since there aren't universally accepted definitions.

Answer 3

I agree with the other answers already given. But here I'll try to answer the specific questions that you had:

1. Black box testing also does functional testing just as this, so what's the difference between above and functional test?. In blackbox, testers unaware of the inside code, writes test cases to test the functionalities specific to requirements. How above such kind of tests are different than that? So does gcov generated coverage on this test suite, can be trusted or misleading?

Since a blackbox test does not concern itself with inner workings of a program, and thus should have no knowledge of the source code, code coverage should not apply to it. Of course gcov could generate coverage if you have the source code, but changing the blackbox tests based on it would mean those tests are not really blackbox any more.

2. Apparently, gcov code coverage data is based on test suite with all technically incorrect unit tests, does it mean the actual code coverage may be even zero?

If you know a test is bad, it should be fixed or removed, so that the final code coverage numbers reflect the amount of code covered by good tests. It doesn't make sense to say you've got 100% coverage if your tests are bad.

3. In unit test, we mock function calls using google mock-like framework rather than calling actual calls Purpose of unit test is to test the code itself, by smallest unit wise. But above tests, seemingly more like functional tests, can gcov generate reliable code coverage data based on it??

Yes, the code coverage data is reliable. Note that you can have 100% coverage with solitary and/or sociable unit tests, and still miss bugs. If you don't have 100% coverage, then that means you need more tests, but it doesn't tell you what kind of tests you need.