I like this anecdote because it points out the difference between being able to handle and process bizarre conditions, as if they were something that should work, which is maybe not that helpful, vs. detecting them and doing something reasonable, like failiing with a "limit exceeded" message. A silent, insidious failure down the line because a limit was exceeded is never good. If "fuzz testing" helps exercise limits and identifies places where software hasn't realized it has exceeded its limits, has run off the end of a table, etc., that seems like a good thing to me.

On Tue, May 21, 2024 at 12:09 AM Serissa <stewart@serissa.com> wrote:

Well this is obviously a hot button topic.  AFAIK I was nearby when fuzz-testing for software was invented. I was the main advocate for hiring Andy Payne into the Digital Cambridge Research Lab.  One of his little projects was a thing that generated random but correct C programs and fed them to different compilers or compilers with different switches to see if they crashed or generated incorrect results.  Overnight, his tester filed 300 or so bug reports against the Digital C compiler.  This was met with substantial pushback, but it was a mostly an issue that many of the reports traced to the same underlying bugs.

Bill McKeemon expanded the technique and published "Differential Testing of Software" https://www.cs.swarthmore.edu/~bylvisa1/cs97/f13/Papers/DifferentialTestingForSoftware.pdf

 

In the mid-late 1980s Bill Mckeeman worked with DEC's compiler product teams to introduce fuzz testing into our testing process.  As with the C compiler work at DEC Cambridge, fuzz testing for other compilers (Fortran, PL/I) also found large numbers of bugs.

The pushback from the compiler folks was mainly a matter of priorities.  Fuzz testing is very adept at finding edge conditions, but most failing fuzz tests have syntax that no human programmer would ever write.  As a compiler engineer you have limited time to devote to bug testing.  Do you spend that time addressing real customer issues that have been reported or do you spend it fixing problems with code that no human being would ever write?  To take an example that really happened, a fuzz test consisting of 100 nested parentheses caused an overflow in a parser table (it could only handle 50 nested parens).  Is that worth fixing?

As you pointed out, fuzz test failures tend to occur in clusters and many of the failures eventually are traced to the same underlying bug.  Which leads to the counter-argument to the pushback.  The fuzz tests are finding real underlying bugs.  Why not fix them before a customer runs into them?  That very thing did happen several times.  A customer-reported bug was fixed and suddenly several of the fuzz test problems that had been reported went away.  Another consideration is that, even back in the 1980s, humans weren't the only ones writing programs.  There were programs writing programs and they sometimes produced bizarre (but syntactically correct) code.

-Paul W.