Once we've migrated away from svn.python.org for Windows build dependencies, there should be no reason to continue running svn.python.org.

The reason to keep it running is for all the existing people who are building from source. Unless you're porting all the old versions over and providing instructions for updating old sources? I wouldn't bother - just leave the server up for a few more years.

Ben probably knows - can we add a custom message to svn requests? Then we could start publicising an end of life date for it.

The box svn is running on is at risk of falling over any day. Plus with whatever ancient software it's running, it's surely a security risk to keep running an unmaintained svn instance. I certainly don't want to keep it running for years. I don't think we've ever advertised the svn windows deps as a service we provide the community.

Off the top of my head, I'm not sure how to announce that deprecation to svn client users. If you find something on the internets, I can put it on the server.

> I don't think we've ever advertised the svn windows deps as a service we provide the community.

Except in the dev guide :) ([after looking] well, linked from the dev guide https://github.com/python/cpython/blob/master/PCbuild/readme.txt#L230 )

Maybe we just need to pick a date and publicise it separately. Is this process PEP-worthy? Considering the amount of decisions we had to make regarding putting code in git, it might be worth writing up how it works in a brief PEP and including an EOL for svn.p.o

How far along are you with removing the svn.python.org dependency. Does  Jan 1 2018 sound okay for the end of svn.python.org?

I think a PEP is up to you. Doesn't seem like something that needs to be approved.

I could imagine people wanting to build old releases.  But we have Windows binaries up here:
https://www.python.org/ftp/python/

Therefore I don't think it's a problem to discontinue svn.python.org once new releases are migrated to github (but only once those new releases are out).

I just disabled viewvc on svn.python.org because of a security problem.

New changeset 51599e2bdd10ab77212a7cbb41a13ea70ee13da8 by Zachary Ware in branch 'master':
bpo-30450: Pull Windows dependencies from GitHub rather than svn (GH-1783)
https://github.com/python/cpython/commit/51599e2bdd10ab77212a7cbb41a13ea70ee13da8

New changeset cb8c0485ab1e3aeb4f64e76177aabbc6eba5a0ef by Zachary Ware in branch 'master':
bpo-30450: Add NEWS and whatsnew (GH-2236)
https://github.com/python/cpython/commit/cb8c0485ab1e3aeb4f64e76177aabbc6eba5a0ef

New changeset 04431c926a88be433203f89e3577addd0d41b602 by Zachary Ware in branch '3.6':
bpo-30450: Pull Windows dependencies from GitHub rather than svn (GH-1783) (GH-2237)
https://github.com/python/cpython/commit/04431c926a88be433203f89e3577addd0d41b602

The usage text in build.bat still mentions that svn.exe is required for the '-e' flag. I think it should be updated to reflect the changes made.

Zach, I just spotted this build log, that appears to have failed to download nuget.exe: http://buildbot.python.org/all/builders/AMD64%20Windows7%20SP1%203.6/builds/339/steps/compile/logs/stdio

Guessing the bot is only succeeding because it currently has all the dependencies. Maybe we should consider just checking in nuget.exe?

Looks like that error is coming from the fact that the Powershell on that buildbot is outdated. As the documentation notes:

https://msdn.microsoft.com/en-us/powershell/reference/5.1/microsoft.powershell.utility/invoke-webrequest

> This cmdlet was introduced in Windows PowerShell 3.0.


Before running the web request command maybe do a `powershell.exe -Command $PSVersionTable.PSVersion` in order to debug this issue on any other machines?

The buildbot failures currently being blamed on issue30916 are actually due to this issue - we actually need Python 3.6 on these machines in order to download the new externals, and they don't have it and we can't get it via Powershell.

My PR adds a small script to use requests.get/urlretrieve with any version of Python to get nuget.exe.

Theoretically, we *could* use this to download everything, but I'd rather minimize our exposure to insecure downloads for machines that don't already have Python 3.6 available.

Compilation failed on:

* http://buildbot.python.org/all/builders/x86%20Windows7%203.x/builds/885
* http://buildbot.python.org/all/builders/AMD64%20Windows7%20SP1%203.x/builds/750

New changeset 588836d3e646c2bcb3473cda7c5f6a1e0ff2c2e9 by Steve Dower in branch 'master':
bpo-30450: Adds alternate download approach for nuget.exe (#2737)
https://github.com/python/cpython/commit/588836d3e646c2bcb3473cda7c5f6a1e0ff2c2e9

Compilation still fails, as some of the buildbots don't have py.exe either.

At this stage, our choices are either to fix the buildbots, most easily by copying py.exe into PATH, or to check nuget.exe (4MB) into the source repository. (I guess we can cover a few more % of cases by checking in py.exe, but that still won't get us 100%)

Any preferences?

If the dependencies are in git, why don't you use "git clone" instead of trying a HTTP fetch?

For example "git clone --depth 1 https://github.com/python/cpython-bin-deps --branch openssl-bin-1.0.2k myopenssldir"

We can't assume git.exe is available either (build from sdist/hggit/.zip), though I guess we can also use it as a fallback. At least on the buildbots it'll be there.

I think it's reasonable to mandate the presence of a git install to fetch CPython externals.  After all we used to mandate the presence of a svn install...

> We can't assume git.exe is available either

How do you get CPython source code if git is not available? Also using
HTTP to get a tarball? I think that it's ok to require git to build
CPython on Windows. Previously, we required: svn, perl and git :-)

> we used to mandate the presence of a svn install

And we regretted that so much that we changed away from it :)

When discussing this changeover plan, Zach and I decided we needed a fallback requiring only OS dependencies. In this case, the Powershell dependency fails on Windows 7 since it does not have the Invoke-WebRequest command (unless you've been installing all your updates).

We decided against using git first because then the case that is more reliable (download and extract a .zip file from a URL) would go unused/untested in many cases.

We also use the potentially installed python.exe in other places in the build, so it's far from a waste to grab it, and we already required *any* Python dependency to do a full build (docs and/or installer), so it isn't really adding anything there. The bit that was overlooked was the PowerShell on old systems limitation, and the absence of py.exe when you don't have Python 3 installed.

> How do you get CPython source code if git is not available?

Go to GitHub and click the "Download" button, or select any sdist from our downloads page, or use Mercurial+hggit, or Dulwich directly, or a mirror/fork on any VCS you like. There are plenty of ways that don't require having git.exe on your PATH.

> And we regretted that so much that we changed away from it :)

I'm not sure what your reasons were but from my POV (as someone who used to build on Windows from time to time), the two problems were:

1) svn was only used for this while we had migrated to hg

2) svn is darn slow, which neither hg nor git (especially with shallow clones, see "git clone --depth 1") are

I don't think the annoyance itself of having to install git is much to worry about, since 1) it's a one time cost 2) it's quite easy 3) many developers or packagers are bound to already use git for other projects.

Ah ok, I see, I was focused on buildbots and other CIs, sorry.

> Maybe we should consider just checking in nuget.exe?

Where? In Python source code? I don't want to how to maintain yet another dependencies :-( I prefer to try a few tools, or explain how to get them. And write good documentation how to build Python on Windows ;-)

Unfortunately, "install <anything>" is often a blocking requirement for many developers (typically the ones who get paid well to do this), so I'd rather have no requirements.

Currently we achieve this if PowerShell is up to date or Python+py.exe are installed, both of which are likely even in the "I can't install anything" case. Soon we'll also succeed here if git is installed and on PATH, which likely covers a few more % of users (but especially our Win7 buildbots). I feel confident we've covered 99% of cases transparently without adding the need for more documentation - still just get_externals.bat && build.bat to build.

Once Travis has finished its build I'll merge and we'll see if the buildbots like it.

New changeset efa26bcd5085279fc4e9ae96d052272a5214c2bd by Steve Dower in branch 'master':
bpo-30450: Fall back to git.exe if no Python is found. (#2739)
https://github.com/python/cpython/commit/efa26bcd5085279fc4e9ae96d052272a5214c2bd

The buildbots have already successfully built, so I'm declaring this and issue30916 resolved.

Or, use the preexisting convention of the HOST_PYTHON envvar that was used prior to the recent merged PRs

> use the preexisting convention of the HOST_PYTHON envvar that was used prior to the recent merged PRs

Is that an actual convention? I didn't see any other references, so I figured Zach had made it up.

I think that's used on Unix when cross-compiling, but I don't know about Windows...

> In this case, the Powershell dependency fails on Windows 7 since it does not have the Invoke-WebRequest command (unless you've been installing all your updates).

Just to note, PowerShell must be updated *manually* (at least on Win7).  Plus you need to know what to update to get it. In this case, the Windows Management Framework:

https://www.microsoft.com/en-us/download/details.aspx?id=40855

I've installed this on my buildbot apparently concurrently with this conversation.  Personally, I'd lean toward recommending this update with Windows 7 machines as it stays closer to "naked" OS + Python checkout for building a new Python. (Although the latest merge kinda makes this a mute point).

> Is that an actual convention? I didn't see any other references, so I figured Zach had made it up.

It has existed in the Windows build files since 2.5, when x64
supported was initially added by MvL.

I just got a download failure on an AppVeyor build:
https://ci.appveyor.com/project/python/cpython/build/3.7.0a0.4571

Downloading nuget...
Invoke-WebRequest : The operation has timed out.
At line:1 char:1
+ Invoke-WebRequest https://aka.ms/nugetclidl -OutFile 'C:\projects\cpy ...
+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    + CategoryInfo          : InvalidOperation: (System.Net.HttpWebRequest:Htt 
   pWebRequest) [Invoke-WebRequest], WebException
    + FullyQualifiedErrorId : WebCmdletWebResponseException,Microsoft.PowerShe 
   ll.Commands.InvokeWebRequestCommand
 
  File "<string>", line 1
    C:\projects\cpython\PCbuild\\urlretrieve.py
     ^
SyntaxError: invalid syntax
Installing Python via nuget...

Hmm... looks like there's another way to fail I didn't account for that goes down a different path in the batch file. I'll try being less clever.

Yeouch, it's been causing AppVeyor builds to get stuck, and then they time out after an hour.

Luckily mine is up next in a minute or so, so I'll try to merge quickly and avoid anyone else getting held up.

I also restored HOST_PYTHON in the PR, as a fallback for "where do I find a python.exe I can download stuff with" - specific version doesn't matter, as long as it has requests or urlretrieve.

Ah, there's also a nuget.org outage affecting AppVeyor (but not me because I'm in Italy and so connecting to the EU mirror automatically) - https://appveyor.statuspage.io/

When that recovers it should be fine again. The workaround for those without py.exe is to set PYTHON (or, post-PR, HOST_PYTHON) to a Python 3.6 or later executable.

New changeset 5feda33a35d9413e2073411b848dc49d94c57497 by Steve Dower in branch 'master':
bpo-30450: Fix logic for retrying nuget.exe download (#2744)
https://github.com/python/cpython/commit/5feda33a35d9413e2073411b848dc49d94c57497

Zach - did this get backported to 2.7 or 3.5? It doesn't look like it.

New changeset e99d3a52a50b3f836fb9fb88f317aacddd494858 by Steve Dower in branch '3.6':
[3.6] bpo-30450: Improved logic for obtaining dependencies (#2751)
https://github.com/python/cpython/commit/e99d3a52a50b3f836fb9fb88f317aacddd494858

Has this actually been backported to 3.6? It looks like a lot of Tcl and OpenSSL stuff is being compiled on AppVeyor yet:
https://ci.appveyor.com/project/python/cpython/build/3.6.1+.5134

The prebuilt binaries have not, and the OpenSSL change has functional implications and should not be (unless someone wants to make the security justification, but right now I think it's a theoretical rather than an actual benefit). The TclTk build could be, but I haven't gone to the effort to split it up.

There's also a separate issue for that part - this bug is for the location of sources.

New changeset 986b7ffc650919b3022ccaa458a843bb8a95d2bd by Zachary Ware in branch '2.7':
[2.7] bpo-30450: Pull Windows dependencies from GitHub rather than SVN (GH-1783) (GH-3306)
https://github.com/python/cpython/commit/986b7ffc650919b3022ccaa458a843bb8a95d2bd

> New changeset 986b7ffc650919b3022ccaa458a843bb8a95d2bd by Zachary Ware in branch '2.7':
> [2.7] bpo-30450: Pull Windows dependencies from GitHub rather than SVN (GH-1783) (GH-3306)

This change broke Python 2.7 compilation on Windows XP.

Pending fix: PR 3330.

> New changeset 986b7ffc650919b3022ccaa458a843bb8a95d2bd by Zachary Ware in branch '2.7':
> [2.7] bpo-30450: Pull Windows dependencies from GitHub rather than SVN (GH-1783) (GH-3306)

This change also broke compilation on AMD64 Windows10 2.7 buildbot:
http://buildbot.python.org/all/builders/AMD64%20Windows10%202.7/builds/279



"D:\buildarea\2.7.bolen-windows10\build\PCbuild\pcbuild.proj" (Build target) (1) ->
"D:\buildarea\2.7.bolen-windows10\build\PCbuild\pythoncore.vcxproj" (Build target) (2) ->
(Manifest target) -> 
  C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\Microsoft.CppCommon.targets(574,5): error MSB6006: "mt.exe" exited with code -1073741819. [D:\buildarea\2.7.bolen-windows10\build\PCbuild\pythoncore.vcxproj]

    260 Warning(s)
    1 Error(s)

New changeset 8905fb831cf7c400c479b79bb2f90bfbe9c71337 by Zachary Ware in branch '2.7':
bpo-30450: Don't use where, XP doesn't have it (GH-3348)
https://github.com/python/cpython/commit/8905fb831cf7c400c479b79bb2f90bfbe9c71337

New changeset 1911cf3dd2ae67d600c166ba52872fdcf3e85824 by Zachary Ware in branch '2.7':
[2.7] bpo-30450: Fall back on the old env.bat (GH-3443)
https://github.com/python/cpython/commit/1911cf3dd2ae67d600c166ba52872fdcf3e85824

> This change also broke compilation on AMD64 Windows10 2.7 buildbot:
> http://buildbot.python.org/all/builders/AMD64%20Windows10%202.7/builds/279
> (...) error MSB6006: "mt.exe" exited with code -1073741819. (...)

I created the bug bpo-31430 to track this regression.

This has been backported and working for nearly 3 months now; closing.

As for whether to continue operating svn.python.org now that we're officially away from it: that can be determined elsewhere, probably by the box dying and deciding for us :)

Noooo! I loved Subversion so much! ... just kidding. Thanks for getting rid of Subversion.

It became much easier to build CPython on Windows nowadays!