In the summer of 2020, I worked with the contributors of pip, trying to improve the networking performance of the package manager. Admittedly, at the end of the internship period, the benchmark said otherwise; though I really hope the clean-up and minor fixes I happened to be doing to the codebase over the summer, in addition to the implementation of parallel utils and lazy wheel, might actually help the project.

Personally, I learned a lot: not just about Python packaging and networking stuff, but also on how to work with others. I am really grateful to @pradyunsg (my mentor), @chrahunt, @uranusjr, @pfmoore, @brainwane, @sbidoul, @xavfernandez, @webknjaz, @jaraco, @deveshks, @gutsytechster, @dholth, @dstufft, @cosmicexplorer and @ofek. While this feels like a long shout-out list, it really isn't. These people are the maintainers, the contributors of pip and/or other Python packaging projects, and more importantly, they have been more than helpful, encouraging and patient to me throughout my every activities, showing me the way when I was lost, fixing me when I was wrong, putting up with my carelessness and showing me support across different social media.

To best serve the community, below I have tried my best to document what I have done, how I've done it and why I've done it for over the last three months. At the time of writing, some work is still in progress, so these also serve as a reference point for myself and others to reason about decisions in relevant topics.

The Main Story

The storyline can be divided into the following four main acts.

Act One: Parallelization Utilities

In this first act, I ensured the portibility of parallelization measures for later use in the final act. Multithreading and multiprocessing map were properly fellback on platforms without full support.

• GH-8320: Add utilities for parallelization (close GH-8169)

• GH-8538: Make utils.parallel tests tear down properly

• GH-8504: Parallelize pip list --outdated and --uptodate (using GH-8320)

Act Two: Lazy Wheels

As proposed by @cosmicexplorer in GH-7819, it is possible to only download a portion of a wheel to obtain metadata during dependency resolution. Not only that this would reduce the total amount of data to be transmitted over the network in case the resolver needs to perform heavy backtracking, but also it would create a synchronization point at the end of the resolution progress where parallel downloading can be applied to the needed wheels (some wheels solely serve their metadata during dependency backtracking and are not needed by the users).

• GH-8467: Add utitlity to lazily acquire wheel metadata over HTTP

• GH-8584: Revise lazy wheel and its tests

• GH-8681: Make range requests closer to chunk size (help GH-8670)

• GH-8716 and GH-8730: Disable caching for range requests

Act Three: Late Downloading

During this act, the main works were refactoring to integrate the lazy wheel into pip's codebase and clean up the way for download parallelization.

• GH-8411: Refactor operations.prepare.prepare_linked_requirement

• GH-8629: Abstract away AbstractDistribution in higher-level resolver code

• GH-8442, GH-8532 and GH-8588 (later reworked by @chrahunt in GH-8685): Use lazy wheel to obtain dependency information for the new resolver

• GH-8743: Test hash checking for fast-deps

• GH-8804: Check download directory before making range requests

Act Four: Batch Downloading in Parallel

The final act is mostly about the UI of the parallel download. My work involved around how the progress should be displayed and how other relevant information should be reported to the users.

• GH-8710: Revise method fetching metadata using lazy wheels

• GH-8722: Dedent late download logs (fix GH-8721)

• GH-8737: Add a hook for batch downloading

• GH-8771: Parallelize wheel download

The Side Quests

In order to keep the wheel turning (no pun intended) and avoid wasting time waiting for the pull requests above to be reviewed, I decided to create even more PRs (as I am typing this, many of the patches listed below are nowhere near being merged).

• GH-7878: Fail early when install path is not writable

• GH-7928: Fix rst syntax in Getting Started guide

• GH-7988: Fix tabulate col size in case of empty cell

• GH-8137: Add subcommand alias mechanism

• GH-8143: Make mypy happy with beta release automation

• GH-8248: Fix typo and simplify ireq call

• GH-8332: Add license requirement to _vendor/README.rst

• GH-8423: Nitpick logging calls

• GH-8435: Use str.format style in logging calls

• GH-8456: Lint src/pip/_vendor/README.rst

• GH-8568: Declare constants in configuration.py as such

• GH-8571: Clean up Configuration.unset_value and nit __init__

• GH-8578: Allow verbose/quiet level to be specified via config files and environment variables

• GH-8599: Replace tabs by spaces for consistency

• GH-8614: Use monkeypatch.setenv to mock environment variables

• GH-8674: Fix tests/functional/test_install_check.py, when run with new resolver

• GH-8692: Make assertion failure give better message

• GH-8709: List downloaded distributions before exiting (fix GH-8696)

• GH-8759: Allow py2 deprecation warning from setuptools

• GH-8766: Use the new resolver for test requirements

• GH-8790: Mark tests using remote svn and hg as xfail

• GH-8795: Reformat a few spots in user guide

The Plot Summary

Every Monday throughout the Summer of Code, I summarized what I had done in the week before in the form of either a short blog or an (even shorter) check-in. These write-ups often contain handfuls of popular culture references and was originally hosted on Python GSoC.

• First Check-In

• Unexpected Things When You're Expecting

• Second Check-In

• The Wonderful Wizard of O'zip

• Third Check-In

• I'm Not Drowning On My Own

• Fourth Check-In

• I've Walked 500 Miles…

• Fifth Check-In

• Sorting Things Out

• Sixth Check-In

• Parallelizing Wheel Downloads

• Final Check-In

• Outro

Steamed fish was amazing, matter of fact
Let me get some jerk chicken to go
Grabbed me one of them lemon pie theories
And let me get some of them benchmarks you theories too

The Look

At the time of writing, implementation-wise parallel download is ready:

Does this mean I've finished everything just-in-time? This sounds to good to be true! And how does it perform? Welp...

The Benchmark

Here comes the bad news: under a decent connection to the package index, using fast-deps does not make pip faster. For best comparison, I will time pip download on the following cases:

Average Distribution

For convenience purposes, let's refer to the commands to be used as follows

$ pip --no-cache-dir download {requirement}  # legacy-resolver
$ pip --use-feature=2020-resolver \
   --no-cache-dir download {requirement}  # 2020-resolver
$ pip --use-feature=2020-resolver --use-feature=fast-deps \
   --no-cache-dir download {requirement}  # fast-deps

In the first test, I used axuy and obtained the following results

Funny enough, running pip download with fast-deps in a directory with downloaded files already took around 7-8 seconds. This is because to lazily download a wheel, pip has to make many requests which are apparently more expensive than actual data transmission on my network.

Large Distribution

In this test, I used TensorFlow as the requirement and obtained the following figures:

Distribution with Conflicting Dependencies

Some requirement that will trigger a decent amount of backtracking by the current implementation of the new resolver oslo-utils==1.4.0:

What Now?

I don't know, to be honest. At this point I'm feeling I've failed my own (and that of other stakeholders of pip) expectation and wasted the time and effort of pip's maintainers reviewing dozens of PRs I've made in the last three months.

On the bright side, this has been an opportunity for me to explore the codebase of package manager and discovered various edge cases where the new resolver has yet to cover (e.g. I've just noticed that pip download would save to-be-discarded distributions, I'll file an issue on that soon). Plus I got to know many new and cool people and idea, which make me a more helpful individual to work on Python packaging in the future, I hope.

And now it's clear as this promise
That we're making
Two progress bars into one

Hello there! It has been raining a lot lately and some mosquito has given me the Dengue fever today. To whoever reading this, I hope it would never happen to you.

Download Parallelization

I've been working on pip's download parallelization for quite a while now. As distribution download in pip was modeled as a lazily evaluated iterable of chunks, parallelizing such procedure is as simple as submitting routines that write files to disk to a worker pool.

Or at least that is what I thought.

Progress Reporting UI

pip is currently using customly defined progress reporting classes, which was not designed to working with multithreading code. Firstly, I want to try using these instead of defining separate UI for multithreaded progresses. As they use system signals for termination, one must the progress bars has to be running the main thread. Or sort of.

Since the progress bars are designed as iterators, I realized that we can call next on them. So quickly, I throw in some queues and locks, and prototyped the first working implementation of progress synchronization.

Performance Issues

Welp, I only said that it works, but I didn't mention the performance, which is terrible. I am pretty sure that the slow down is with the synchronization, since the map_multithread call doesn't seem to trigger anything that may introduce any sort of blocking.

This seems like a lot of fun, and I hope I'll get better tomorrow to continue playing with it!

Hi! I really hope that everyone reading this is still doing okay, and if that isn't the case, I wish you a good day!

pip 20.2 Released!

Last Wednesday, pip 20.2 was released, delivering the 2020-resolver as well as many other improvements! I was lucky to be able to get the fast-deps feature to be included as part of the release. A brief description of this experimental feature as well as testing instruction can be found on Python Discuss.

The public exposure of the feature also remind me of some further optimization to make on the lazy wheel. Hopefully without download parallelization it would not be too slow to put off testing by concerned users of pip.

Preparation for Download Parallelization

As of this moment, we already have:

• Multithreading pool fallback working

• An opt-in to use lazy wheel to optain dependency information, and thus getting a list of wheels at the end of resolution ready to be downloaded together

What's left is only to interject a parallel download somewhere after the dependency resolution step. Still, this struggles me way more than I've ever imagined. I got so stuck that I had to give myself a day off in the middle of the week (and study some Rust), then I came up with something what was agreed upon as difficult to maintain.

Indeed, a large part of this is my fault, for not communicating the design thoroughly with pip's maintainers and not carefully noting stuff down during (verbal) discussions with my mentor. Thankfully Chris Hunt came to the rescue and did a refactoring that will make my future work much easier and cleaner.

... and I would walk 500 more
Just to be the man who walks a thousand miles
To fall down at your door

The Main Road

Hi, have you met fast-deps? It's (going to be) the name of pip's experimental feature that may improve the speed of dependency resolution of the new resolver. By avoid downloading whole wheels to just obtain metadata, it is especially helpful when pip has to do heavy backtracking to resolve conflicts.

Thanks to Chris Hunt's review on GH-8537, my mentor Pradyun Gedam and I worked out a less hacky approach to inteject the call to lazy wheel during the resolution process. A new PR GH-8588 was filed to implement it—I could have just worked on top of the old PR and rebased, but my git skill is far from gud enuff to confidently do it.

Testing this one has been a lot of fun though. At first, integration tests were added as a rerun of the tests for the new resolver, with an additional flag to use feature fast-deps. It indeed made me feel guilty towards Travis, who has to work around 30 minutes more every run. Per Chris Hunt's suggestion, in the new PR, I instead write a few functional tests for the area relating the most to the feature, namely pip's subcommands wheel, download and install.

It was also suggested that a mock server with HTTP range requests support might be better (in term of performance and reliablilty) than for testing. However, I have yet to be able to make Werkzeug do it.

Why did I say I'm half way there? With the parallel utilities merged and a way to quickly get the list of distribution to be downloaded being really close, what left is only to figure out a way to properly download them in parallel. With no distribution to be added during the download progress, the model of this will fit very well with the architecture in my original proposal. A batch downloader can be implemented to track the progress of each download and thus report them cleanly as e.g. progress bar or percentage. This is the part I am second-most excited about of my GSoC project this summer (after the synchronization of downloads written in my proposal, which was then superseded by fast-deps) and I can't wait to do it!

The Side Quests

As usual, I make sure that I complete every side quest I see during the journey:

• GH-8568: Declare constants in configuration.py as such

• GH-8571: Clean up Configuration.unset_value and nit the class' __init__

• GH-8578: Allow verbose/quite level to be specified via config file and env var

• GH-8599: Replace tabs by spaces for consistency

Snap Back to Reality

A bit about me, I actually walked 500 meters earlier today to a bank and walked 500 more to another to prepare my Visa card for purchasing the upcoming PinePhone prototype. It's one of the first smartphones to fully support a GNU/Linux distribution, where one can run desktop apps (including proper terminals) as well as traditional services like SSH, HTTP server and IPFS node because why not? Just a few hours ago, I pre-ordered the postmarketOS community edition with additional hardware for convergence.

If you did not come here for a PinePhone ad, please take my apologies though d-; and to ones reading this, I hope you all can become the person who walks a thousand miles to fall down at the door opening to all what you ever wished for!

Cold Water

Hello there! My schoolyear is coming to an end, with some final assignments and group projects left to be done. I for sure underestimated the workload of these and in the last (and probably next) few days I'm drowning in work trying to meet my deadlines.

One project that might be remotely relevant is cheese-shop, which tries to manage the metadata of packages from the real Cheese Shop. Other than that, schoolwork is draining a lot of my time and I can't remember the last time I came up with something new for my GSoC project )-;

Warm Water

On the bright side, I received a lot of help and encouragement from contributors and stakeholders of pip. In the last week alone, I had five pull requests merged:

• GH-8332: Add license requirement to _vendor/README.rst

• GH-8320: Add utilities for parallelization

• GH-8504: Parallelize pip list --outdated and --uptodate

• GH-8411: Refactor operations.prepare.prepare_linked_requirement

• GH-8467: Add utitlity to lazily acquire wheel metadata over HTTP

In addition to helping me getting my PRs merged, my mentor Pradyun Gedam also gave me my first official feedback, including what I'm doing right (and wrong too!) and what I should keep doing to increase the chance of the project being successful.

GH-7819's roadmap (Danny McClanahan's discoveries and works on lazy wheels) is being closely tracked by hatch's maintainter Ofek Lev, which really makes me proud and warms my heart, that what I'm helping build is actually needed by the community!

Learning How To Swim

With GH-8467 and GH-8530 merged, I'm now working on GH-8532 which aims to roll out the lazy wheel as the way to obtain dependency information via the CLI flag --use-feature=lazy-wheel.

GH-8532 was failing initially, despite being relatively trivial and that the commit it used to base on was passing. Surprisingly, after rebasing it on top of GH-8530, it suddenly became green mysteriously. After the first (early) review, I was able to iterate on my earlier code, which used the ambiguous exception RuntimeError.

The rest to be done is just adding some functional tests (I'm pretty sure this will be either overwhelming or underwhelming) to make sure that the command-line flag is working correctly. Hopefully this can make it into the beta of the upcoming release this month.

In other news, I've also submitted a patch improving the tests for the parallelization utilities, which was really messy as I wrote them. Better late than never!

Metaphors aside, I actually can't swim d-:

Diving Plan

After GH-8532, I think I'll try to parallelize downloads of wheels that are lazily fetched only for metadata. By the current implementation of the new resolver, for pip install, this can be injected directly between the resolution and build/installation process.

Never give up... No one knows what's going to happen next.

Preface

Greetings and best wishes! I had a lot of fun during the last week, although admittedly nothing was really finished. In summary, these are the works I carried out in the last seven days:

• Finilizing utilities for parallelization

• Continuing experimenting on using lazy wheels or dependency resolution

• Polishing up the patch refactoring operations.prepare.prepare_linked_requirement

• Adding flake8-logging-format to the linter

• Splitting the linting patch from the PR adding the license requirement to vendor README

The multiprocessing[.dummy] wrapper

Yes, you read it right, this is the same section as last fortnight's blog. My mentor Pradyun Gedam gave me a green light to have GH-8411 merged without support for Python 2 and the non-lazy map variant, which turns out to be troublesome for multithreading.

The tests still needs to pass of course and the flaky tests (see failing tests over Azure Pipeline in the past) really gave me a panic attack earlier today. We probably need to mark them as xfail or investigate why they are undeterministic specifically on Azure, but the real reason I was all caught up and confused was that the unit tests I added mess with the cached imports and as pip's tests are run in parallel, who knows what it might affect. I was so relieved to not discover any new set of tests made flaky by ones I'm trying to add!

The file-like object mapping ZIP over HTTP

This is where the fun starts. Before we dive in, let's recall some background information on this. As discovered by Danny McClanahan in GH-7819, it is possible to only download a potion of a wheel and it's still valid for pip to get the distribution's metadata. In the same thread, Daniel Holth suggested that one may use HTTP range requests to specifically ask for the tail of the wheel, where the ZIP's central directory record as well as where usually dist-info (the directory containing METADATA) can be found.

Well, usually. While PEP 427 does indeed recommend

Archivers are encouraged to place the .dist-info files physically at the end of the archive. This enables some potentially interesting ZIP tricks including the ability to amend the metadata without rewriting the entire archive.

one of the mentioned tricks is adding shared libraries to wheels of extension modules (using e.g. auditwheel or delocate). Thus for non-pure Python wheels, it is unlikely that the metadata lie in the last few megabytes. Ignoring source distributions is bad enough, we can't afford making an optimization that doesn't work for extension modules, which are still an integral part of the Python ecosystem )-:

But hey, the ZIP's directory record is warrantied to be at the end of the file! Couldn't we do something about that? The short answer is yes. The long answer is, well, yessssssss! That, plus magic provided by most operating systems, this is what we figured out:

1. We can download a realatively small chunk at the end of the wheel until it is recognizable as a valid ZIP file.

2. In order for the end of the archive to actually appear as the end to zipfile, we feed to it an object with seek and read defined. As navigating to the rear of the file is performed by calling seek with relative offset and whence=SEEK_END (see man 3 fseek for more details), we are completely able to make the wheels in the cloud to behave as if it were available locally.

   

3. For large wheels, it is better to store them in hard disks instead of memory. For smaller ones, it is also preferable to store it as a file to avoid (error-prony and often not really efficient) manual tracking and joining of downloaded segments. We only use a small potion of the wheel, however just in case one is wonderring, we have very little control over when tempfile.SpooledTemporaryFile rolls over, so the memory-disk hybrid is not exactly working as expected.

4. With all these in mind, all we have to do is to define an intermediate object check for local availability and download if needed on calls to read, to lazily provide the data over HTTP and reduce execution time.

The only theoretical challenge left is to keep track of downloaded intervals, which I finally figured out after a few trials and errors. The code was submitted as a pull request to pip at GH-8467. A more modern (read: Python 3-only) variant was packaged and uploaded to PyPI under the name of lazip_. I am unaware of any use case for it outside of pip, but it's certainly fun to play with d-:

What's next?

I have been falling short of getting the PRs mention above merged for quite a while. With pip's next beta coming really soon, I have to somehow make the patches reach a certain standard and enough attention to be part of the pre-release—beta-testing would greatly help the success of the GSoC project. To other GSoC students and mentors reading this, I also hope your projects to turn out successful!

Hi everyone, I hope that you are all doing well and wishes you all good health! The last week has not been really kind to me with a decent amount of academic pressure (my school year is lasting until early Jully). It would be bold to say that I have spent 10 hours working on my GSoC project since the last check-in, let alone the 30 hours per week requirement. That being said, there were still some discoveries that I wish to share.

The multiprocessing[.dummy] wrapper

Most of the time I spent was to finalize the multi{processing,threading} wrapper for map function that submit tasks to the worker pool. To my surprise, it is rather difficult to write something that is not only portable but also easy to read and test.

By the latest commit, I realized the following:

1. The multiprocessing module was not designed for the implementation details to be abstracted away entirely. For example, the lazy map's could be really slow without specifying suitable chunk size (to cut the input iterable and distribute them to workers in the pool). By suitable, I mean only an order smaller than the input. This defeats half of the purpose of making it lazy: allowing the input to be evaluated lazily. Luckily, in the use case I'm aiming for, the length of the iterable argument is small and the laziness is only needed for the output (to pipeline download and installation).

2. Mocking import for testing purposes can never be pretty. One reason is that we (Python users) have very little control over the calls of import statements and its lower-level implementation __import__. In order to properly patch this built-in function, unlike for others of the same group, we have to monkeypatch the name from builtins (or __builtins__ under Python 2) instead of the module that import stuff. Furthermore, because of the special namespacing, to avoid infinite recursion we need to alias the function to a different name for fallback.

3. To add to the problem, multiprocessing lazily imports the fragile module during pools creation. Since the failure is platform-specific (the lack of sem_open), it was decided to check upon the import of the pip's module. Although the behavior is easier to reason in human language, testing it requires invalidating cached import and re-import the wrapper module.

4. Last but not least, I now understand the pain of keeping Python 2 compatibility that many package maintainers still need to deal with everyday (although Python 2 has reached its end-of-life, pip, for example, will still support it for another year).

The change in direction

Since last week, my mentor Pradyun Gedam and I set up weekly real-time meeting (a fancy term for video/audio chat in the worldwide quarantine era) for the entire GSoC period. During the last session, we decided to put parallelization of download during resolution on hold, in favor of a more beneficial goal: partially download the wheels during dependency resolution.

As discussed by Danny McClanahan and the maintainers of pip, it is feasible to only download a few kB of a wheel to obtain enough metadata for the resolution of dependency. While this is only applicable to wheels (i.e. prebuilt packages), other packaging format only make up less than 20% of the downloads (at least on PyPI), and the figure is much less for the most popular packages. Therefore, this optimization alone could make the upcoming backtracking resolver's performance par with the legacy one.

During the last few years, there has been a lot of effort being poured into replacing pip's current resolver that is unable to resolve conflicts. While its correctness will be ensured by some of the most talented and hard-working developers in the Python packaging community, from the users' point of view, it would be better to have its performance not lagging behind the old one. Aside from the increase in CPU cycles for more rigorous resolution, more I/O, especially networking operations is expected to be performed. This is due to the lack of a standard and efficient way to acquire the metadata. Therefore, unlike most package managers we are familiar with, pip has to fetch (and possibly build) the packages solely for dependency informations.

Fortunately, PEP 427#recommended-archiver-features recommends package builders to place the metadata at the end of the archive. This allows the resolver to only fetch the last few kB using HTTP range requests_ for the relevant information. Simply appending Range: bytes=-8000 to the request header in pip._internal.network.download makes the resolution process lightning fast. Of course this breaks the installation but I am confident that it is not difficult to implement this optimization cleanly.

One drawback of this optimization is the compatibility. Not every Python package index support range requests, and it is not possible to verify the partial wheel. While the first case is unavoidable, for the other, hashes checking is usually used for pinned/locked-version requirements, thus no backtracking is done during dependency resolution.

Either way, before installation, the packages selected by the resolver can be downloaded in parallel. This warranties a larger crowd of packages, compared to parallelization during resolution, where the number of downloads can be as low as one during trail of different versions of the same package.

Unfortunately, I have not been able to do much other than a minor clean up. I am looking forward to accomplishing more this week and seeing what this path will lead us too! At the moment, I am happy that I'm able to meet the blog deadline, at least in UTC!