I hope you don't mind me throwing in some random comments on this. I
think there are two types of testing that need to be distinguished,
system and unit.
A system test looks that the external behavior of the whole system, so
what it is testing changes only when the API changes, and tests should
apply across many source code revisions. I can see separating those out.
However, I feel River has been weak on unit tests, tests that check the
implementation of e.g. a data structure against its javadoc comments.
Those tests need to change with internal interface changes.
Testing e.g. the multi-thread consistency of a complex data structure
using only external system tests can be a mistake. It may take a very
large configuration or many days of running to bring out a bug that
could be found relatively fast by a unit test that hits the data
structure rapidly.
I'm a little concerned that the reconfiguration of the tests may
represent an increased commitment to only doing system tests, and not
doing any unit tests.
Patricia
On 12/2/2012 10:21 AM, Dan Creswell wrote:
...
On 30 November 2012 19:53, Gregg Wonderly <ge...@cox.net> wrote:
I still wonder why it doesn't feel right that the test suite be in the same branch as the associated
"release". Some of the new code needs new test that demonstrate "functionality" while
other tests that demonstrate compatibility will be ran on each release without change. It seems to me, that
in the end, when a release goes out the door, the tests that validated that release, are part of that
"release".
I have some similar disquiet, here's what I'm thinking at the moment
(subject to change faster than I can type!)...
Compatibility and similar is really "compliance test" and is closely
linked to the APIs defined by the specs. Two flavours here:
(1) "Well-behaved service" tests - does a service do join properly etc.
(2) Compliance tests - do the APIs behave right etc.
These are kind of slow moving as are the APIs at least for now. I feel
right now like (1) might be a subproject applied to our own "built-in"
services as well as others. I'm tempted to say the same about (2) save
for the fact that if we give up on the idea someone else is going to
build a River clone this stuff becomes part of the release/test phase
for the core.
Any other testing we're doing over and above what falls into (1) and
(2) above is part of tests for core and ought to be living in the same
branch and run as part of release. However, that's a little
uncomfortable when one wishes to freeze development of core to do
major work on the test harness etc. You branch core and test suite to
work purely on the suite.
Manageable I guess well, until you have the trunk moving on and
breaking your already seriously under construction test suite where
everything in trunk is "old style" and will be a b*stard to merge
across but if you don't your branched test suite is gonna break for
nuisance reasons.
If we need two different types of tests, and a migration path from "functionality tests"
into "compatibility tests", then maybe we really need two trees for development of each
release, and branching the whole test suite would be one branch an the new release would be the
other.
Is that how you guys are thinking about this?
You have my (current) thinking above...
Gregg Wonderly
On Nov 30, 2012, at 9:43 PM, Peter Firmstone <j...@zeus.net.au> wrote:
On 30/11/2012 12:27 AM, Dan Creswell wrote:
On 29 November 2012 13:11, Peter Firmstone<j...@zeus.net.au> wrote:
The last passing trunk versions:
Jdk6 Ubuntu 1407017
Solaris x86 1373770
Jdk7 Ubuntu 1379873
Windows 1373770
Revision 1373770 looks the most stable, I think all platforms were passing
on this, 1407017 only passed on Ubuntu jdk6, nothing else.
If we can confirm 1373770 as stable, maybe we should branch a release off
that, buying some time to stabilise what we're working on now.
I think we should do that. I'm also tempted to suggest we consider limiting
our development until we've fixed these tests up. Or alternatively control
the rate of patch merging so we can pace it and make sure the tests get
focus.
That's a bit sledgehammer but...
Ok, sounds like a plan, how do you think we should best approach the task?
Create a branch in skunk, just for qa and run tests against released jars?
Regards,
Peter.