metaprogramming and politics

Decentralize. Take the red pill.

Archive for the ‘testing’ Category

Running tests against multiple devices/resources (in parallel)

with one comment

devicesHow to best distribute tests against multiple devices or resources with pytest? This interesting question came up during my training in Lviv (Ukraine) at an embedded systems company. Distributing tests to processes can serve two purposes:

  • running the full test suite against each device to verify they all work according to the test specification
  • distributing the test load to several devices of the same type in order to minimize overall test execution time.

The solution to both problems is easy if you use two pytest facilities:

  • the general fixture mechanism: we write a fixture function which provides a device object which is pre-configured for use in tests.
  • the pytest-xdist plugin: we use it to run subprocesses and communicate configuration data for the device fixture from the master process to the subprocesses.

To begin with, let’s configure three devices that are each reachable by a separate IP address. We create a list of ip addresses in a file:

# content of devices.json
["", "", ""]

We now create a local pytest plugin which reads the configuration data, implements a per-process device fixture and the master-to-slave communication to configure each subprocess according to our device list:

# content of

import pytest

def read_device_list():
    import json
    with open("devices.json") as f:
        return json.load(f)

def pytest_configure(config):
     # read device list if we are on the master
     if not hasattr(config, "slaveinput"):
        config.iplist = read_device_list()

def pytest_configure_node(node):
    # the master for each node fills slaveinput dictionary
    # which pytest-xdist will transfer to the subprocess
    node.slaveinput["ipadr"] = node.config.iplist.pop()

def device(request):
    slaveinput = getattr(request.config, "slaveinput", None)
    if slaveinput is None: # single-process execution
        ipadr = read_device_list()[0]
    else: # running in a subprocess here
        ipadr = slaveinput["ipadr"]
    return Device(ipadr)

class Device:
    def __init__(self, ipadr):
        self.ipadr = ipadr

    def __repr__(self):
        return "<Device ip=%s>" % (self.ipadr)

We can now write tests that simply make use of the device fixture by using its name as an argument to a test function:

# content of
import time

def test_device1(device):
    time.sleep(2)  # simulate long test time
    assert 0, device

def test_device2(device):
    time.sleep(2)  # simulate long test time
    assert 0, device

def test_device3(device):
    time.sleep(2)  # simulate long test time
    assert 0, device

Let’s first run the tests in a single-process, only using a single device (also using some reporting option to shorten output):

$ py.test -q --tb=line
================================= FAILURES =================================
/tmp/doc-exec-9/ AssertionError: <Device ip=>
/tmp/doc-exec-9/ AssertionError: <Device ip=>
/tmp/doc-exec-9/ AssertionError: <Device ip=>
3 failed in 6.02 seconds

As to be expected, we get six seconds execution time (3 tests times 2 seconds each).

Now let’s run the same tests in three subprocesses, each using a different device:

$ py.test --tx 3*popen --dist=each -q --tb=line
gw0 I / gw1 I / gw2 I
gw0 [3] / gw1 [3] / gw2 [3]

scheduling tests via EachScheduling
================================= FAILURES =================================
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
9 failed in 6.52 seconds

We just created three subprocesses each running three tests. Instead of 18 seconds execution time (9 tests times 2 seconds per test) we roughly got 6 seconds, a 3-times speedup. Each subprocess ran in parallel three tests against “its” device.

Let’s also run with load-balancing, i.e. distributing the tests against three different devices so that each device executes one test:

$ py.test --tx 3*popen --dist=load -q --tb=line
gw0 I / gw1 I / gw2 I
gw0 [3] / gw1 [3] / gw2 [3]

scheduling tests via LoadScheduling
================================= FAILURES =================================
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
E   AssertionError: <Device ip=>
3 failed in 2.50 seconds

Here each test runs in a separate process against its device, overall more than halfing the test time compared to what it would take in a single-process (3*2=6 seconds). If we had many more tests than subproceses than load-scheduling would distribute tests in real-time to the process which has finished executing other tests.

Note that the tests themselves do not need to be aware of the distribution mode. All configuration and setup is contained in the file.

To summarize the behaviour of the hooks and fixtures in

  • pytest_configure(config) is called both on the master and each subprocess. We can distinguish where we are by checking for presence of config.slaveinput.
  • pytest_configure_node(node) is called for each subprocess. We can fill the slaveinput dictionary which the subprocess slave can then read via its config.slaveinput dictionary.
  • the device fixture only is called when a test needs it. In distributed mode, tests are only collected and executed in a subprocess. In non-distributed mode, tests are run single-process. The Device class is just a stub — it will need to grow methods for actual device communication. The tests can then simply use those device methods.

I’d like to thank Anton and the participants of my three day testing training in Lviv (Ukraine) for bringing up this and many other interesting questions.


I am giving another such professional testing course 25-27th November at the Python Academy in Leipzig. There are still two seats available. Me and other trainers can also be booked for on-site/in-house trainings worldwide.

Written by holger krekel

November 12, 2013 at 7:43 am

metaprogramming in Python: What CPython, PyPy, Pyramid, pytest and politics have in common …

leave a comment »

Metaprogramming in Python too often revolves around metaclasses, which are just a narrow application of the “meta” idea and not a great one at that. Metaprogramming more generally deals with reasoning about program code, about taking a “meta” stance on it.  A metaprogram takes a program as input, often just partial programs like functions or classes. Here are a few applications of metaprogramming:

  • CPython is a metaprogram written in C. It takes Python program code as input and interprets it, so that it runs at a higher level than C.
  • PyPy is a metaprogramm written in Python. It takes RPython program code as input and generates a C-level metaprogram (the PyPy interpreter) which itself interprets Python programs and takes another meta stance by generating Assembler pieces for parts of the interpreation execution. If you like, PyPy is a metaprogram generating metaprograms whereas CPython and typical compilers like GCC are “just” a metaprogram.
  • Pyramid is a metaprogram that takes view, model definitions and http-handling code as input and executes them, thereby raising code on a higher level to implement the “Pyramid application” language.
  • pytest is a metaprogram written in Python, taking test, fixture and plugin functions as input and executing them in a certain manner, thereby implementing a testing language.
  • metaclasses: in Python they allow to intercept class creation and introspect methods and attributes, amending their behaviour. Because metaclass-code usually executes at import time, it often uses global state for implementing non-trivial meta aspects.

Apart from these concrete examples, language compilers, testing tools and web frameworks all have metaprogramming aspects. Creating big or small “higher” level or domain-specific languages within Python is as a typical example of metaprogramming. Python is actually a great language for metaprogramming although it could be better.

In future blog posts i plan to talk about some good metaprogramming practise, particularly:

  • keep the layers/levels separate by good naming and API design
  • define a concise “language” for the programs you take as input
  • avoid creating global state in your metaprograms (and elsewhere)
    which can easily happen with meta-classes executing at import time

Lastly, i see metaprogramming at work not only when coding in a computer language. Discussing the legal framing for executing programs on the internet is some kind of metaprogramming, especially if you consider licensing and laws as human-interpreted code which affects how programs can be written, constructed and executed. In reverse, web applications increasingly affect how we interact with each other other, thereby implementing rules formerly dealt with in the arena of politics. Therefore, metaprogramming and politics are fundamentally connected topics.

have metafun, i. e. take fun stuff as input to generate more of it :) holger

Written by holger krekel

November 22, 2012 at 3:04 pm

If i were to design a new programming language …

with 17 comments

I’d see to base syntax and semantics on Python3, but strip and rebase it:

  • no C: implement the interpreter in RPython, get a JIT for free and implementation bits from PyPy’s Python interpreter (parsing, IO, etc.)
  • no drags-you-down batteries: lean interpreter core and a standard battery distro which is tested against the last N interpreter versions + current
  • no yield: use greenlets to implement all of what yield provides and more
  • no underlying blocking on IO: base it all on event loop, yet provide synchronous programming model through greenlets
  • no c-level API nor ctypes: use cffi to interface with c-libraries
  • no global state: just support state bound to execution context/stack
  • no GIL: support free threading and Automatic Mutual Exclusion for dealing with shared state
  • no have a thought-through story and tools from the start for packaging, installation, depending/interfacing between packages
  • no import, no sys.modules: provide an object with which you can access other packages’s objects and introspect/interact with one’s own package
  • no testing as an afterthought: everything needs to be easily testable, empowered assert statement and branch-coverage supported from the core.
  • no extensibility as an afterthought: support plugins and loose coupling through builtin 1:N calling mechanism (event notification on steroids)
  • no unsafe code: support IO/CPU/RAM sandboxing as a core feature
  • no NIH syndrome: provide a bridge to a virtualenv’ed Python interpreter allowing to leverage existing good crap

Anything else? Probably! Discussion needed? Certainly. Unrealistic? Depends on who would participate — almost all of the above has projects, PEPs and code showcasing viability.

Btw, did you know that when we started PyPy we initially did this under the heading of “Minimal Python”? Some of the above ideas above and their underlying motivations were already mentioned when I invited to the first PyPy sprint almost 10 years ago:

I learned since then that Python has more complex innards than it seems but i still believe it could be both simpler and more powerful.


Written by holger krekel

November 13, 2012 at 3:29 pm

pylib 1.0.0 released: the testing-with-python innovations continue

with 5 comments

Took a few betas but finally i uploaded a 1.0.0 py lib release, featuring the mature and powerful py.test tool and "execnet-style" elastic distributed programming. With the new release, there are many new advanced automated testing features – here is a quick summary:

  • funcargs – pythonic zero-boilerplate fixtures for Python test functions :
    • totally separates test code, test configuration and test setup
    • ideal for integration and functional tests
    • allows for flexible and natural test parametrization schemes
  • new plugin architecture, allowing easy-to-write project-specific and cross-project single-file plugins. The most notable new external plugin is oejskit which naturally enables running and reporting of javascript-unittests in real-life browsers.
  • many new features done in easy-to-improve default plugins, highlights:
    • xfail: mark tests as "expected to fail" and report separately.
    • pastebin: automatically send tracebacks to pocoo paste service
    • capture: flexibly capture stdout/stderr of subprocesses, per-test …
    • monkeypatch: safely monkeypatch modules/classes from within tests
    • unittest: run and integrate traditional tests
    • figleaf: generate html coverage reports with the figleaf module
    • resultlog: generate buildbot-friendly reporting output
  • distributed testing and elastic distributed execution:
    • new unified "TX" URL scheme for specifying remote processes
    • new distribution modes "–dist=each" and "–dist=load"
    • new sync/async ways to handle 1:N communication
    • improved documentation

The py lib continues to offer most of the functionality used by the testing tool in independent namespaces.

Some non-test related code, notably greenlets/co-routines and api-generation now live as their own projects which simplifies the installation procedure because no C-Extensions are required anymore.

The whole package should work well with Linux, Win32 and OSX, on Python 2.3, 2.4, 2.5 and 2.6. (Expect Python3 compatibility soon!)

For more info, see the py.test and py lib documentation:

have fun, holger

Written by holger krekel

August 4, 2009 at 10:05 am

Parametrizing Python tests, generalized.

with one comment

Parametrizing test runs is a kind of a hot topic with Python test tools. py.test recently grew a new pytest_generate_tests hook to parametrize tests. I am going to introduce it by providing ports of Michael Foord‘s recent experiments with parametrizing test cases and an example from Rob Collins testscenarios unittest extension. The gist of the new hook is that it allows to easily implement and combine these schemes. It builds on the general idea of allowing python test functions to receive function arguments ("funcargs") – and defining mechanisms on how to provide them.

The parametrizer example, ported

The idea of Michael Foord‘s Parametrizer example is to define multiple sets of parameters and have specified test functions receive those arguments. Here is a direct port of Michael’s example to use py.test’s new hook:

import py

def pytest_generate_tests(metafunc):
    # called once per each test function
    for funcargs in metafunc.cls.params[metafunc.function.__name__]:
        # schedule a new test function run with applied **funcargs

class TestClass:
    params = {
        'test_equals': [dict(a=1, b=2), dict(a=3, b=3), dict(a=5, b=4)],
        'test_zerodivision': [dict(a=1, b=0), dict(a=3, b=2)],

    def test_equals(self, a, b):
        assert a == b

    def test_zerodivision(self, a, b):
        py.test.raises(ZeroDivisionError, "a/b")

py.test automatically discovers both the pytest_generate_tests hook and the two test functions. For each test function it calls the hook, passing it a metafunc object which provides meta information about the test function and allows to add new tests during collection. Let’s see what just collecting the tests produces:

$ py.test --collectonly

<Module ''>
  <Class 'TestClass'>
    <Instance '()'>
      <FunctionCollector 'test_equals'>
        <Function 'test_equals[0]'>
        <Function 'test_equals[1]'>
        <Function 'test_equals[2]'>
      <FunctionCollector 'test_zerodivision'>
        <Function 'test_zerodivision[0]'>
        <Function 'test_zerodivision[1]'>

So we collected 5 actual runs of test functions. Let now run the test functions:

$ py.test

========================= test session starts =========================
python: platform linux2 -- Python 2.6.2
test object 1: F.F.F

============================== FAILURES ===============================
________________ TestClass.test_equals.test_equals[0] _________________

self = <test_parametrize.TestClass instance at 0x994f8ac>, a = 1, b = 2

    def test_equals(self, a, b):
>       assert a == b
E       assert 1 == 2 AssertionError
________________ TestClass.test_equals.test_equals[2] _________________

self = <test_parametrize.TestClass instance at 0x994f8ac>, a = 5, b = 4

    def test_equals(self, a, b):
>       assert a == b
E       assert 5 == 4 AssertionError
__________ TestClass.test_zerodivision.test_zerodivision[1] ___________

self = <test_parametrize.TestClass instance at 0x994f8ac>, a = 3, b = 2

    def test_zerodivision(self, a, b):
>       py.test.raises(ZeroDivisionError, "a/b")
E       ExceptionFailure: 'DID NOT RAISE' ExceptionFailure
================= 3 failed, 2 passed in 0.13 seconds =================

You can easily see the failing tests and the parameters that the tests received. It also showcases py.test traceback reporting but that’s for another discussion.

The parametrizer example, decorated

So, you say, what about having a decorator specifying test parameters? Here is the same example, letting our hook implement a decorator scheme:


import py

def params(funcarglist):
    def wrapper(function):
        function.funcarglist = funcarglist
        return function
    return wrapper

def pytest_generate_tests(metafunc):
    for funcargs in getattr(metafunc.function, 'funcarglist', ()):

# actual test code, above support code can live elsewhere

class TestClass:
    @params([dict(a=1, b=2), dict(a=3, b=3), dict(a=5, b=4)], )
    def test_equals(self, a, b):
        assert a == b

    @params([dict(a=1, b=0), dict(a=3, b=2)])
    def test_zerodivision(self, a, b):
        py.test.raises(ZeroDivisionError, "a/b")

This variant leaves the "test specification" tightly coupled. Running it with py.test provides the some output as for the first example port.

A quick port of "testscenarios"

Finally, let’s also port Rob Collin’s testscenario example. Here is the implementation of the full mechanism with py.test and the tests in funcarg-style:


def pytest_generate_tests(metafunc):
    for scenario in metafunc.cls.scenarios:
        metafunc.addcall(id=scenario[0], funcargs=scenario[1])

scenario1 = ('basic', {'attribute': 'value'})
scenario2 = ('advanced', {'attribute': 'value2'})

class TestSampleWithScenarios:
    scenarios = [scenario1, scenario2]

    def test_demo(self, attribute):
        assert isinstance(attribute, str)

Let’s run it:

$ py.test -v

================================ test session starts ================================
python: platform linux2 -- Python 2.6.2 -- /usr/bin/python
test object 1: TestSampleWithScenarios.test_demo[basic] PASS TestSampleWithScenarios.test_demo[advanced] PASS

============================= 2 passed in 0.06 seconds ==============================

Easy, isn’t it?

Playing yourself

If you want to play with the examples yourself, you can use hg clone and install it. In the example/parametrize/ direcrory you can tweak and run the test examples. Let me know of comments or problems you may encounter.

Conclusion: deprecating "yield"

The three ports show that pytest_generate_tests is a hook that allows to implement many custom parametrization schemes. You can implement the hook in a test module or in local or global plugin, sharing it in your project or in the community. The hook also integrates well with other usages of funcargs, see the extensive pytest funcarg documentation.

The new way to parametrize test is meant to substitute yield usage of test-functions aka "generative tests", also used by nosetests. yield-style Generative tests have received criticism and despite being the one who invented them, i mostly agree and recommend not using them anymore.

I’d like to thank Samuele Pedroni and Ronny Pfannschmitt who helped to evolve the new hook and pushed me for implementing it. Oh, and did i emphasize that working feedback-based and documentation driven is so much better than going wild on hypothetical usages?

have fun, holger

Written by holger krekel

May 13, 2009 at 2:55 pm

py.test: shrinks code, grows plugins and integration testing

leave a comment »

Just before Pycon i uploaded the lean and mean py.test 1.0.0b1 beta release. A lot of code got moved out, most notably the greenlets C-extension. This simplifies packaging and increases the py lib’s focus on test facilities. It now has a pluginized architecture and provides funcargs which tremendously help with writing functional and integration tests. One such example are py.test’s own acceptance tests checking behaviour of the command line tool from a user perspective. Other features include a zero-install mechanism for distributing tests which also allow to conveniently drive cross-platform integration tests.

Unittesting, functional and integration testing are now official targets. No doubt, Test category naming is a slippery subject and it’s a good idea to consider test category names as labels rather than "either-or" categories. In the end, tests are about being useful for software development which today means coding for a wide variety of environments and involving integration and deployment issues on every corner. I think that testing tools yet have to develop their full potential. In my opinion, automated testing and deployment techniques are to fully integrate with each other, and i consider coding of distributed integration test scenarios as key to that.

The upcoming final py.test 1.0 release i’d like to make a starting point for facilitating the integration of many more test methods and test mechanisms via plugins. Some people already contributed a pytest_figleaf (for coverage testing) and pytest_twisted (for running twisted style tests) although i am still only finalizing API details and writing up docs. So I am very happy how things are turning out and also motivated by the positive feedback on the two testing tutorials that Brian Dorsey and me gave at Pycon (see his writeup).

Btw, if you use the quickstart and encounter any problems, please use the brand new issue tracker on bitbucket. I started hosting a mercurial py.test trunk repository and so far it’s been a positive experience and I guess I fully switch to mercurial soon. Alas, i probably drop setuptools before i go 1.0 with py.test – it simply causes too many troubles. py.test’s trunk has a straightforward and i intend to release a second beta with refined docs and removed setuptools. Stay tuned for many more news in May – right now, i am looking forward to some 1-week offline holiday :)

cheers, holger

Written by holger krekel

April 18, 2009 at 9:05 pm

new: simultanously test your code on all platforms

with 5 comments

It is now convenient to do test runs that transparently distribute a normal test run to multiple CPUs or processes. You can do local code changes and immediately distribute tests of your choice simultanously on all available platforms and python versions.

A small example. Suppose you have a Python pkg directory containing the typical and a test file with this content:

import sys, os

def test_pathsep():
    assert os.sep == "/"

def test_platform():
    assert sys.platform != "darwin"

Without further ado (no special configuration files, no remote installations) i can now run:

py.test pkg/ --dist=each --rsyncdir=pkg --tx socket=  --tx ssh=noco --tx popen//python=python2.4

This will rsync the "pkg" directory to the specified test execution places and then run tests on my windows machine (reachable through the specified socket-connection), my mac laptop (reachable through ssh) and in a local python 2.4 process. Here is the full output of the distributed test run which shows 6 tests (4 passed, 2 failed), i.e. our 2 tests multiplied by 3 platforms. It shows one expected failure like so:

[2] ssh=noco -- platform darwin, Python 2.5.1-final-0 cwd: /Users/hpk/pyexecnetcache

    def test_platform():
>       assert sys.platform != "darwin"
E       assert 'darwin' != 'darwin'
E        +  where 'darwin' = sys.platform

/Users/hpk/pyexecnetcache/pkg/ AssertionError

Hope you find the output obvious enough. I’ve written up some docs in the new distributing tests section.

I also just uploaded a 1.0 alpha py lib release so that you might type "easy_install -U py" to get the alpha release (use at your own risk!). Did i mention that it passes all of its >1000 tests on all platforms simultanously? :)

This is all made possible by py.execnet which is the underlying mechanism for instantiating local and remote processes through SSH- or socket servers and executing code in them, without requiring any prior installation on the remote sides. zero installation really means that you only need a working python interpreter, nothing more. You can make use of this functionality without using py.test. In fact, i plan to soon separate py.test and make the py lib smaller and smaller …

so, enjoy, hope it makes as much sense to you as it makes to me :) And hope to see some of you at Pycon … btw, anybody interested to join me for a Drum and Base party thursday night? cheers, holger

Written by holger krekel

March 23, 2009 at 6:12 pm


Get every new post delivered to your Inbox.