Considering Bens mails, and some personal discussions
yesterday, I refine variant 2, and drop variant 1 (which
actually was an extremely simplified subset of variant 2).
1) We define a bunch of test categories.
- The number of categories should be small and well-arranged. To much
categories will just be confusing and unmaintainable.
- Some suggested categories:
- Excessive (Tests which excessively test an isolated part of the code which is
unlikely to break by unrelated changes, and take
a lot of time to execute, like the #ifdef-ed test in svn_io right now).
- Manual (Tests which require manual interaction, currently only 1 candidate)
- WorkingCopy (Tests which require a working copy)
- NoRepository (Some C tests which do not involve the repository at all)
- RaSerf, RaFile, RaSvn (Tests checking specific functionality of a RA layer)
- FsFs, FsBdb (Tests checking specific functionality of an FS implementation)
- Local (Tests intended to check local working copy functionality, only
accessing a repo as a side effect)
- Repository (Tests intended to check access to the repository via Ra layer)
- Server (Tests just covering server side functionality, no client / working
2) Each test case gets annotated with one or more test
3) Selection of tests:
- When the user does not specify anything, all tests except the ones declared
"Excessive" and/or "Manual" are selected.
- When the user does explicitly specify test numbers and/or categories, the
selection covers all tests which have a given test number or are marked with
at least one of the given categories, or are not marked with any category at
all. Using the special category "default" selects the default set, the
special category "all" selects all tests including the "Excessive" and
- Additionally, the user may specify a list of excluded test numbers and
categories, which are then excluded from the selection as defined by the
three cases above.
4) Calling Syntax:
For both Python and C test executables, I propose that we just allow
to-be-selected test categories mentioned in addition to the test numbers.
The tests to be excluded are preceded by the --exclude option.
Example derived from Bens use-case:
foo.py default excessive --exclude NoRepository
This runs all non-manual tests which actually use an RA layer.
This can be used for the 2nd or 3rd test run when alternating ra layers,
to not run the ra-independent tests twice.
For make check, those lists are to be passed via the CATEGORIES and
(I'm not sure yet whether we allow to also pass single tests there, this
would need some syntax combining the test executable name and number,
For the UI, I suggest category names are case insensitive, but case preserving.
5) Implementation details:
In Python, I'd define an decorator @categories which one can use to
assing categories per test case. In addition, one can assign per-module
default categories which apply to all tests in the test_list which don't
have explicit categories declared. The categories as well as the decorator
will be defined in main.py.
For the C tests, I'd define an enum for the test categories using bit
flags. The svn_test_descriptor_t will gain an additional const field containing
(This email was sent from a mobile device...)
CODESYSŪ a trademark of 3S-Smart Software Solutions GmbH
Inspiring Automation Solutions
3S-Smart Software Solutions GmbH
Dipl.-Inf. Markus Schaber | Product Development Core Technology
Memminger Str. 151 | 87439 Kempten | Germany
Tel. +49-831-54031-979 | Fax +49-831-54031-50
E-Mail: m.schaber_at_codesys.com | Web: codesys.com
CODESYS internet forum: forum.codesys.com
Managing Directors: Dipl.Inf. Dieter Hess, Dipl.Inf. Manfred Werner | Trade register: Kempten HRB 6186 | Tax ID No.: DE 167014915
Von: Ben Reser [ben_at_reser.org]
Gesendet: Donnerstag, 13. Juni 2013 15:39
An: Markus Schaber
Cc: Subversion Dev (dev_at_subversion.apache.org)
Betreff: Re: Proposal for separating Tests into groups
On Thu, Jun 13, 2013 at 3:20 PM, Markus Schaber <m.schaber_at_codesys.com> wrote:
> This are two alternative proposals for the test suite:
> Rationale: Developers restrain from implementing some tests because they just take to much time to run at every commit.
I don't think that's a problem with Subversion. I can run the full
test suite against a single fs layer + ra layer in 5 minutes.
Depending on what I'm touching I may decide to run one or more but
even if I test all 3 ra layers that's only 15 minutes.
I don't recall anyone ever saying I didn't write a test for that
because it would take too long to run. However, I can certainly say
people have avoided writing tests in this project because the tests
would take too long to write (especially our C level tests vs cmdline
tests). I can also say that people have avoided writing tests because
our test harness for the server side doesn't support changing the
server configuration per test.
> Other test systems like JUnit, NUnit or the CODESYS Test Manager come with ways to select unit tests by category, so we could implement something similar with our tests.
> 1) Just two test sets: simple and extended tests:
> Tests which take a lot of time and cover areas which are unlikely to break can be marked as extended tests. For Python, there will be an @extended decorator, for the C tests we'll have macros like SVN_TEST_EXTENDED_PASS.
> Then running the test with an command line option --simple will skip those tests. Explicitly mentioning extended tests by number will still run them, and can be combined with the --simple flag.
I really don't see a reason to do this.
> Continous integration systems and the tests before signing an release should still execute the full test suite.
> But before a small, not destabilizing commit (use common sense here), only running the non-extended tests is mandatory (and maybe the extended tests covering that specific area.)
There is absolutely no way to enforce a test run in this project. So
the entire concept of a mandatory test run before committing is
pointless. What tests a developer runs is ALWAYS going to be a matter
of the developer using their own judgement. For the most part I don't
see too many broken things being committed that are broken even with
our current test suite situation. When broken things are committed
it's usually because the developer didn't understand their change was
impacted by ra or fs differences and the only thing that would have
prevented it would have been more testing, not less.
> For make check, it would be an SIMPLE=true variable.
> 2) Test Categories:
> A set of categories is defined in a central place.
> Examples for such categories could be:
> - Smoke: For smoke tests, only the most important.
> - Fsfs: Tests covering only the FSFS specific code.
> - Offline: Tests which do not contact the server.
> - Repository: Tests which cover the repository, without a client involved (e. G. svnadmin)
> Each test then gets attributed with the categories which are valid for this test.
> When running the tests, one could pass a parameter to run only the tests which are attributed with at least one of the given flags. For example, if you changed something in FSFS, "--categories=Smoke,Fsfs" would run the smoke tests and the FSFS tests. A second "--exclude=Repository,5,7" switch could be used to exclude test categories as well as single tests by number.
> For make check, we'd have a CATEGORIES and EXCLUDE variables.
I'm more in the favor of something like this because right now some
tests don't use an RA layer or even an FS layer (i.e. some C tests).
If you run tests across all FS and RA layers you end up running these
tests multiple times. Granted that most of these tests are relatively
fast, there is still duplication.
Received on 2013-06-14 13:54:06 CEST