/brz/remove-bazaar

*********************

Bazaar Release Cycles

*********************

:status: Current policy, as of 2009-08. 

:blueprint: <https://blueprints.launchpad.net/bzr/+spec/6m-cycle>

Our users want easy access to bug fixes without other changes to the

core product. They also want a Just Works experience across the full

Bazaar ecosystem. To deliver the first and enable the second, we're

adopting some standard process patterns: a 6 monthly release cycle and a

stable branch. These changes will also have other benefits, including

better availability of bug fixes in OS distributions, more freedom to

remove old code, and less work for in packaging.

See also:

* `Bazaar Developer Document Catalog <index.html>`_

* `Releasing Bazaar <releasing.html>`_ -- the process for actually making 

  a release or release candidate.

The Problem Situation

*********************

Bazaar makes a release every month, preceded by a one-week

release-candidate test.  

In any release we may fix bugs, add formats, change the default format,

improve performance, add new commands or change command line behaviour,

change the network protocol, or deprecate APIs.  We sometimes also

introduce new bugs, regress existing behaviour or performance, remove

existing features or formats, or break behaviour or APIs depended upon by

plugins, external programs or users.

Some users are happy upgrading every month and consider the overall

positive balance of changes is worth some amount of churn.  But there are

some serious problems:

* You cannot get bug fixes without also getting disruptive changes.

* Bazaar is seen as unstable.

* Many releases cause some plugin breakage.

* One month is not a very long window for dependent programs or systems

  to catch up to changes in Bazaar before the release goes out of date.

* There's no clear indication to distributions which version they should

  package.

* If people (or their distributions) just pick an arbitrary version, they

  may all be on different arbitrary versions, therefore they will have

  different behaviour and different bugs, and sometimes interoperation

  problems.

* Any effort we, or distributors, wanted to put into backporting fixes

  would be dissipated across many possible backport target releases.

* When in the past we've tried either stalling releases for particular

  features, or having trunk frozen for some weeks, it causes churn and

  waste.  People rush features in, or already landed features wait a long

  time to be released, or branches go out of date because they cannot

  land.

The Process

************

Bazaar will make a major release every six months, which will be supported

at least until the time of the next major release.  During this support

period, we'll make incremental releases which fix bugs, but which do not

change network or disk formats or command syntax, and which do not require

updates to plugins.

We will also run a development series, which will become the next major

release.  We'll make a beta release from this every four weeks.  The

beta releases will be as stable as our current monthly releases and

completely suitable for everyday use by users who can tolerate changes

from month to month.

Having the stable series isn't a reason to cut back on QA or to make the

trunk or development releases unstable, which would only make our job

harder.  We keep our trunk in an always-releasable state, and that should

continue: any beta release could potentially be supported in the long

term, but we identify particular releases that actually will be supported.

The trunk will never be frozen: changes that pass review, other quality

checks and that are agreed amongst the developers can always be landed

into trunk.  The only restrictions will be on branches specifically

targeted at a release.

Schedule

--------

::

 2.0 --- 2.0.1 -- 2.0.2 -- ...

  \

   +--2.1beta1 -- 2.1beta2 -- ... -- 2.1rc1 -- 2.1 -- 2.1.1 -- ...

                                                \

                                                 \

                                                  +-- 3.0beta1 ...

Starting from the date of a major release: 

At four-week intervals we make a new beta release.  There will be no

separate release candidate, but if a serious problem is discovered we may

do the next beta ahead of schedule or make a point release.  There will be

about five or six releases in that series.

In parallel with this, bugs targeted to the previous major release are

merged into its branch.  We will make bugfix releases from that branch as

appropriate to the accumulation of changes, perhaps monthly, perhaps more

often if there are serious bugs, perhaps much less often if no new changes

have landed.

We will then make a release candidate for the next major release, and at

this point create a release branch for it.  We will iterate release

candidates at approximately weekly intervals until there are no bugs

blocking the final major release. 

Compared to the current process this has approximately the same amount of

release-related work, because the extra releases from the stable branch

are "paid for" by not doing RCs for the development series.

We will synchronize our major releases with Ubuntu, so that they come out

in sufficient time for some testing and margin of error before Ubuntu's

upstream freeze.

Regularity

----------

We value regular releases.  We prefer to slip a feature or fix to

a later release rather than to make a release late.  We will normally only

slip a release to fix a critical bug.

Numbering

---------

The number for a six-month cycle is chosen at the start, with an increment

to either the first field (3.0) or second field (3.1) depending on what we

expect to be the user impact of the release.  We expect releases that

culminate in a new disk format or that require changes in how people use

the tool will get a new major number.  We can change (forward only) if it

turns out that we land larger changes than were expected.

Terminology

-----------

Major releases (2.0 or 2.1)

    The big ones, every six months, intended to ship in distributions and

    to be used by stability-oriented users.

Release candidate (2.0rc1)

    A preview of a major release, made one or a few weeks beforehand at

    the time the release branch is created.  There should be few if any

    changes from the rc to the stable release.  We should avoid the

    confusing phrasing "release candidate 2.0rc1 is released"; instead use

    "available."

Bugfix releases (2.0.1)

    Based on the previous major release or bugfix; contains only bugfixes

    and perhaps documentation or translation corrections.

Stable series

    A major release and its descendant bugfix releases.

Stable release

    Either a major release or a bugfix release.

Beta release (3.0beta1)

    Made from trunk every month, except for the month there's a major

    release.  Stable and suitable for users who want the latest code and

    can live with some changes from month to month.

Development series

    The development releases leading up to a stable release.

Bug Work

--------

Bug fixes should normally be done first against the stable branch, 

reviewed against that branch, and then merged forward to trunk.

It may not always be easy to do this, if fixing the bug requires large

changes or the affected code is different in the stable and development

branches.  If the tradeoff does not seem worthwhile the bug can be fixed

only in the development branch, at least in the first instance.  If users

later want the fix backported we can discuss it.

Developers can merge the release branch into trunk as often as they like,

only asking for review if they're making nontrivial changes or feel review

is needed.

Feature and Performance Work

----------------------------

Features can be landed to the development branch at any time, and they'll

be released for testing within a month.

Performance bugs, although important, will generally not be landed in a

stable series.  Fixing performance bugs well often requires nontrivial

code changes or new formats.  These are not suitable for a stable series.

Performance bugs that can be fixed with a small safe patch can be

considered for the stable series.

Plugins

-------

Plugins that want to cooperate with this should make a series and a branch

that matches each bzr stable series, and follow similar rules in making

releases from their stable branch.  We'd expect that plugins will make a

release between the last development release of a series and the major

release candidate.

Within a stable series, anything that breaks any known plugin is

considered an API break and will be avoided.  Before

making each bugfix release, we'll test that code against important

plugins.

Within a development series, the focus is on helping plugin authors keep

up to date by giving clear error messages when an interface is removed.

We will no longer focus on letting old plugin code work with new versions

of bzrlib, which is an elusive target in Python.

This may mean that in cases where today a plugin would keep running but

give warnings, it will now fail altogether with an error.   

In return we expect more freedom to change and cleanup bzrlib code without

needing to keep old code around, or write extra compatibility shims, or

have review turnarounds related to compatibility.  Some changes, such as

removing module-global variables, that are hard to do now, will be

possible to do safely.

Discussion of plugins here includes programs that import and use bzrlib

but that aren't technically plugins.  The same approach, though the

technical considerations are different, should apply to other extensions

such as programs that use bzr through the shell interface.

Data and Network Formats

------------------------

Any development release should be able to interoperate with the previous

stable release, and any stable release should be able to interoperate with

the previous stable release.  This is a minimum and normally releases will be

able to interoperate with all previous releases as at present.

Each major release will have one recommended data format which will be the

default.  The name of the format will indicate which release series (not

specific release) it comes from: '2a' is the first supported format for

the 2.0 series, '2b' the second, etc.  We don't mention the particular

release that introduced it so as to avoid problems predicting precisely

when it will land.

During a development series we may have a series of experimental formats.

We will not leave people stranded if they test these formats, but we also

won't guarantee to keep supporting them in a future release.  If something

inserted in one development release turns out to be bad it can just be

removed in the next.

Hosting Services

-----------------

The guarantees made above about format and network interoperation

mean that hosting services such as Launchpad, Savannah, FedoraHosted,

and Sourceforge could choose to run either the stable or beta versions.

They might find it useful to run the beta version on their own beta

server.

Simultaneous Installation

-------------------------

Some people may want to simultaneously install and use both a stable

release and development release.

This can be handled in various ways either at the OS packaging or the

Python level.  We don't propose to directly address it in the upstream

source.  (For example, we will not change the bzrlib library name from one

release to the next.)  

The issue already exists with people who may want to use for example the

previous bzr release and the trunk.  There is a related issue that plugins

may be compatible with only some of the Bazaar versions people want to use

at the same time, and again that is something that can be handled

separately.

OS Distributions

----------------

OS distributors will be recommended to ship the bzr stable release that

fits their schedule, the betas leading up to that release during their own

beta period, and the bugfix releases following on from it.  They might

also choose to offer the beta releases as an alternative package.

Packaging

---------

At present we have three upstream-maintained PPAs containing Ubuntu

packages of Bazaar: ``~bzr-nightly-ppa``, ``~bzr-beta-ppa`` (rcs and

releases) and ``~bzr`` (ie stable).  We will keep these PPAs, and reorient

beta to contain the monthly beta releases, and the stable PPA to contain

stable releases, their release candidates, and bugfixes to those releases.

Some platforms with relatively less active packagers may choose to ship

only the stable releases.  This is probably better than having them only

intermittently or slowly ship the monthly releases.

Binary installers should use a version number like '2.0-1' or '2.0beta1-1'

so that the last component just reflects the packaging version, and can be

incremented if a new installer is made with no upstream source changes.

Code Freeze vs Announcement

---------------------------

We will separate the code freeze for a particular release from its actual

announcement, allowing a window of approximately one week for plugins to

be released and binary installers to be built.  On the date the

announcement is published, people will be able to easily install it.

Weekly Metronome Mail

---------------------

Every week the release manager should send a mail to the Bazaar list

covering these points (as appropriate):

* Early communication about changing dependencies or defaults

* Reminder re lifecycle and where we're up to right now, in particular the 

  dates for the next release and/or candidate.

* Summary of recent successes and pending work.

* Reminder re release objectives

* Reminder re things needing attention, e.g. bug triage, reviews, testing

  of certain things, etc.

Questions

*********

Do users actually want this?  

    Apparently yes, because it's often requested and often raised as a

    problem.

Would this confuse users?  

    It shouldn't, because it's a fairly standard scheme.

Won't it take more time to fix bugs in multiple places?

    It shouldn't, because we'll only do this when the stable bugfix seems

    economical.  When we fix bugs today in both trunk and release branches

    it normally does not take much more time.

What about bzr in Ubuntu LTS, with a five-year support life?

    Most bugs are either fixed within six months, or not fixed at all, or

    not very important, or fixed as part of a large rework of the code

    that would be too large to backport.  However, if there are fixes that

    are especially desired in an old release and feasible to do, we can do

    them without making a general commitment.

Will anyone test the beta releases?

    Probably yes, our most active users will run them, but if people would

    really rather not test them, forcing them is not helpful.

Isn't this a step backwards to a slower, less-agile process?

    No, our trunk stays releasable, and we ship every month.  We're just 

    cutting out things that hold us back (continuous rather than episodic

    API stability; RCs every month) and giving users what they demand.

How about calling the monthly releases "milestone" or "next" not "beta"?

    Those words are less scary but they also have less clear meanings.

Expected Benefits

*****************

If this plan works, we'll expect to see the following changes.  If they

don't occur, we'll think again:

* We see a distribution curve of users and bug reports across nightly, monthly

  and stable releases, indicating that each has value.

* API changes are easier or safer to make during beta periods, without

  being held back by fears of compatibility or 

* The stable releases are actually stable and don't introduce regressions

  or break plugins.

* Many bugs are fixed in stable branches, without developers feeling this

  is a waste of time.

* Distributions ship the stable releases in their stable releases and the

  bugfix releases in their bugfix releases.

* Plugin authors follow this policy, making their own bugfix releases.

* Users like it.

References

**********

#. List thread "[rfc] six-month stable release cycles", July 2009.

..

   vim: filetype=rst textwidth=74 ai shiftwidth=4