After a
very long porting journey,
Launchpad is finally running on Python 3 across
all of our systems.
I wanted to take a bit of time to reflect on why my emotional responses to
this port differ so much from those of some others who ve done large ports,
such as the
Mercurial
maintainers.
It s hard to deny that we ve had to burn a lot of time on this, which I m
sure has had an opportunity cost, and from one point of view it s
essentially running to stand still: there is no single compelling feature
that we get solely by porting to Python 3, although it s clearly a
prerequisite for tidying up old compatibility code and being able to use
modern language facilities in the future. And yet, on the whole, I found
this a rewarding project and enjoyed doing it.
Some of this may be because by inclination I m a maintenance programmer and
actually enjoy this sort of thing. My default view tends to be that
software version upgrades may be a pain but it s much better to get that
pain over with as soon as you can rather than trying to hold back the tide;
you can certainly get involved and try to shape where things end up, but
rightly or wrongly I can t think of many cases when a righteously indignant
user base managed to arrange for the old version to be maintained in
perpetuity so that they never had to deal with the new thing (
OK, maybe Perl
5 counts here).
I think a more compelling difference between Launchpad and Mercurial,
though, may be that very few other people really had a vested interest in
what Python version Launchpad happened to be running, because it s all
server-side code (aside from some client libraries such as
launchpadlib, which were ported
years ago). As such, we weren t trying to do this with the internet having
Strong Opinions at us. We were doing this because it was obviously the only
long-term-maintainable path forward, and in more recent times because some
of our library dependencies were starting to drop support for Python 2 and
so it was obviously going to become a practical problem for us sooner or
later; but if we d just stayed on Python 2 forever then fundamentally hardly
anyone else would really have cared directly, only maybe about some indirect
consequences of that. I don t follow Mercurial development so I may be
entirely off-base, but if other people were yelling at me about how late my
project was to finish its port, that
in itself would make me feel more
negatively about the project even if I thought it was a good idea. Having
most of the pressure come from ourselves rather than from outside meant that
wasn t an issue for us.
I m somewhat inclined to think of the process as an extreme version of
paying down technical debt. Moving from Python 2.7 to 3.5, as we just did,
means skipping over multiple language versions in one go, and if similar
changes had been made more gradually it would probably have felt a lot more
like the typical dependency update treadmill. I appreciate why not everyone
might want to think of it this way: maybe this is just my own rationalization.
Reflections on porting to Python 3
I m not going to defend the Python 3 migration process; it was pretty rough
in a lot of ways. Nor am I going to spend much effort relitigating it here,
as it s already been done to death elsewhere, and as I understand it the
core Python developers have got the message loud and clear by now. At a
bare minimum, a lot of valuable time was lost early in Python 3 s lifetime
hanging on to flag-day-type porting strategies that were impractical for
large projects, when it should have been providing for bilingual
strategies (code that runs in both Python 2 and 3 for a transitional period)
which is where most libraries and most large migrations ended up in
practice. For instance, the early advice to library maintainers to maintain
two parallel versions or perhaps translate dynamically with
2to3 was
entirely impractical in most non-trivial cases and wasn t what most people
ended up doing, and yet the idea that
2to3 is all you need still floats
around Stack Overflow and the like as a result. (These days, I would
probably point people towards something more like
Eevee s porting
FAQ
as somewhere to start.)
There are various fairly straightforward things that people often suggest
could have been done to smooth the path, and I largely agree: not removing
the
u'' string prefix only to put it back in 3.3, fewer gratuitous
compatibility breaks in the name of tidiness, and so on. But if I had a
time machine, the number one thing I would ask to have been done differently
would be introducing type annotations in Python 2 before Python 3 branched
off. It s true that it s
technically
possible
to do type annotations in Python 2, but the fact that it s a different
syntax that would have to be fixed later is offputting, and in practice it
wasn t widely used in Python 2 code. To make a significant difference to
the ease of porting, annotations would need to have been introduced early
enough that lots of Python 2 library code used them so that porting code
didn t have to be quite so much of an exercise of manually figuring out the
exact nature of string types from context.
Launchpad is a complex piece of software that interacts with multiple
domains: for example, it deals with a database,
HTTP, web page rendering,
Debian-format archive publishing, and multiple revision control systems, and
there s often overlap between domains. Each of these tends to imply
different kinds of string handling. Web page rendering is normally done
mainly in Unicode, converting to bytes as late as possible; revision control
systems normally want to spend most of their time working with bytes,
although the exact details vary;
HTTP is of course bytes on the wire, but
Python s
WSGI interface has some
string type
subtleties.
In practice I found myself thinking about at least four string-like types
(that is, things that in a language with a stricter type system I might well
want to define as distinct types and restrict conversion between them):
bytes, text, ordinary native strings (
str in either language, encoded to
UTF-8 in Python 2), and native strings with
WSGI s encoding rules. Some of
these are emergent properties of writing in the intersection of Python 2 and
3, which is effectively a specialized language of its own without coherent
official documentation whose users must intuit its behaviour by comparing
multiple sources of information, or by referring to unofficial porting
guides: not a very satisfactory situation. Fortunately much of the
complexity collapses once it becomes possible to write solely in Python 3.
Some of the difficulties we ran into are not ones that are typically thought
of as Python 2-to-3 porting issues, because they were changed later in
Python 3 s development process. For instance, the
email module was
substantially improved in around the 3.2/3.3 timeframe to handle Python 3 s
bytes/text model more correctly, and since Launchpad sends quite a few
different kinds of email messages and has some quite picky tests for exactly
what it emits, this entailed a lot of work in our email sending code and in
our test suite to account for that. (It took me a while to work out whether
we should be treating raw email messages as bytes or as text; bytes turned
out to work best.) 3.4 made some tweaks to the implementation of
quoted-printable encoding that broke a number of our tests in ways that took
some effort to fix, because the tests needed to work on both 2.7 and 3.5.
The list goes on. I got quite proficient at digging through Python s git
history to figure out when and why some particular bit of behaviour had changed.
One of the thorniest problems was parsing
HTTP form data. We mainly rely on
zope.publisher for this, which
in turn relied on
cgi.FieldStorage; but
cgi.FieldStorage is
badly broken in some
situations on Python 3. Even if that
bug were fixed in a more recent version of Python, we can t easily use
anything newer than 3.5 for the first stage of our port due to the version
of the base
OS we re currently running, so it wouldn t help much. In the
end I fixed some minor issues in the
multipart module (and was kindly
given co-maintenance of it) and
converted zope.publisher to use
it. Although
this took a while to sort out, it seems to have gone very well.
A couple of other interesting late-arriving issues were around
pickle. For most things
we normally prefer safer formats such as
JSON, but there are a few cases
where we use pickle, particularly for our session databases. One of my
colleagues pointed out that I needed to remember to tell
pickle to
stick
to protocol
2,
so that we d be able to switch back and forward between Python 2 and 3 for a
while; quite right, and we later ran into a similar problem with
marshal too. A more
surprising problem was that
datetime.datetime objects pickled on Python 2
require special care when unpickling
on Python 3; rather than the approach that ended up being implemented and
documented
for Python 3.6, though, I preferred a
custom
unpickler,
both so that things would work on Python 3.5 and so that I wouldn t have to
risk affecting the decoding of other pickled strings in the session database.
General lessons
Writing this over a year after Python 2 s end-of-life date, and certainly
nowhere near the leading edge of Python 3 porting work, it s perhaps more
useful to look at this in terms of the lessons it has for other large
technical debt projects.
I mentioned in my
previous article that
I used the approach of an enormous and frequently-rebased git branch as a
working area for the port, committing often and sometimes combining and
extracting commits for review once they seemed to be ready. A port of this
scale would have been entirely intractable without a tool of similar power
to
git rebase, so I m very glad that we finished migrating to git in 2019.
I relied on this right up to the end of the port, and it also allowed for
quick assessments of how much more there was to land.
git
worktree was also helpful, in that I
could easily maintain working trees built for each of Python 2 and 3 for comparison.
As is usual for most multi-developer projects, all changes to Launchpad need
to go through code review, although we sometimes make exceptions for very
simple and obvious changes that can be self-reviewed. Since I knew from the
outset that this was going to generate a lot of changes for review, I
therefore structured my work from the outset to try to make it as easy as
possible for my colleagues to review it. This generally involved keeping
most changes to a somewhat manageable size of 800 lines or less (although
this wasn t always possible), and arranging commits mainly according to the
kind of change they made rather than their location. For example, when I
needed to fix issues with
/ in Python 3 being true division rather than
floor division, I did so in
one
commit
across the various places where it mattered and took care not to mix it with
other unrelated changes. This is good practice for nearly any kind of
development, but it was especially important here since it allowed reviewers
to consider a clear explanation of what I was doing in the commit message
and then skim-read the rest of it much more quickly.
It was vital to keep the codebase in a working state at all times, and
deploy to production reasonably often: this way if something went wrong the
amount of code we had to debug to figure out what had happened was always
tractable. (Although I can t seem to find it now to link to it, I saw an
account a while back of a company that had taken a flag-day approach instead
with a large codebase. It seemed to work for them, but I m certain we
couldn t have made it work for Launchpad.)
I can t speak too highly of Launchpad s test suite, much of which originated
before my time. Without a great deal of extensive coverage of all sorts of
interesting edge cases at both the unit and functional level, and a
corresponding culture of maintaining that test suite well when making new
changes, it would have been impossible to be anything like as confident of
the port as we were.
As part of the porting work, we split out a couple of substantial chunks of
the Launchpad codebase that could easily be decoupled from the core: its
Mailman integration and its
code import
worker. Both of these had substantial
dependencies with complex requirements for porting to Python 3, and
arranging to be able to do these separately on their own schedule was
absolutely worth it. Like disentangling balls of wool, any opportunity you
can take to make things less tightly-coupled is probably going to make it
easier to disentangle the rest. (I can see a tractable way forward to
porting the code import worker, so we may well get that done soon. Our
Mailman integration will need to be rewritten, though, since it currently
depends on the Python-2-only Mailman 2, and Mailman 3 has a different architecture.)
Python lessons
Our
database layer was already in pretty good
shape for a port, since at least the modern bits of its table modelling
interface were already strict about using Unicode for text columns. If you
have any kind of pervasive low-level framework like this, then making it be
pedantic at you in advance of a Python 3 port will probably incur much less
swearing in the long run, as you won t be trying to deal with quite so many
bytes/text issues at the same time as everything else.
Early in our port, we established a standard set of
__future__ imports
and started incrementally converting files over to them, mainly because we
weren t yet sure what else to do and it seemed likely to be helpful.
absolute_import was definitely reasonable (and not often a problem in our
code), and
print_function was annoying but necessary. In hindsight I m
not sure about
unicode_literals, though. For files that only deal with
bytes and text it was reasonable enough, but as I mentioned above there were
also a number of cases where we needed literals of the language s native
str type, i.e. bytes in Python 2 and text in Python 3: this was
particularly noticeable in
WSGI contexts, but also cropped up in
some other
surprising
places. We
generally either omitted
unicode_literals or used
six.ensure_str in such
cases, but it was definitely a bit awkward and maybe I should have listened
more to people telling me it might be a bad idea.
A lot of Launchpad s early tests used
doctest, mainly in the
style
where you have text files that interleave narrative commentary with
examples. The development team later reached consensus that this was best
avoided in most cases, but by then there were far too many doctests to
conveniently rewrite in some other form. Porting doctests to Python 3 is
really annoying. You run into all the little changes in how objects are
represented as text (particularly
u'...' versus
'...', but plenty of
other cases as well); you have next to no tools to do anything useful like
skipping individual bits of a doctest that don t apply; using
__future__
imports requires the rather obscure approach of adding the relevant names to
the doctest s globals in the relevant
DocFileSuite or
DocTestSuite;
dealing with many exception tracebacks requires something like
zope.testing.renormalizing;
and whatever code refactoring tools you re using probably don t work
properly. Basically, don t have done that. It did all turn out to be
tractable for us in the end, and I managed to avoid using much in the way of
fragile doctest extensions aside from the aforementioned
zope.testing.renormalizing, but it was not an enjoyable experience.
Regressions
I know of nine regressions that reached Launchpad s production systems as a
result of this porting work; of course there were various other regressions
caught by
CI or in manual testing. (Considering the size of this project, I
count it as a resounding success that there were only nine production
issues, and that for the most part we were able to fix them quickly.)
Equality testing of removed database objects
One of the things we had to do while porting to Python 3 was to
implement
the
__eq__,
__ne__, and
__hash__ special methods for all our database
objects. This was quite conceptually fiddly, because doing this requires
knowing each object s primary key, and that may not yet be available if
we ve created an object in Python but not yet flushed the actual
INSERT
statement to the database (most of our primary keys are auto-incrementing
sequences). We thus had to take care to flush pending
SQL statements in
such cases in order to ensure that we know the primary keys.
However, it s possible to have a problem at the other end of the object
lifecycle: that is, a Python object might still be reachable in memory even
though the underlying row has been
DELETEd from the database. In most
cases we don t keep removed objects around for obvious reasons, but it can
happen in caching code, and buildd-manager
crashed as a result (in
fact while it was still running on Python 2). We had to
take extra
care
to avoid this problem.
Debian imports crashed on non-
UTF-8 filenames
Python 2 has some
unfortunate
behaviour around passing
bytes or Unicode strings (depending on the platform) to
shutil.rmtree, and
the combination of some
porting
work
and a particular source package in Debian that contained a non-
UTF-8 file
name caused us to run into this. The
fix
was to ensure that the argument passed to
shutil.rmtree is a
str
regardless of Python version.
We d actually run into
something
similar
before: it s a subtle porting gotcha, since it s quite easy to end up
passing Unicode strings to
shutil.rmtree if you re in the process of
porting your code to Python 3, and you might easily not notice if the file
names in your tests are all encoded using
UTF-8.
lazr.restful ETags
We eventually got far enough along that we could switch one of our four
appserver machines (we have quite a number of other machines too, but the
appservers handle web and
API requests) to Python 3 and see what happened.
By this point our extensive test suite had shaken out the vast majority of
the things that could go wrong, but there was always going to be room for
some interesting edge cases.
One of the Ubuntu kernel team reported that they were seeing an increase in
412 Precondition
Failed errors in some
of their scripts that use our webservice
API. These can happen when you re
trying to modify an existing resource: the underlying protocol involves
sending an
If-Match header with the
ETag that the client thinks the
resource has, and if this doesn t match the
ETag that the server calculates
for the resource then the client has to refresh its copy of the resource and
try again. We initially thought that this might be legitimate since it can
happen in normal operation if you collide with another client making changes
to the same resource, but it soon became clear that something stranger was
going on: we were getting inconsistent
ETags for the same object even when
it was unchanged. Since we d recently switched a quarter of our appservers
to Python 3, that was a natural suspect.
Our
lazr.restful package provides the framework for our webservice
API,
and roughly speaking it generates
ETags by serializing objects into some
kind of canonical form and hashing the result. Unfortunately the
serialization was dependent on the Python version in a few ways, and in
particular it serialized lists of strings such as lists of bug tags
differently: Python 2 used
[u'foo', u'bar', u'baz'] where Python 3 used
['foo', 'bar', 'baz']. In
lazr.restful 1.0.3 we
switched to using
JSON
for this, removing the Python version dependency and ensuring consistent
behaviour between appservers.
Memory leaks
This problem took the longest to solve. We noticed fairly quickly from our
graphs that the appserver machine we d switched to Python 3 had a serious
memory leak. Our appservers had always been a bit leaky, but now it wasn t
so much a small hole that we can bail occasionally as the boat is sinking rapidly :
(Yes, this got in the way of working out what was going on with
ETags for
a while.)
I spent ages messing around with various attempts to fix this. Since only
a quarter of our appservers were affected, and we could get by on 75%
capacity for a while, it wasn t urgent but it was definitely annoying.
After spending some quality time with
objgraph, for
some time I thought
traceback reference
cycles
might be at fault, and I sent a number of fixes to various upstream projects
for those (e.g.
zope.pagetemplate).
Those didn t help the leaks much though, and after a while it became clear
to me that this couldn t be the sole problem: Python has a cyclic garbage
collector that will eventually collect reference cycles as long as there are
no strong references to any objects in them, although it might not happen
very quickly. Something else must be going on.
Debugging reference leaks in any non-trivial and long-running Python program
is extremely arduous, especially with ORMs that naturally tend to end up
with lots of cycles and caches. After a while I formed a hypothesis that
zope.server might be keeping a
strong reference to something, although I never managed to nail it down more
firmly than that. This was an attractive theory as we were already in the
process of migrating to
Gunicorn for
other reasons anyway, and Gunicorn also has a convenient
max_requests
setting that s good at mitigating memory leaks. Getting this all in place
took some time, but once we did we found that everything was much more stable:
This isn t completely satisfying as we never quite got to the bottom of the
leak itself, and it s entirely possible that we ve only papered over it
using
max_requests: I expect we ll gradually back off on how frequently we
restart workers over time to try to track this down. However,
pragmatically, it s no longer an operational concern.
Mirror prober
HTTPS proxy handling
After we switched our script servers to Python 3, we had several reports of
mirror probing
failures. (Launchpad
keeps lists of Ubuntu archive and image mirrors, and probes them every so
often to check that they re reasonably complete and up to date.) This only
affected
HTTPS mirrors when probed via a proxy server, support for which is
a relatively recent feature in Launchpad and involved some code that we
never managed to unit-test properly: of course this is exactly the code that
went wrong. Sadly I wasn t able to sort out that gap, but at least the
fix
was simple.
Non-
MIME-encoded email headers
As I mentioned above, there were substantial changes in the
email package
between Python 2 and 3, and indeed between minor versions of Python 3. Our
test coverage here is pretty good, but it s an area where it s very easy to
have gaps. We noticed that a script that processes incoming email was
crashing on messages with headers that were non-
ASCII but not
MIME-encoded (and
indeed then crashing again when it tried to send a notification of the
crash!). The only examples of these I looked at were spam, but we still
didn t want to crash on them.
The
fix
involved being somewhat more careful about both the handling of headers
returned by Python s email parser and the building of outgoing email
notifications. This seems to be working well so far, although I wouldn t be
surprised to find the odd other incorrect detail in this sort of area.
Failure to handle non-
ISO-8859-1
URL-encoded form input
Remember how I said that parsing
HTTP form data was thorny? After we
finished upgrading all our appservers to Python 3, people started reporting
that they
couldn t post Unicode comments to
bugs, which turned out
to be only if the attempt was made using JavaScript, and was because I
hadn t quite managed to get
URL-encoded form data working properly with
zope.publisher and
multipart. The current standard describes the
URL-encoded format for form data as
in many ways an aberrant
monstrosity ,
so this was no great surprise.
Part of the problem was some
very strange
choices in
zope.publisher dating back to 2004 or earlier, which I attempted to
clean
up and simplify.
The rest was that Python 2 s
urlparse.parse_qs unconditionally decodes
percent-encoded sequences as
ISO-8859-1 if they re passed in as part of a
Unicode string, so
multipart needs to
work around
this on Python 2.
I m still not completely confident that this is correct in all situations,
but at least now that we re on Python 3 everywhere the matrix of cases we
need to care about is smaller.
Inconsistent marshalling of Loggerhead s disk cache
We use
Loggerhead for providing web
browsing of Bazaar branches. When we upgraded one of its two servers to
Python 3, we immediately noticed that the one still on Python 2 was failing
to read back its revision information cache, which it stores in a database
on disk. (We noticed this because it caused a deployment to fail: when we
tried to roll out new code to the instance still on Python 2, Nagios checks
had already caused an incompatible cache to be written for one branch from
the Python 3 instance.)
This turned out to be a similar problem to the
pickle issue mentioned
above, except this one was with
marshal, which I didn t think to look for
because it s a relatively obscure module mostly used for internal purposes
by Python itself; I m not sure that Loggerhead should really be using it in
the first place. The fix was
relatively
straightforward,
complicated mainly by now needing to cope with throwing away unreadable
cache data.
Ironically, if we d just gone ahead and taken the nominally riskier path of
upgrading both servers at the same time, we might never have had a problem here.
Intermittent bzr failures
Finally, after we upgraded one of our two Bazaar codehosting servers to
Python 3, we had a
report of intermittent
bzr branch hangs. After some digging I found this in our logs:
Traceback (most recent call last):
...
File "/srv/bazaar.launchpad.net/production/codehosting1-rev-20124175fa98fcb4b43973265a1561174418f4bd/env/lib/python3.5/site-packages/twisted/conch/ssh/channel.py", line 136, in addWindowBytes
self.startWriting()
File "/srv/bazaar.launchpad.net/production/codehosting1-rev-20124175fa98fcb4b43973265a1561174418f4bd/env/lib/python3.5/site-packages/lazr/sshserver/session.py", line 88, in startWriting
resumeProducing()
File "/srv/bazaar.launchpad.net/production/codehosting1-rev-20124175fa98fcb4b43973265a1561174418f4bd/env/lib/python3.5/site-packages/twisted/internet/process.py", line 894, in resumeProducing
for p in self.pipes.itervalues():
builtins.AttributeError: 'dict' object has no attribute 'itervalues'
I d seen this before in our git hosting service: it was a bug in Twisted s
Python 3 port,
fixed after
20.3.0 but unfortunately after the last release that supported Python 2, so
we had to backport that patch. Using the same backport dealt with this.
Onwards!
Armadillo is a powerful and expressive C++ template library for linear algebra aiming towards a good balance between speed and ease of use with a syntax deliberately close to a Matlab. RcppArmadillo integrates this library with the R environment and language and is widely used by (currently) 912 other packages on CRAN.
This new release brings us Armadillo 10.7.0 released this morning by Conrad. Leading up to this were three runs of reverse dependencies the first of which uncovered the need for a small PR for 
For a customer project, I recently needed to take a closer look at best strategies of deploying Chromium settings to thrillions of client machines in a corporate network.
Unfortunately, the information on how to deploy site-wide Chromium
browser policies are a little scattered over the internet and the
intertwining of Chromium preferences and Chromium policies required
deeper introspection.
Here, I'd like to provide the result of that research, namely a list of references
that has been studied before setting up Chromium policies for the customer's proof-of-concept.
Difference between Preferences and Policies
Chromium can be controlled via preferences (mainly user preferences) and
administratively rolled-out policy files.
The difference between preferences and policies are explained here:
I have uploaded most of my Machine Learning lecture to YouTube.
The slides are in English, but the audio is in German.
Some very basic contents (e.g., a demo of standard k-means clustering) were left out from this advanced class,
and instead only a link to recordings from an earlier class were given. In this class, I wanted to focus
on the improved (accelerated) algorithms instead. These are not included here (yet).
I believe there are some contents covered in this class you will find nowhere else (yet).
The first unit is pretty long (I did not split it further yet). The later units are shorter recordings.
ML F1: Principles in Machine Learning
screenshot showing power consumption of server
