[TIP] unitteset outcomes (pass/fail/xfail/error), humans and automation
Mark Sienkiewicz
sienkiew at stsci.edu
Mon Dec 14 07:53:12 PST 2009
I think the idea of user-defined statuses is a good one, but sometimes
people are too quick to make up a new test status. You need to make an
easier way to get more information out of a test result.
There is a benefit to having a small number of possible statuses: They
group into meaningful categories. A test that reports a status of
"error" is doing something fundamentally different from a test that
reports "fail".
But it is not true that "status" is the only way to get information out
of a test. To use an example mentioned on the list, I don't think
"error" because of HTTP response 400 is so fundamentally different from
"error" because of HTTP response 500 that it deserves a different status.
If you define too many status, at some point you dilute the usefulness.
Let it run wild (say, by having many developers each making up new ways
for a test to fail/error for a few years), and you might end up with the
standard pass, fail, error... plus error_http_500, error_http_400,
error_http_timeout, fail_floating_point_exception, fail_data_format,
fail_syntax, fail_off_by_10_percent, fail_off_by_20_percent, ...
You could end in the interesting situation where you have 0 pass, 0
fail, and 0 error, but 3 error_http_500, 1 error_http_400, 2 fail_fpe, 2
fail_floating_point_exception (different from fpe how?), blah blah blah.
But you still need the information about the failure. That is why we
added the idea of "attributes" to pandokia. The record that reports a
test result can contain a "test result attribute" that conveys some
information about the result. So, for example, you might have
status=E
tra_http_response=400
or
status=E
tra_http_response=timeout
I see no particular reason why you couldn't add something like the
pandokia attributes to unittest. The nice thing about attributes is
that they are a way to get arbitrary information out of the test,
without the test system having to know anything about what it means.
That is, I don't have to _define_ a new attribute -- I just have to use it:
tra['http_response'] = str(httpobj.response)
If you formalize this idea and clearly document it, people will have an
alternative for getting test-specific data in their result. That means
that they do not have an incentive to try to overload the status field
by storing two (or more) different kinds of data in it.
In pandokia, we made Test Definition Attributes (TDA, something about
the test inputs), Test Result Attributes (TRA, something computed during
the test), and Test Configuration Attributes (TCA, something about the
environment we are running in; not fully implemented yet, though).
There are three name spaces because, in our work flow, they are
maintained independently. For example, I wouldn't want the test
administrator (me) to set a new environment variable, add it to the
attributes, then find that the name conflicts with an attribute computed
by one of our thousands of nightly tests. By keeping TCA and TRA in
different name spaces, you never have to worry about that collision.
The name of an attribute is an arbitrary non-case-sensitive string[1],
and the value is a string. In nosetests in Python, we collect the test
results through a nose plugin that recognizes variables tda, tra, and
tca either as globals in the test module or as attributes on the
unittest object, depending on which kind of test nose is running. I
didn't work on the nose plugin, but I seem to remember that all you
would need to add tda/tra/tca to unittest is a way to report them when
the test is done.
So, to summarize my point: Yes, you should offer a way to create new
values for the test status, but you should offer an even easier way to
get information about what happened in a particular test, so that people
are not tempted to create a new test status to encode information that
doesn't really belong to the status.
Mark S.
[1] Why non-case-sensitive? I had problems with people using
inconsistent camel casing. And can you really argue that it is a good
idea to name one field "someRaNdomFieldName" and another
"somEraNdomFieldName"?
More information about the testing-in-python
mailing list