Sage Release Management

These are some notes on various aspects of the Sage release process. It is also worth reading Eric S. Raymond's article Software Release Practice HOWTO.

sage -merge/sage-apply-ticket

This is a new script written by Craig Citro and Nick Alexander. It tries to automate much of the process of downloading patches from trac, merging them, and running doctests. Here's a quick rundown of usage:

• sage -merge -c or sage -merge --candidates

  • List all candidates for merging, i.e. all trac tickets with positive review.

• sage -merge <ticket_number> <options>

  • Download patches from trac for the given ticket number, merge them, run tests, and report the results.

• sage -merge -a <options> or sage -merge --all <options>

  • For each ticket on trac with a positive review, download the patches, apply them, and run doctests. At the end, report which tickets passed, which failed, and which didn't have any files to doctest (as is commonly the case with tickets for new spkg's).

Notes

• This script makes heavy use of the queues extension (mq) to Mercurial. It is recommended that you understand how mq works and know how to use it for managing patches. This is especially relevant if you consider using the merge script for developing Sage or managing a release. For an introductory tutorial on mq, see the Mq Tutorial from Selenic. Bryan O'Sullivan's book "Mercurial: The Definitive Guide" is a handy reference on using Mercurial for everyday development. Chapters 12 and 13 carefully explain how to use mq to manage patches. Appendix B of the book is a reference on features of mq.

• Currently, the merge script requires that the patch queue be empty to start. After any test, it will pop all changes, regardless of outcome, unless an option (as below) asks it to do otherwise. Hint: hg update && hg qpop -a && hg qrm `hg qser` kills the queue.

• If there are multiple patches on a ticket, the merge script will spawn an editor and allow you to choose which patches should be applied, in what order. It looks for the $EDITOR environment variable, and if it can't find it, tries pico. If both of those fail, the script exits.

More options

• -t <option>, --test <option>: none, files, directory, long

  • This determines which files to doctest after applying the patches. none will doctest nothing, files will doctest all files touched by the patches, directory will doctest all directories containing a file that was changed, and long will call make ptestlong.

• -f, --finish: On completion of testing, if all tests succeeded, commit all changes to all repositories. Incompatible with -l.

• -l, --leave-in-queue: On the completion of testing, leave all patches in queues (whether tests failed or succeeded), as long as all patches applied correctly. Incompatible with -f.

• -n <number>, --num-threads <number>: number of threads to use for parallel doctesting.

• -r <repo>, --repository <repo>: default repository to commit patches to

• -d <dir>, --directory <dir>: directory to store patch files in (defaults to a temp dir)

• -o, --overwrite: if storing files to a different directory, whether or not to overwrite files when downloading

Standard use cases

• sage -merge <ticket_number> -t long -n 12: This will take a ticket, merge patches from that ticket, and run long doctests on the whole library with 12 threads. This is a very common case. Adding -l will let you inspect the patches and commit afterwards, and -f will simply commit if everything worked.

• sage -merge -a -f -t long: This basically creates a new alpha release. :) It merges patches with positive review from trac, running long doctests after each, and committing those that pass all doctests.

sage -merge todo list

• sage -merge seems to fail when run from certain directories (such as inside the sage tree). Look into this and fix it -- likely just add an os.chdir somewhere at the start of the script (maybe to $SAGE_ROOT?).

• It would be very easy to make sage -merge -c take an optional argument to show tickets which need review, or tickets that need work. (In fact, all_tickets takes an argument called report for exactly this reason.)

• Add code to test coverage of the code after the patches are applied.
• Add automatic logging to a file, with the default name including the ticket number, and a command line option to change it.

• Currently, one can't use this script if there are patches sitting in the queue. This wouldn't be too hard to add -- one needs to use the hg qsave and hg qrestore options to keep track of where to move back to in the queue. See hg help qsave for more info on how this works.

• Add more options for which tests to run -- in particular, something for optional doctests is needed.
• The list of what files to doctest (for the choice of doctesting everything) is currently hard-coded, and was copied out of the top-level makefile for Sage. This shouldn't be hard-coded.
• Make the warnings from the docbuilding process turn into errors, so that no new patches which cause the docs not to build will get merged unnoticed.
• Add an option to use pylint, pyflakes, pychecker, and possibly other tools to the new code. (Since the sage code base doesn't currently have this, one might need to call it on the files getting changed before they're patched, then call it again after, and diff the two.)

• Make an option for sage -merge -a to intelligently apply patches in sequence, so that one can see what goes on with applying all the patches at once (not just one at a time). One can do this with -f, but that will commit things to your repository. (Of course, you could use a clone and throw it away if it fails ...)

• Make better options for allowing people to apply patches from the same ticket to different repositories (i.e. one to sage-main, one to sage-scripts, etc.). This could involve coming up with a good way to notate this when editing the list of patches.

  • Or, one could try each patch on every repository until it works on one, assuming that no patch would cleanly apply to multiple repositories.
• Add a system to make default choices for which patches to merge, so that no interaction from the user is needed. (This would make batch processing all positive reviewed patches possible.)
• Display the trac file comments as comments after patches. For example,

  • http://trac.sagemath.org/sage_trac/raw-attachment/ticket/6273/trac_6273.patch
    http://trac.sagemath.org/sage_trac/raw-attachment/ticket/6273/trac_6273-replacement.patch
    http://trac.sagemath.org/sage_trac/raw-attachment/ticket/6273/trac_6273_new.patch

    should be

    http://trac.sagemath.org/sage_trac/raw-attachment/ticket/6273/trac_6273.patch
    http://trac.sagemath.org/sage_trac/raw-attachment/ticket/6273/trac_6273-replacement.patch
    http://trac.sagemath.org/sage_trac/raw-attachment/ticket/6273/trac_6273_new.patch
    # Replaces both previous

    (also, strongly encourage people to add comments to files when uploading in trac)

Doctesting on the sage.math cluster

When you run doctests with "make ptest" or "make ptestlong", Sage will by default use a number of threads equal to the number of cores. On sage.math, this is often inadvisable, so when running doctests, always use something like make ptest NUM_THREADS=16 (replace 16 with a sensible number; 24 minus the current load average will usually be considered "sensible")

Building on boxen.math

The machine boxen.math hosts a number of virtual machines that run various 32- and 64-bit Linux distributions. Here is a list of supported Linux distributions on boxen.math:

• CentOS 5.3 (32- and 64-bit) --- centos5.3-32 and centos5.3-64
• Debian 5.0 Lenny (32- and 64-bit) --- debian5-32 and debian5-64
• Fedora 11 (32- and 64-bit) --- fedora11-32 and fedora11-64
• Mandriva 2009.1 (32- and 64-bit) --- mandriva2009.1-32 and mandriva2009.1-64
• openSUSE 11.1 (32- and 64-bit) --- opensuse-11.1-32 and opensuse-11.1-64
• Ubuntu 9.04 (32- and 64-bit) --- ubuntu9.04-32 and ubuntu9.04-64

The following Windows XP virtual machines are also hosted on boxen.math:

• winxp1
• winxp2
• winxp3

Before starting the build process on the build farm, first login to boxen.math using either one of two methods shown below:

$ ssh -C -x -a <username>@boxen                      # use this syntax if you're 
                                                     # already login to the Sage cluster

$ ssh -C -x -a <username>@boxen.math.washington.edu  # this syntax is required if 
                                                     # you're not already login to the Sage cluster

Replace <username> with your username. Next, ensure that you have the privilege to run as the username wstein. If you are able to sudo as username wstein, then do so as follows:

$ sudo -H -u wstein bash

The option -H would set your home directory to that of username wstein, the option -u wstein is the username you want to sudo as (in this case it's wstein), and bash would login with the Bash shell. Delete all source tarballs in /home/wstein/farm/src and place the source tarball you want to build in /home/wstein/farm/src. Navigate to the directory /home/wstein/farm and run ./go, which would actually start the build process on the Linux virtualized guests. The results would appear in /home/wstein/farm/out/<version_number>/ as they complete. You can watch the build progress on all virtualized guests using

$ tail -f /home/wstein/farm/out/<version_number>/*

or watch the progress of a virtualized guest as follows:

$ tail -f /home/wstein/farm/out/<version_number>/<guest_name>.out

Press Control-C to quit watching.

Known issues:

• Both of the 32- and 64-bit Mandriva machines would fail trying to build Python, because the Python spkg won't build under nohup. Start a screen session, ssh into both boxes, and restart it by hand -- or, better yet, diagnose and fix the problem.

Building on SkyNet

Use /home/wstein/farm/go_skynet to start the build on SkyNet. The files appear in /home/wstein/build/logs/<version_number>. As of 4.0.2, eno and cleo works fine, but the rest would fail. You can restart the builds on menas, cicero, and iras by hand, and they should complete just fine.

Releasing source tarballs and binaries

You can use sage -sdist <version> to create a new source tarball, and sage -bdist <version> to create a new binary. These scripts are very particular about the formatting of version strings, see trac #6338. Both sage -sdist <version> and sage -bdist <version> take care of updating the Sage banner for you so you don't need to manually edit the file SAGE_ROOT/local/bin/sage-banner.

NOTE: When cutting a new rc or alpha release, make sure to always produce binaries for sage.math. This also applies to any stable release -- several people will make use of these. The download page for 64-bit Linux contains a text file called sage-x.y.z-sage.math.washington.edu-x86_64-Linux.txt. This file lists directories on sage.math where sage.math account holders are able to get a copy of a binary specifically built for that machine.

NOTE: After an sdist, one needs to call sage -br to rebuild the library. This is annoying, and should be fixed.

Release checklist

Here's a list of things you want to make sure to do before every release:

• Check that sage starts (you can use local/bin/sage-starts or check by hand)

• Check that notebook() works

• Check that notebook(secure=True) works

• Check that wiki() works

• Check that the documentation builds, both the HTML and PDF versions.

• Check that sage -startuptime works, and isn't worse than the last release.

• Make sure to Valgrind Sage late in the release cycle if this is a major release.

• Use sage -coverage to test that the coverage hasn't gone down.

• If any patches add files to the sage library itself, the release manager should check that the rules in MANIFEST.in include these files, or they will be missing from sage-x.x.x.spkg.

• When a new package is added to the standard spkg repository, this usually involve corresponding changes in the files SAGE_ROOT/spkg/standard/deps and SAGE_ROOT/spkg/install. The updated standard spkg repository needs to be copied over to the master server on sagemath.org. Keep in mind that the standard spkg repository contains the file SAGE_ROOT/spkg/standard/deps so copying that repository over to the master server automatically copies the deps dependencies file. However, the file SAGE_ROOT/spkg/install must be manually copied over to the master server so it can be mirrored out.

• We should incorporate max timings into the automerge scripts. Maybe reject a ticket if it triples the testing time or something. It should also say "oh, and doctesting these files took more than twice as long as normal."
• Double check the release note and release tour.

• When releasing the source version of a stable release, send a release announcement to sage-devel and sage-release. The release announcement should be accompanied with the release note. Upload the release note to http://www.sagemath.org/src/announce/ and prepend the HISTORY.txt file with that latest release note. This cumulative HISTORY.txt file should also be up on http://www.sagemath.org/src/announce/. Make sure to put this cumulative HISTORY.txt under the SAGE_ROOT of the source distribution.

• When releasing binaries for a particular stable version, send a release announcement to sage-announce. That release announcement should be accompanied with the release note, and it should also contain a link to the release tour.