[Avocado-devel] RFC: Nested tests (previously multi-stream test) [v5]
Ademar Reis
areis at redhat.com
Wed May 25 21:36:20 UTC 2016
On Wed, May 25, 2016 at 04:18:38PM -0300, Cleber Rosa wrote:
>
>
> On 05/24/2016 11:53 AM, Lukáš Doktor wrote:
> > Hello guys,
> >
> > this version returns to roots and tries to define clearly the single
> > solution I find teasing for multi-host and other complex tests.
> >
> > Changes:
> >
> > v2: Rewritten from scratch
> > v2: Added examples for the demonstration to avoid confusion
> > v2: Removed the mht format (which was there to demonstrate manual
> > execution)
> > v2: Added 2 solutions for multi-tests
> > v2: Described ways to support synchronization
> > v3: Renamed to multi-stream as it befits the purpose
> > v3: Improved introduction
> > v3: Workers are renamed to streams
> > v3: Added example which uses library, instead of new test
> > v3: Multi-test renamed to nested tests
> > v3: Added section regarding Job API RFC
> > v3: Better description of the Synchronization section
> > v3: Improved conclusion
> > v3: Removed the "Internal API" section (it was a transition between
> > no support and "nested test API", not a "real" solution)
> > v3: Using per-test granularity in nested tests (requires plugins
> > refactor from Job API, but allows greater flexibility)
> > v4: Removed "Standard python libraries" section (rejected)
> > v4: Removed "API backed by cmdline" (rejected)
> > v4: Simplified "Synchronization" section (only describes the
> > purpose)
> > v4: Refined all sections
> > v4: Improved the complex example and added comments
> > v4: Formulated the problem of multiple tasks in one stream
> > v4: Rejected the idea of bounding it inside MultiTest class
> > inherited from avocado.Test, using a library-only approach
> > v5: Avoid mapping ideas to multi-stream definition and clearly
> > define the idea I bear in my head for test building blocks
> > called nested tests.
> >
> >
> > Motivation
> > ==========
> >
> > Allow building complex tests out of existing tests producing a single
> > result depending on the complex test's requirements. Important thing is,
> > that the complex test might run those tests on the same, but also on a
> > different machine allowing simple development of multi-host tests. Note
> > that the existing tests should stay (mostly) unchanged and executable as
> > simple scenarios, or invoked by those complex tests.
> >
> > Examples of what could be implemented using this feature:
> >
> > 1. Adding background (stress) tasks to existing test producing
> > real-world scenarios.
> > * cpu stress test + cpu hotplug test
> > * memory stress test + migration
> > * network+cpu+memory test on host, memory test on guest while
> > running migration
> > * running several migration tests (of the same and different type)
> >
> > 2. Multi-host tests implemented by splitting them into components and
> > leveraging them from the main test.
> > * multi-host migration
> > * stressing a service from different machines
> >
> >
> > Nested tests
> > ============
> >
> > Test
> > ----
> >
> > A test is a receipt explaining prerequisites, steps to check how the
> > unit under testing behaves and cleanup after successful or unsuccessful
> > execution.
> >
>
> You probably meant "recipe" instead of "receipt". OK, so this is an
> abstract definition...
>
> > Test itself contains lots of neat features to simplify logging, results
> > analysis and error handling evolved to simplify testing.
> >
>
> ... while this describes concrete conveniences and utilities that users of
> the Avocado Test class can expect.
>
> > Test runner
> > -----------
> >
> > Is responsible for driving the test(s) execution, which includes the
> > standard test workflow (setUp/test/tearDown), handle plugin hooks
> > (results/pre/post) as well as safe interruption.
> >
>
> OK.
>
> > Nested test
> > -----------
> >
> > Is a test invoked by other test. It can either be executed in foreground
>
> I got from this proposal that a nested test always has a parent. Basic
> question is: does this parent have to be a regular (that is, non-nested)
> test?
>
> Then, depending on the answer, the following question would also apply: do
> you believe a nesting level limit should be enforced?
Let's not introduce yet another concept here. I don't think there
would be a need for a "non-nested parent" rule.
Ditto for nesting level limits, I see no reason to break the
current abstraction. If we introduce enforcement, then the test
will have to be run in an environment that knows if it's nested
or not (and at which level). Proper error handling will require
this information and/or even worse: test writers may have access
to this variable and start using it.
>
> > (while the main test is waiting) or in background along with the main
> > (and other background tests) test. It should follow the default test
> > workflow (setUp/test/tearDown), it should keep all the neat test feature
> > like logging and error handling and the results should also go into the
> > main test's output, with the nested test's id as prefix. All the
> > produced files of the nested test should be located in a new directory
> > inside the main test results dir in order to be able to browse either
> > overall results (main test + nested tests) or just the nested tests ones.
> >
>
> Based on the example given later, you're attributing to the NestedRunner the
> responsibility to put the nested test results "in the right" location. It
> sounds appropriate. The tricky questions are really how they show up in the
> overall job/test result structure, because that reflects how much the
> NestedRunner looks like a "Job".
Since it's a test, which kind of visibility does the job have
about it? Is the result interpretation entirely up to the parent
test? Would nested-tests results be considered arbitrary data?
(think of test-results storage, or a database).
>
> > Resolver
> > --------
> >
> > Resolver is an avocado component resolving a test reference into a list
> > of test templates compound of the test name, params and other
> > `avocado.Test.__init__` arguments.
> >
> > Very simple example
> > -------------------
> >
> > This example demonstrates how to use existing test (SimpleTest
> > "/usr/bin/wget example.org") in order to create a complex scenario
> > (download the main page from example.org from multiple computers almost
> > concurrently), without any modifications of the `SimpleTest`.
This won't sound as a surprise to you, but although this is a
valid use-case, I don't think nested tests is the proper solution
for it. :-)
> >
> > import avocado
> >
> > class WgetExample(avocado.Test):
> > def test(self):
> > # Initialize nested test runner
> > self.runner = avocado.NestedRunner(self)
> > # This is what one calls on "avocado run"
> > test_reference = "/usr/bin/wget example.org"
> > # This is the resolved list of templates
> > tests = avocado.resolver.resolve(test_reference)
> > # We could support list of results, but for simplicity
> > # allow only single test.
> > assert len(tests) == 1, ("Resolver produced multiple test "
> > "names: %s\n%s" % (test_reference,
> > tests)
> > test = tests[0]
> > for machine in self.params.get("machines"):
> > # Query a background job on the machine (local or
> > # remote) and return test id in order to query for
> > # the particular results or task interruption, ...
> > self.runner.run_bg(machine, test)
Here we're missing something: you've described what a nested-test
is, but now you're introducing another concept: the ability to
run (some of these) sub-tests on different machines or
environments.
Which is precisely where it gets ugly, as it brings to the layer
of tests concepts which belong to a job. You should write at
least one section in your RFC to describe what you have in
mind in this case.
> > # Wait for all background tasks to finish, raise exception
> > # if any of them fails.
> > self.runner.wait(ignore_errors=False)
You also didn't say anything about synchronization, although I'm
sure you do have something in mind. Do you expect nested-tests to
communicate with, or depend on, each other?
> >
>
> Just for accounting purposes at this point, and not for applying judgment,
> let's take note that this approach requires the following sets of APIs to
> become "Test APIs":
>
> * avocado.NestedRunner
> * avocado.resolver
>
> Now, doing a bit of judgment. If I were an Avocado newcomer, looking at the
> Test API docs, I'd be intrigued at how these belong to the same very select
> group that includes only:
>
> * avocado.Test
> * avocado.fail_on
> * avocado.main
> * avocado.VERSION
>
> I'm not proposing a different approach or a different architecture. If the
> proposed architecture included something like a NestedTest class, then
> probably the feeling is that it would indeed naturally belong to the same
> group. I hope I managed to express my feeling, which may just be
> overreaction. If others share the same feeling, then it may be a sign of a
> red flag.
>
> Now, considering my feeling is not an overreaction, this is how such an
> example could be written so that it does not put NestedRunner and resolver
> in the Test API namespace:
>
> from avocado import Test
> from avocado.utils import nested
>
> class WgetExample(Test):
> def test(self):
> reference = "/usr/bin/wget example.org"
> tests = []
> for machine in self.params.get("machines"):
> tests.append(nested.run_test_reference(self, reference, machine))
> nested.wait(tests, ignore_errors=False)
>
> This would solve the crossing (or pollution) of the Test API namespace, but
> it has a catch: the test reference resolution is either included in
> `run_test_reference` (which is a similar problem) or delegated to the remote
> machine. Having the reference delegated sounds nice, until you need to
> identify backing files for the tests and copy them over to the remote
> machine. So, take this as food for thought, and not as a fail proof
> solution.
>
> > When nothing fails, this usage has no benefit over the simple logging
> > into a machine and firing up the command. The difference is, when
> > something does not work as expected. With nested test, one get a runner
> > exception if the machine is unreachable. And on test error he gets not
> > only overall log, but also the per-nested-test results simplifying the
> > error analysis. For 1, 2 or 3 machines, this makes no difference, but
> > imagine you want to run this from hundreds of machines. Try finding the
> > exception there.
> >
>
> I agree that it's nice to have the nested tests' logs. What you're
> proposing is *core* (as in Test API) convenience, over something like:
>
> from avocado import Test
> from avocado.utils import nested
>
> class WgetExample(Test):
> def test(self):
> reference = "/usr/bin/wget example.org"
> tests = []
> for machine in self.params.get("machines"):
> tests.append(nested.run_test_reference(self, reference, machine))
> nested.wait(tests, ignore_errors=False)
> nested.save_results(tests,
> os.path.join(self.resultsdir, "nested"))
I agree it should not be part of the core API.
>
> > Yes, you can implement the above without nested tests, but it requires a
> > lot of boilerplate code to establish the connection (or raise an
> > exception explaining why it was not possible and I'm not talking about
> > "unable to establish connection", but granularity like "Invalid
> > password", "Host is down", ...). Then you'd have to setup the output
> > logging for that particular task, add the prefix, run the task (handling
> > all possible exceptions) and interpret the results. All of this to get
> > the same benefits very simple avocado test provides you.
> >
>
> Having boiler plate code repeatedly written by users is indeed not a good
> thing. And a well thought out API for users is the way to prevent boiler
> plate code from spreading around in tests.
>
> The exception handling, that is, raising exceptions to flag failures in the
> nested tests execution is also a given IMHO.
>
> > Advanced example
> > ----------------
> >
> > Imagine a very complex scenario, for example a cloud with several
> > services. One could write a big-fat test tailored just for this scenario
> > and keep adding sub-scenarios producing unreadable source code.
> >
> > With nested tests one could split this task into tests:
> >
> > * Setup a fake network
> > * Setup cloud service
> > * Setup in-cloud service A/B/C/D/...
> > * Test in-cloud service A/B/C/D/...
> > * Stress network
> > * Migrate nodes
I don't understand your motivation here. Do you mean that setting
up a fake network as a (sub-)test would be a positive thing?
> >
> > New variants could be easily added, for example DDoS attack to some
> > nodes, node hotplug/unplug, ... by invoking those existing tests and
> > combining them into a complex test.
> >
> > Additionally note that some of the tests, eg. the setup cloud service
> > and setup in-cloud service are quite generic tests, what could be reused
> > many times in different tests. Yes, one could write a library to do
> > that, but in that library he'd have to handle all exceptions and provide
> > nice logging, while not clutter the main output with unnecessary
> > information.
Or one could create a job that runs the individual tests as
needed.
For this particular use-case, a custom job has many advantages.
To mention just one: the multiplexer.
> >
> > Job results
> > -----------
> >
> > Combine (multiple) test results into understandable format. There are
> > several formats, the most generic one is file format:
> >
> > .
> > ├── id -- id of this job
> > ├── job.log -- overall job log
> > └── test-results -- per-test-directories with test results
> > ├── 1-passtest.py:PassTest.test -- first test's results
> > └── 2-failtest.py:FailTest.test -- second test's results
> >
> > Additionally it contains other files and directories produced by avocado
> > plugins like json, xunit, html results, sysinfo gathering and info
> > regarding the replay feature.
> >
>
> OK, this is pretty much a review.
>
> > Test results
> > ------------
> >
> > In the end, every test produces results, which is what we're interested
> > in. The results must clearly define the test status, should provide a
> > record of what was executed and in case of failure, they should provide
> > all the information in order to find the cause and understand the failure.
> >
> > Standard tests does that by providing test log (debug, info, warning,
> > error, critical), stdout, stderr, allowing to write to whiteboard and
> > attach files in the results directory. Additionally due to structure of
> > the test one knows what stage(s) of the test failed and pinpoint exact
> > location of the failure (traceback in the log).
> >
> > .
> > ├── data -- place for other files produced by a test
> > ├── debug.log -- debug, info, warn, error log
> > ├── remote.log -- additional log regarding remote session
> > ├── stderr -- standard error
> > ├── stdout -- standard output
> > ├── sysinfo -- provided by sysinfo plugin
> > │ ├── post
> > │ ├── pre
> > │ └── profile
> > └── whiteboard -- file for arbitrary test data
> >
> > I'd like to extend this structure of either a directory "subtests", or
> > convention for directories intended for nested test results `r"\d+-.*"`.
> >
>
> Having them on separate sub directory is less intrusive IMHO. I'd even
> argue that `data/nested` is the way to go.
+1.
>
> > The `r"\d+-.*"` reflects the current test-id notation, which nested
> > tests should also respect, replacing the serialized-id by
> > in-test-serialized-id. That way we easily identify which of the nested
> > tests was executed first (which does not necessarily mean it finished as
> > first).
So the nested-tests will have "In-Test-Test-IDs", which are different
than "Test-IDs".
> >
> > In the end nested tests should be assigned a directory inside the main
> > test's results (or main test's results/subtests) and it should produce
> > the data/debug.log/stdout/stderr/whiteboard in there as well as
> > propagate the debug.log with a prefix to the main test's debug.log (as
> > well as job.log).
> >
> > └── 1-parallel_wget.py:WgetExample.test -- main test
> > ├── data
> > ├── debug.log -- contains main log + nested logs with prefixes
> > ├── remote.log
> > ├── stderr
> > ├── stdout
> > ├── sysinfo
> > │ ├── post
> > │ ├── pre
> > │ └── profile
> > ├── whiteboard
> > ├── 1-_usr_bin_wget\ example.org -- first nested test
> > │ ├── data
> > │ ├── debug.log -- contains only this nested test log
> > │ ├── remote.log
> > │ ├── stderr
> > │ ├── stdout
> > │ └── whiteboard
> > ├── 2-_usr_bin_wget\ example.org -- second nested test
> > ...
> > └── 3-_usr_bin_wget\ example.org -- third nested test
> > ...
And with the above, a test-ID is not unique anymore in logs and
in the results directory. For example, when looking for
"1-foobar.py", I may find:
- foobar.py, the first test run inside the job
AND
- multiple foobar.py, run as a nested test inside an arbitrary
parent test.
That's why I said you would need "In-Test-Test-IDs" (or
"Nested-Test-IDs").
> >
> > Note that nested tests can finish with any result and it's up to the
> > main test to evaluate that. This means that theoretically you could find
> > nested tests which states `FAIL` or `ERROR` in the end. That might be
> > confusing, so I think the `NestedRunner` should append last line to the
> > test's log saying `Expected FAILURE` to avoid confusion while looking at
> > results.
> >
>
> This special injection, and special handling for that matter, actually makes
> me more confused.
Agree. This is something to add to the parent log (which is
waiting for the nested-test result).
>
> > Note2: It might be impossible to pass messages in real-time across
> > multiple machines, so I think at the end the main job.log should be
> > copied to `raw_job.log` and the `job.log` should be reordered according
> > to date-time of the messages. (alternatively we could only add a contrib
> > script to do that).
You probably mean debug.log (parent test), not job.log.
I'm assuming the nested tests would run in "jobless" mode (is
that the case? If yes, you need to specify what it means).
> >
>
> Definitely no to another special handling. Definitely yes to a post-job
> contrib script that can reorder the log lines.
+1
>
> >
> > Conclusion
> > ==========
> >
> > I believe nested tests would help people covering very complex scenarios
> > by splitting them into pieces similarly to Lego. It allows easier
> > per-component development, consistent results which are easy to analyze
> > as one can see both, the overall picture and the specific pieces and it
> > allows fixing bugs in all tests by fixing the single piece (nested test).
> >
>
> It's pretty clear that running other tests from tests is *useful*, that's
> why it's such a hot topic and we've been devoting so much energy to
> discussing possible solutions. NestedTests is one to do it, but I'm not
> confident we have enough confidence to make it *the* way to do it. The
> feeling that I have at this point, is that maybe we should prototype it as
> utilities to:
>
> * give Avocado a kickstart on this niche/feature set
> * avoid as much as possible user-written boiler plate code
> * avoid introducing *core* test APIs that would be set in stone
>
> The gotchas that we have identified so far, are IMHO, enough to restrain us
> from forcing this kind of feature into the core test API, which we're in
> fact, trying to clean up.
>
> With user exposition and feedback, this, a modified version or a completely
> different solution, can evolve into *the* core (and supported) way to do it.
>
I tend to disagree. I think it should be the other way around:
maybe, once we have a Job API, we can consider the possibilities
of supporting nested-tests, reusing some of the other concepts.
Nested tests (as in: "simply running tests inside tests") is
relatively OK to digest. Not that I like it, but it's relatively
simple.
But what Lukas is proposing involves at least three more features
or APIs, all of which relate to a Job and should be implemented
there before being considered in the context of a test:
- API and mechanism for running tests on different machines or
environments (at least at first, a Job API)
- API and mechanism for running tests in parallel (ditto)
- API and mechanism to allow tests to synchronize and wait for
barriers (which might be useful once we can run tests in
parallel).
To me the idea of "nested tests than can be run in multiple
machines, under different configurations and with synchronization
between them" is fundamentally flawed. It's a huge layer
violation that brings all kinds of architectural problems.
Thanks.
- Ademar
--
Ademar Reis
Red Hat
^[:wq!
More information about the Avocado-devel
mailing list