diagnosing flaky tests

Victor (here, here) grabbed me on IRC yesterday for some testr help. He had a super frustrating bug in glance: about one in thirty unit test runs, a test would fail. He’d spent hours tracking it down and still couldn’t reliably reproduce it.

Part of that was due to the glance tests taking a few minutes to run, so each iteration was slow, but another part was a lack of familiarity with the test tooling we use in OpenStack which can give rich data to help analyse such things.

I helped him out – and this post is a step by step handbook of what I did so that I can point people at it :)


  1. start by duplicating the environment
  2. setup automation so you are only doing the interesting (or at least not time consuming) bits
  3. bisect and bisect and bisect

Firstly, I pulled down exactly the same code he was working on:

cd glance; git review -d 250083

This let me try to reproduce the thing. However, my normal reproduction facility couldn’t be used because the glance testr configuration depended on invoking testr within lockutils-wrapper. I’m still working through the implications, but for the short term I moved that to be testr’s problem.

So now I could make a python 34 venv and run testr directly:

tox -epy34 --notest; . .tox/py34/bin/activate; testr run --parallel

This is pretty important – it lets me get under the setup.py wrapper that projects use and now I have more control over what is happening. Plus I’m not dealing with tox recreating the venv or anything like that.

It turned out that only the unit tests had been ported to Python3, so I needed to filter down to just those tests. And because I didn’t want to sit here watching it, I set testr off to find a reproduction example on its own:

testr run --parallel --until-failure tests.unit

This runs the same set of tests – whatever you’ve specified in the normal way – in parallel, in a loop. It specifically reschedules and starts new backends (processes that are actually executing test code) each time around, so its very close to just scripting it in shell around testr, with only minor differences (such as not re-querying all the tests each time, because testr knows the full set already).

After an hour or so I had toggled back to look at the terminal, and there was a lovely backtrace and information on the failure. It looks something like this:

lockutils-wrapper \
${PYTHON:-python} -m subunit.run discover -t ./ ./glance/tests --load-list /tmp/tmpafpyzyd5
Ran 2 tests in 0.485s (+0.011s)
PASSED (id=1614)
lockutils-wrapper \
${PYTHON:-python} -m subunit.run discover -t ./ ./glance/tests --load-list /tmp/tmpafpyzyd5
Traceback (most recent call last):
glance.common.exception.NotFound: b'Image not found'
FAIL: glance.tests.unit.v1.test_api.TestGlanceAPI.test_upload_image_http_nonexistent_location_url
tags: worker-0
Traceback (most recent call last):
File "/home/robertc/work/openstack/glance/glance/tests/unit/v1/test_api.py", line 1149, in test_upload_image_http_nonexistent_location_url
self.assertEqual(404, res.status_int)
File "/home/robertc/work/openstack/glance/.tox/py34/lib/python3.4/site-packages/testtools/testcase.py", line 350, in assertEqual
self.assertThat(observed, matcher, message)
File "/home/robertc/work/openstack/glance/.tox/py34/lib/python3.4/site-packages/testtools/testcase.py", line 435, in assertThat
raise mismatch_error
testtools.matchers._impl.MismatchError: 404 != 201
Ran 2 tests in 0.419s (-0.061s)
FAILED (id=1615, failures=1 (+1))

Except that the id was much lower, there were multiple concurrent processes being run on each iteration, and the number of tests run much higher – 506 in fact. The important bit to me was the id, because with that we could get programmatic on the problem.

While testr has support for automatically isolating faults, this depends on deterministic behaviour- there isn’t [yet] any support for things that fail 1 time in N when doing bisection, looking for cross-test interactions and so forth. So normally, one would just run:

testr run --analyze-isolation

And testr would churn away and give a useful answer, in this case I needed to do it by hand. I think there is room for getting really smart about dealing with this sort of situation, but the simplest method is to just repeat each test (a run of some X tests looking for a failure) until some confidence level is reached, rather than assuming a pass is actually a pass.

To do that we needed to do two things: we needed a set of tests to run, and a way to reduce the set and repeat. We could use –until-failure as a way to repeat a given test, and stop it after a couple of hours if it hadn’t failed.

Extracting the tests that a given backend ran is straightforward, if not something you’d just luck upon. If the run id you want to investigate is 6, and the backend that you want to report on is worker-0 (see the test tag in the error report above):

cat .testr/6 | subunit-1to2 | subunit-filter -s --xfail --with-tag=worker-0 | subunit-ls > worker-0

This takes the subunit stream from the repository, which is in a legacy format 1, upgrades it to subunit v2, then includes successful tests and expected failures, but only includes tests run on worker-0, pulls out the test ids (thats the subunit-ls bit) and writes it into a file ‘worker-0’.

To run just those tests:

testr run –load-list worker-0

More interestingly though, lets start by not running all the tests that took place after our failure. Inside the file it looks like this:


Note that the test we’re interested in is in the middle there – though the file looks sorted, thats due to the test backend, what we have is the actual order tests executed in (and we don’t need to worry about concurrency because we pulled out just one backend process, and in Python unittest thats single-threaded).

First step then is to delete all the tests after the one we care about:


And then we don’t want to run all the tests: we’re assuming there is an interaction with a single other test leaving a stale process or something, so we want to run the one that failed, and half of the earlier tests; if they end up being reliable, we switch to the half we hadn’t been running, and then repeat the process – take half, run until we’re satisfied its reliable, repeat.

The way I do this by hand is to just edit the text file, making a new copy each step, so that I can backtrack easily. So delete half the preceeding lines, save to new file, then run:

testr run --load-list newfile --until-failure

Walk away, do something else, and then come back in a couple of hours.

This worked: from 506 tests on the worker, I then had about 300 that ran before the failing test after the first trim, so 150 was my first bisection which ran for a couple hours before failing. Then 75, then 35, then 18, then 9, then 5, then 2, then 1, then – 0 – thats right, eventually we found that the failing test could fail on its own!

And from that Victor was able to dig deep into what it was doing with confidence – he found a race condition in the test server setup stuff (I haven’t looked closely – I’m parroting what he said on IRC) – and is confident he’s found the bug. Yay!

I also ran one of the smaller sets overnight using Python 2.7, and that didn’t fail at all, so I suspect the failure is in some area that was masked by Python 2.7/eventlet on 2.7 handling of *something*. We saw a bug in subunit of that nature earlier this year, where a different (but legitimate) behaviour in eventlet on 3.4 led to subunit dropping writes silently. Thats fixed now in both eventlet and subunit :)

signalling via exit status in Python

A common idiom in non-trivial command line tools is to have more than two return codes. For instance, diff uses 0 for ‘same inputs’, 1 for ‘different inputs’, 2 for ‘trouble’.

Doing that in Python is a little harder though, and since I’ve gotten it wrong in the past, I want to write it down for both myself and anyone else contemplating it.

The issue is that both your program and the Python VM itself can fail, and so if you attempt to use a common status code with those the Python VM uses for failures, you have to make sure that the meanings are at least broadly compatible. There’s also a bug in existing Python releases that will cause an exit status of 0 sometimes when an error is actually appropriate.

I’ve only researched this on CPython, its possible that other Python VM’s behave differently, and as far as I know this is not a language spec issue (but perhaps it should be).


  1. Always flush stdout and stderr yourself, even when signalling errors.
  2. Never use status 1 or 2 for non-error conditions.
  3. (Provisional) don’t use status 120 at all.


CPython exits with 0 when the interpreter cleanup code fails to flush stdout/stderr, even though that would be an error if it happened earlier. To address that, add an explicit flush of both streams before your program ends. We may end up making CPython exit with 120 when the stdout/err flushing fails. There’s also a possibility that a very early threading error may result in a 0 exit code, though I haven’t managed to make this actually happen yet.

CPython exits with 1 when site.py fails to import, so using 1 for non-error conditions makes it hard for callers to discriminate between your meaning and site.py failures.

Cpython exits with 2 when CLI arguments fail to parse, so using 2 for non-error conditions is similar there. optparse also uses 2 for this, so even if you are using a different interpreter, it is not a safe status code to reuse with different semantics.


OpenStack Mitaka debrief

Well, last week was the 6-monthly OpenStack summit in Tokyo. It was fantastic to catch up with many folk, but with 5000 attendees, there are many more that I didn’t see than those that I did. Yet I find the sheer volume of face-to-face stuff nearly overwhelming. I wish it was quite a bit longer and less intense.

Over the next cycle I’ve committed to a few things…

  1. Kicking off TC leadership of scaling for OpenStack. That is, sparking the conversation with the broader community about what scaling means for us, and ensuring each project is paying some attention to it – in the same way that each project already pays attention to e.g. backwards compatibility – they can choose how much, and implementation and so on, but the basic user expectations and framework for thinking about it are shared across OpenStack. The performance working group is certainly related to this but scaling is different to performance.
  2. Replacing the oslo incubator process with one that creates the package straight away. This will go up as a spec for approval of course. The crux of the issue will be finding a way to preserve the freedom of early refactorings without API commitments, without breaking everything. The current approach in my head is to use versioned submodules within the package during the pre-1.0.0 phase, and liberally copy-paste things when API breaks are needed.
  3. Helping the app catalog folk a little bit by doing a review of their review guidelines – looking specifically for gaps (e.g. like the currently unsecured http attack vector).
  4. Start a broad discussion over changing the way we use minimum versions of requirements. Today we raise the minimum version of most requirements quite eagerly. Yet for some like libvirt we instead use feature detection and degrade gracefully when non-latest versions are installed. It seems likely that it would increase compatibility with distributions if we took that approach more widely, but we’d need some care to think through the ramifications.
  5. Kicking off a discussion about leadership training for TC & PTL members. We vote folk into these rolls, but leading isn’t a innate skill. With our constituency of over two thousand developers, spending some money on good leadership training seems like a sound investment. If the TC agrees that its a good idea, my plan is to seek funding from the Board, and aim to make the training be a pre-summit event. This was suggested to me by Colette Alexander.
  6. Seek some more eyeballs on the olso.messaging Kafka driver spec from the HP folk that have been working with Kafka.
  7. Establish connections between Yahoo & HP’s iLO team – they’re seeing the same sort of lockups we did with IPMI on the TripleO test cloud (and the infra-cloud folk are still seeing that) – so I want to see if we can get the bug fixed for everyone.
  8. Work up a clear spec on refactoring the testrepository and subunit2sql layers so that we have all the data store backends in one common repository, an HTTP REST API for consumers like openstack-health, and still have a good experience for CLI users.
  9. Lastly, but not least, work up a formal stabilisation cycle proposal to try and give everyone (product working group, users, core developers) what they want which we seem deadlocked on not doing today. The basic thing to me seems to be fear of the consequences of saying no to feature patches – for pretty good reason; many developers have their income directly tied to achieving things upstream, and when upstream says no, the ensuing discussion is fraught (and there is often information asymmetry present). What we probably need to do is find some balance point – and then socialise the plan very broadly – including the Board, so they can encourage member companies to look after their developers properly.

If any of these things are of interest to you, please feel free to reach out to me :).

Graceful introduction of test servers

A test server acts as a little RPC server where we can ask it to run some tests without paying a full new-process startup cost each time. They are a necessary precondition to online scheduling of tests (because without them the latency of scheduling a test will be orders of magnitude more time than executing the test), as well as potentially enabling better debugger glue by providing an explicit out of band interface.

It’s vital that we don’t break existing users of subunit.run or testrepository when we bring this in – folk don’t react well to having their environment broken. Breaks could occur several different ways – but lets assume that an unmodified .testr.conf will not result in the server code being activated. (It would be nice in theory to Just Work and make things better, but there are lots of ways it could fail, starting with the fact that we have no negotiation step with the things we’re running, and anything else (e.g. exported environment variables) stands a high chance of being eaten by intermediaries like ssh, tox and so on).

So, assuming a new .testr.conf:

  1. A newer subunit.run running with an old testrepository might drop into server mode and then not actually run any tests.
  2. A newer testrepository with an older subunit.run might not go into server mode but not error cleanly.

Testr’s run command has two key interfaces with test backends. Firstly the list interface, where it queries for tests. This is only done when testr needs to know what tests exist (e.g. for offline scheduling). Secondly, the run interface where tests are executed.

In the server based world. testr will have one invoke-a-process interface, and that will offer the two existing interfaces over the basic RPC layer.

To avoid failure 1, we need to ensure we never ask for subunit.run to go into server mode except when testrepository itself can handle it. That implies that we must not insert whatever change we are making into the run_command in .testr.conf, and instead either use a variable substition, or a whole new command key to configure it. I’m in favour of a new command key, because it places less constraints on implementors of other languages.

To avoid failure 2, we need to be able to rigorously determine if a process has gone into server mode. E.g. the server has to send a handshake command of some sort.

Lets talk about failure modes that can occur once we have .testr.conf configured and new subunit.run and testrepository code.

  1. We might have version skew between releases of subunit.run and testrepository on future updates to the RPC server.
  2. We might have a broken testr -> server channel
  3. We might have a broken server -> testr channel
  4. The server might go off into a busy loop or something

For 3, we should version the RPC protocol carefully so that any semantic differences can be detected. Obviously there is a tonne of prior art and everyone is going to scream ‘use grpc’ (or fav RPC of choice). Thats a very sensible thing to do, and subunit can actually sit on top of pretty arbitrary transports as long as they can handle bytestrings and timestamps. That said, my focus in this iteration is to enable the server, porting subunit’s transport to something else won’t save time there (because the RPC angle is going to be a tiny fraction of the development time). I think a simple (new, old) version scheme will do fine (think autotools library soname calculations). If testr offered (5, 1) it means it can speak all versions between 1 and 5, and subunit.run as long as it speaks one of them, should pick that and use it. If we find the need to drop compatibility entirely with a version at some point, we raise the old to a version up from that and move on.

We can deal with 4 by pre-emptively sending a message from testr to the server – a hello message with the supported versions. Likewise, 5 can be dealt with by not considering the server ‘ok’ until we get its initial hello message with a chosen version.

If the server goes into a busy loop – I think we can largely ignore this for now, as its no different than today. (which is, the user notices, gets annoyed, and hits ctrl-C, or their CI job times out. Being able to discriminate between ‘the server is stuck’ and ‘a test is stuck’ would be good – and remembering that in a routed world we don’t know necessarily know the end point for any server…. it might just be routing.

What else – well one long running thing has been the desire to move away from requiring a clean stdin/stdout for test processes. Being broken when some test code decide to write to stdout is *not cool*. This new feature seems like an ideal time to address that. We can’t assume working networking (because e.g. tunnelling over ssh or a container console are important use cases). We could however write a little proxy that uses stdin/stdout with no test code, and then signals (however we’re doing that) that testr is listening on a local port, and tunnel it backwards. (If we choose something simple enough, it may even be possible to do that via parameterised ssh commands and no proxy at all). That does imply that testr itself still needs to be able to talk stdout/stdin. So – because testr has to keep doing that, I’m going to defer tackling this for now: it’s clearly scope creep and as such a dangerous temptation. Layer wise, it’s up to each server to decide how to be responsive when tests are cranky, and how to keep test output from compromising things. That does put the debugger integration work back (or at least, it leaves it as no better than the status quo) but its not in any way prejuidicial to it that I can tell.

Draft RPC spec

RPC packets will be stock subunit packets. Each packet will be for a test called ‘testrepository-rpc’ and contain a ‘application/json’ file attachment (with utf8 encoded text, per the default). The JSON message will be one of the messages defined below.

There are two endpoints, client (the initiator of the connection) and the server. Messages are not idempotent, and may be sent at any time from the client to the server. If a message requires a reply, the server may do so at any time, in any order. Subunit packets may be sent at any time from the client to the server, or the server to the client.

Overall lifecycle of a server:
  1. Client sends a Hello message.
  2. Server sends a Hello response.
  3. Both ends pick the highest common version to define future messages.
  4. Client sends commands, and server actions them.
  5. Client sends a Goodbye message.
  6. Server terminates itself.
Message definitions (version 1):
  • Hello

    Advises the peer of the protocol versions supported.
    {“msg”: “Hello”, “max”: 1, “min”: 1}

  • Goodbye

    Tells the server the client is finished and does not want to run any more tests. The server should cleanup and stop accepting messages. If the server was e.g. a trapdoor into a longer running process, it is undefined whether that longer running process should also terminate or not. No reply is permitted.
    {“msg”: “Goodbye”}

  • List

    Tells the server to list some tests. A “Done” reply is required after all the tests have been listed. The output from the command should be subunit “exists” packets describing the tests that the server can run that were listed in the message. The tests property is optional – if absent, list all available tests.
    {“msg”: “List”, “tests”: [“testid”, …], “nonce”: “arbitrary string here”}

  • Run

    Tells the server to run some tests. A “Done” reply is required after the tests have completed running. The output from the command should be a normal subunit stream resulting from running the tests specified. If the tests property is missing, run all available tests. Tests may be run in whatever order is most useful to the server.
    {“msg”: “List”, “tests”: [“testid”, …], “nonce”: “arbitrary string here”}

  • Done

    Tells the client that some requested command has completed. The nonce must be the nonce for the message that this is in reply to.
    {“msg”: “Done”, “nonce”: “arbitrary string here”}

Implementation sketch

The RPC protocol needs to be accessible to anyone doing this in Python, client *and* server, so subunit seems like the sensible place to define the protocol. It will be pure code – no IO interactions – along with sufficient feature work in subunit’s API to make glueing it into e.g. testrepository and subunit.run straight forward.

In testr, we’ll look for a new command in .testr.conf, expressed much like the run command, and use that to determine that a server mode has been requested. If the server fails to start up, thats an error (e.g. it is up to users to get compatible code in place). When listing and running tests we’ll reuse the server except in isolation modes – both –isolated and –analyze-isolation – where reusing the server would violate the contract they have.

In subunit.run, we’ll add a command line flag to opt-in to the server. In the first implementation, the server is going to just be in-line in the call stack ; no threads or anything. So each command will just be an API call within the existing testtools/unitest2 API with a single subunit packet tacked on the end. We may need to do some ugly stuff to get out of the stock run framework – but I think it is doable.

Testrepository roadmap 2015/16

Testrepository has been moderately successful – its very good at some of the things it aspired to (e.g. debugging sporadic test failures in parallel environments), but other angles have not really been explored.

I’ve set some time aside to correct this, in large part to facilitate some important features for tempest (which has its concurrency currently built on the meta-runner included in testrepository – and I’d like to enable the tempest authors to avoid having to write gnarly concurrency code :))

So my plan is to tackle a few things in the lead up to, and perhaps just after the Tokyo OpenStack summit. I wanted to socialise the proposed changes though, and thus this blog post.


Firstly, a long standing issue is that when one tests several different configurations, testrepository is poor at reporting failures that are configuration specific. For instance, imagine that your test suite is run with both Python 2.7 and 3.4, and both results are loaded into your repository. If a given test ‘X’ fails in the first run, and not the second… after the second run is loaded, it will be reported as ‘passing’.

My proposed fix for this is to call the name of each such run a ‘profile’ and use tags to differentiate between the two runs. So you’d tag the 2.7 run perhaps ‘py27’ and the second ‘py34’, and then tell testrepository that the ‘py27’ and ‘py34’ tags are being used to identify profiles. After that testrepository will only consider two test to apply to the same test if the tags match. Tags that are not specified as being for profiles (e.g. the worker-N tags that the testrepository runner adds to track backends that tests run in) won’t be considered in that comparison. This well then allow testrepository to track that each run was separate and the results are not meant to replace each other. The use of tags allows for test matrices too, in principle– consider python version as one dimension, operating system version as another, and database engine as a third — it would be up to the user. I don’t plan to directly implement a matrix system in the first iteration. A different, more dynamic model is in principle possible: don’t tag things, just log events that will give clues and correlate later – thats not precluded by this tag based approach, and we can always add such a thing later.

The output for the queries of the datastore need to be updated though – we don’t currently report tags in e.g. ‘testr failing –list’. This is a little tricky: the listing format is intended to be a mix of nice-for-humans, and machine consumption. Another approach we considered was to namespace the tests with the profile. This has a couple of disadvantages: it may break an unknown number of deployments if the chosen separator is already in use by people, and secondly, it mixes structured and free-form data in a lossy way. One example of that would be that we’d start interpreting all test ids to see if they are – or are not – namespaced with a profile : thats likely to be fragile, at best. On the other hand it would very easily fit into the list format – which is why it was appealing. On balance though, the fragility and conflation would just add technical debt. Instead, we’ll do the following:

  1. Anything that needs to output a flat list of tests will output that for just one profile. An option will be added to allow querying the profiles for which results might be given. The default will start erroring with a list of available profiles if more than one profile has been specified.
  2. We’ll define a minimal JSON schema for reporting multiple profiles in such places. The excellent jq tool can be used to manipulate that in shell command lines. A command line option will opt into receiving this.

Testrepository has two very related programs inside itself. There is the data store and the various queries it can do – e.g. ‘testr load’ and ‘testr failing’. Then there is the meta-runner, which knows how to run some test processes to execute tests. While strictly speaking this is optional, its been very convenient for working with Python tests to have the meta-runner connected to testr and able to do in-process querying.

The meta-runner will benefit from being updated as well. My intent is to make it capable of running all the tests from all the profiles the user specifies, storing that as one single run in the datastore. Two commands in particular need to change here – `testr list-tests` needs to change in line with the test listing above, and `testr run –load-list` needs to be taught how to deal with multiple profiles. I plan to add a command line option to tell it that JSON is being used, and to select tests across all profiles when a simple list or a test regex is given. Finally the command line can benefit from a command line option to select one or more profiles.


The meta-runner has a crude scheduler – it balances based on historic performance prior to running any backend. An online scheduler will give much greater performance in both unseeded, and skewed data cases- e.g.if many long tests fail due to a bug the run after that will often have some workers finishing well before others – leading to slow test times.

The plan here is to finish the implementation of bidirectional channels to test backends, and then dispatch work to them incrementally

Concurrency plans

Tempest wants to be able to run some tests completely independently, and then others can run together arbitrarily. To facilitate this, the online scheduler will be extended to permit describing an overall plan to run through – e.g. a list of segments, where each segment describes one or more tests that can be run together. The UI to supply that to the scheduler will probably start out as a JSON file listing exact test ids and we can iterate from there based on their experience.

The merits of (careful) impatience

The Python packaging ecosystem has long desired a overhaul and implementation of designed features, but it often stalls on adoption.

I think its time to propose a guiding principle for incremental change.

be carefully impatient

The cautious approach for delivering a new feature in the infrastructure looks like this:

  1. Design the change.
  2. Implement the change(s) needed, in a new major version (e.g Metadata-2.0).
  3. Wait for the new version to be the default everywhere.
  4. Tell users they can use it.

This is frankly terrible. Firstly, we cannot really identify ‘default everywhere’. We can identify ‘default in known distributions’, but behind the firewall setups may lag arbitrarily far behind. Secondly, it makes the cycle time for getting user feedback extraordinarily long: decade plus time windows. Thirdly, as a consequence, we run a large risk of running ahead of our users and delivering less good fixes and improvements than we might do if they were using our latest things and giving us feedback.

So here is how I think we should deliver things instead:

  1. Design the change with specific care that it fails closed and is opt-in.
  2. Implement the change(s) needed, in a new minor version of the tools.
  3. Tell users they can use it.

So, why do I think we can skip waiting for it to be a default?

pip, wheel and setuptools are just as able to be updated as any other Python component. If someone is installing (say) numpy via pip (or easy-install), then by definition they are willing to use things from PyPI, and pip and setuptools are in that category.

And if they are not installing via pip, then the Python packaging ecosystem does not affect them.

If we have opt-in as a design principle, then the adoption process will be bottom up: projects that are willing to say to their users ‘you need new versions of pip and setuptools’ can do so, and use the feature immediately. Projects that want to support users installing their packages with pip but aren’t willing to ask that they also upgrade their pip can hold off.

If we have fails-closed as a design principle, then when a project has opted in, and the user installing the package hasn’t upgraded their pip, things will at least fail rather than silently doing the wrong thing.

I had experience of this in Mock recently: the 1.1.0 and up releases depended on setuptools 17.1. The minimum setuptools we could have depended on (while publishing wheels) was still newer than that in Ubuntu Precise (not to mention RHEL!), so we were forcing an upgrade regardless.

This worked ok but we had two significant issues. Firstly, folk with incorrect Python paths can end up shadowing system installed packages, and for some reason ‘six’ triggered this for multiple users. Secondly, we had a number of different attempts to clearly signal the dependency, as the new features we were using did not fail closed: they were silently ignored by sufficiently old setuptools.

We ended up with a setup_requires="setuptools>17.1" clause in setup.py, which we’re hopeful will fail, or Just Work, consistently.

Bootstrapping developer environments for OpenStack

When I first joined OpenStack, it was incredibly hard to get a development environment working. Its better now in some ways, but not in others.

There are three key things.

Firstly, if you’re going to run devstack (or equally tempest-on-devstack), make a dedicated VM to do it. Use whatever hypervisor you want, but don’t you dare run it on baremetal. Yes it can work on baremetal, but you’ll spend way more time learning about quirks than getting stuff done. Once you’re familiar with devstack you can make an educated choice.

Secondly, if you’re not running devstack [e.g. you’re just hacking on an oslo library or gertty or something] there are a bunch of system packages that will need installing to let various Python packages compile. This was one of the places that was hardest to get going on when I started, and still is pretty hard. I started the bindep project to address this, and its in the process of being rolled out now within OpenStack projects. If the project you’re hacking on has an other-requirements.txt file, then you can just do sudo apt-get install $(bindep -b) – but you’ll need bindep installed first, and its not in distros yet. If there is no such file there is a master fallback one which brings in everything maintained by the OpenStack Infra team. Download it somewhere and add the -f bindep-fallback.txt to your bindep call.

Which brings me to the last point. How do you get bootstrapped with bindep and tox and all the other things that you’ll need. The OpenStack developer docs tell you to install system-wide using apt-get install, and then system-wide install tox. This usually works, but isn’t suitable for multi-user machines, and some distributions don’t include enough metadata for pip to do the right thing, and its confusing to have two different systems managing the same packages.

I prefer to use virtualenv directly, with no reliance on system packages other than truely global things like Python itself. virtualenv is a lovely thing, and it bundles three other important Python tools – pip, wheel and setuptools. All combined this will let you get a fully up to date development environment going without root access – and without confusing your package manager. I still use it even with Python versions that include the venv module, because having the bundled components be up to date is important.

With a virtualenv to host tox and bindep and any other globally-relevant tools, you can run those tools without interacting with your system package manager at all[except when you finally remove the version of Python they use from the system].

Here is how I accomplish this.

  1. Start with Python [any version (2.7 or newer)] installed.
  2. pick a place to make the virtualenv. E.g. ~/devtools
  3. Get a copy of virtualenv to run locally from source. 13.1.0 is a good safe version (at time of writing).
  4. cd into virtualenv-13.1.0 (or whatever number)
  5. Run python ./virtualenv.py ~/devtools
  6. Activate the virtualenv
  7. Install bindep and tox and anything else you need globally using pip. The pip in the virtualenv will be nice and fresh so you’ll get full caching and so on.
  8. (optional) symlink those tools into your user specific path. E.g. from ~/devtools/bin to ~/bin, or some similar directory.
    There are two key gotchas here. Firstly, if you use a bind-dir mounted home in containers, you won’t want this in your user path, since venvs are not very portable. Instead make a per-container path directory you can symlink into.
    Secondly, there is cruft in ~/devtools/bin, so don’t even for a second think of just adding it to your PATH: you’ll end up running the python within the virtualenv by mistake at some point and deeply regretting it.
    If you don’t do this, then just activate the virtualenv any time you want to use the tools.

And now you can circle back, use bindep to install binary dependencies, and off you go!