I'm not too familiar with the Harmony code yet, but since I've had a
bunch of experience on large projects I thought I'd toss my $.02 in here.
1) When dealing with a project as large and with as much surface area
as a VM, your unit tests for the entire project will probably take
several hours to run. The trade off for heavy coverage is totally worth
it, even if it takes a long time. It does indeed mean you need to
manage it.
2) We tended to manage this by breaking up unit tests into Build
Verification Tests(BVTs) and Developer Regression Tests(DRTs).
Developers would be required to run DRTs before checking in, and BVTs
would be run for every build (or with coninous integration, they would
be constantly running every few hours).
3) In the largest projects I've been on DRTs would be broken up further
to be on a component level. When you changed a component that other
components depended on, we tended to depend on the good sense of the
developer to run the DRTs for the related components (and depended on
the CI or daily build to catch the problems that slipped through that
net.) We set a rule that DRTs for a given component could never take
longer than 10 minutes to run.
Again, I'm sorry if this is irrelevent (since I'm not familiar enough
with the Harmony code), but this process was reasonably effective for
us. The real pain ends up being how often changes in core code broke
downstream components, but failing tests are only a symptom (and early
warning system) for this.
The problem is that for core components, it was often important for the
developers to run rather a rather long suite of tests before checking
in, simply because there were so many components using their pieces. We
just bit that bullet.
--Will
Richard Liang wrote:
Dears,
I notice that we put all the test code into one big test method (for
example,
org.apache.harmony.tests.java.util.jar.test_putLjava_lang_ObjectLjava_lang_Object).
This way we will lose some benefits of junit and even unit test:
1. Test code cannot share configuration code through setUp and tearDown
2. We have to add redundant code, such as, "Assert 1:", "Assert 2:
...." to make the test results more comprehensive
3. It makes the test code more complex
Shall we just use small test cases?
You may want to read the description at:
http://www.javaworld.com/javaworld/jw-12-2000/jw-1221-junit_p.html
*Keep tests small and fast*
Executing every test for the entire system shouldn't take hours.
Indeed, developers will more consistently run tests that execute
quickly. Without regularly running the full set of tests, it will be
difficult to validate the entire system when changes are made. Errors
will start to creep back in, and the benefits of unit testing will be
lost. This means stress tests and load tests for single classes or
small frameworks of classes shouldn't be run as part of the unit test
suite; they should be executed separately.
Thanks a lot.
Richard Liang wrote:
Dears,
As I cannot find similar pages about testing convention, I just
create one with my rough ideas
http://wiki.apache.org/harmony/Testing_Convention, so that we can
document our decision timely & clearly.
Geir Magnusson Jr wrote:
Leo Simons wrote:
Gentlemen!
On Mon, Mar 27, 2006 at 11:07:51AM +0200, mr A wrote:
On Monday 27 March 2006 10:14, mr B wrote:
On 3/27/06, mr C wrote:
[SNIP]
[SNIP]
[SNIP]
On 1/1/2006, mr D wrote:
[SNIP]
Hmmm... Lemme support [SNIP]
Now let me support [SNIP].
The ASF front page says
(...) "The Apache projects are characterized by a collaborative,
consensus
based development process, " (...)
That's not just some boilerplate. Consensus is a useful thing.
"How should we organize our tests?" has now been the subject of
debate for
*months* around here, and every now and then much of the same
discussion is
rehashed.
And we're making progress. IMO, it really helped my thinking to
distinguish formally between the implementation tests and the spec
tests, because that *completely* helped me come to terms with the
whole o.a.h.test.* issue.
I now clearly see where o.a.h.test.*.HashMapTest fits, and where
java.util.HashMapTest fits.
I don't think our issues were that obvious before, at least to me.
Now, I see clearly.
I think it would be more productive to look for things to agree on
(such as,
"we don't know, but we can find out", or "we have different ideas
on that,
but there's room for both", or "this way of doing things is not the
best one
but the stuff is still useful so let's thank the guy for his work
anyway")
than to keep delving deeper and deeper into these kinds of
disagreements.
Of course, the ASF front page doesn't say that "apache projects are
characterized by a *productive* development process". Its just my
feeling that
for a system as big as harmony we need to be *very* productive.
You don't think we're making progress through these discussions?
Think about it. Is your time better spent convincing lots of other
people to do
their testing differently, or is it better spent writing better tests?
The issue isn't about convincing someone to do it differently, but
understanding the full scope of problems, that we need to embrace
both approaches, because they are apples and oranges, and we need
both apples and oranges. They aren't exclusionary.
geir