haproxy/BRANCHES

240 lines
14 KiB
Plaintext

HAProxy branches and life cycle
===============================
The HAProxy project evolves quickly to stay up to date with modern features
found in web environments but also takes a great care of addressing bugs which
may affect deployed versions without forcing such users to upgrade when not
needed. For this reason the project is developed in branches.
A branch is designated as two numbers separated by a dot, for example "1.8".
This numbering is historical. Each new development cycle increases the second
digit by one, and after it reaches '9' it goes back to zero and the first digit
increases by one. It effectively grows as a decimal number increased by 0.1 per
version.
The complete version is made of the branch suffixed with "-dev" followed by a
sequence number during development, then by "." followed by a number when the
development of that branch is finished and the branch enters a maintenance
phase. The first release of a branch starts at ".0". Immediately after ".0" is
issued, the next branch is created as "-dev0" as an exact copy of the previous
branch's ".0" version. Thus we observe the following development sequence:
... 1.9-dev10 -> 1.9-dev11 -> 1.9.0 -> 2.0-dev0 -> 2.0-dev1 ... 2.0 -> ...
Occasionally a series of "-rc" versions may be emitted between the latest -dev
and the release to mark the end of development and start of stabilizing, though
it's mostly a signal send to users that the release is approaching rather than
a change in the cycle as it is always hard to categorize patches.
Very often the terms "branch" and "version" will be used interchangeably with
only the first two digits to designate "the latest version of that branch". So
when someone asks you "Could you please try the same on 1.8", it means "1.8.X"
with X as high as possible, thus for example 1.8.20 if this one is available at
this moment.
During the maintenance phase, a maintenance branch is created for the just
released version. The development version remains in the development branch
called "master", or sometimes "-dev". If branches are represented vertically
and time horizontally, this will look like this:
versions branch
1.9-dev10 1.9-dev11 1.9.0 2.0-dev0 2.0-dev1 2.0-dev2
----+--------+---------+-------+---------+---------+----------> master
\
\ 1.9.1 1.9.2
`-----------+-------------+---------> 1.9
Each released version (e.g. 1.9.0 above) appears once in the master branch so
that it is easy to list history of changes between versions.
Before version 1.4, development and maintenance were inter-mixed in the same
branch, which resulted in latest maintenance branches becoming unstable after
some point. This is why versions 1.3.14 and 1.3.15 became maintenance branches
on their own while the development pursued on 1.3 to stabilize again in the
latest versions.
Starting with version 1.4.0, a rule has been set not to create new features
into a maintenance branch. It was not well respected and still created trouble
with certain 1.4 versions causing regressions and confusing users.
Since 1.5.0 this "no new feature" rule has become strict and maintenance
versions only contain bug fixes that are necessary in this branch. This means
that any version X.Y.Z is necessarily more stable than X.Y.W with W<Z.
For this reason there is absolutely no excuse for not updating a version within
your branch, as your version necessarily contains bugs that are fixed in any
later version in that same branch. Obviously when a branch is just released,
there will be some occasional bugs. And once in a while a fix for a recently
discovered bug may have an undesired side effect called a regression. This must
never happen but this will happen from time to time, especially on recently
released versions. This is often presented as an excuse by some users for not
updating but this is wrong, as the risk staying with an older version is much
higher than the risk of updating. If you fear there could be an issue with an
update because you don't completely trust the version in your branch, it simply
means you're using the wrong branch and need an older one.
When a bug is reported in a branch, developers will systematically ask if the
bug is present in the latest version of this branch (since developers don't
like to work on bugs that were already fixed). It's a good practice to perform
the update yourself and to test again before reporting the bug. Note, as long
as you're using a supported branch, as indicated on the haproxy.org web site,
you don't need to upgrade to another branch to report a bug. However from time
to time it may happen that a developer will ask you if you can try it in order
to help narrow the problem down. But this will never be a requirement, just a
question.
Once a bug is understood, it is tested on the development branch and fixed
there. Then the fix will be applied in turn to older branches, jumping from
one to the other in descending order. For example:
FIX
2.0-dev4 HERE 2.0-dev5 2.0-dev6
-----+-------V-------------+-----------+--------------> master
1.9.4 \ 1.9.5 1.9.6 1.9.7
--+------------o-------+---------+-------------+------> 1.9
1.8.18 \ 1.8.19 1.8.20
-----+-----------o------------+-------------+---------> 1.8
This principle ensures that you will always have a safe upgrade path from an
older branch to a newer: under no circumstances a bug that was already fixed
in an older branch will still be present in a newer one. In the diagram above,
a bug reported for 1.8.18 would be fixed between 2.0-dev4 and 2.0-dev5. The
fix will be backported into 1.9 and from there into 1.8. 1.9.5 will be issued
with the fix before 1.8.19 will be issued. This guarantees that for any version
1.8 having the fix, there always exists a version 1.9 with it as well. So if
you would upgrade to 1.8.19 to benefit from the fix and the next day decide
that for whatever new feature you need to upgrade to 1.9, you'll have 1.9.5
available with the same set of fixes so you will not reintroduce a previously
fixed problem.
In practice, it takes longer to release older versions than newer ones. There
are two reasons to this. One is technical: the fixes often require some
adaptations to be done for older versions. The other reason is stability: in
spite of the great care and the tests, there is always a faint risk that a fix
introduces a regression. By leaving fixes exposed in more recent versions
before appearing in older ones, there is a much smaller probability that such a
regression remains undetected when the next version of the older branch is
issued.
So the rule for the best stability is very simple:
STICK TO THE BRANCH THAT SUITS YOUR NEEDS AND APPLY ALL UPDATES.
With other projects, some people developed a culture of backporting only a
selection of fixes into their own maintenance branch. Usually they consider
these fixes are critical, or security-related only. THIS IS TERRIBLY WRONG.
It is already very difficult for the developers who made the initial patch to
figure if and how it must be backported to an older branch, what extra patches
it depends on to be safe, as you can imagine it is impossible for anyone else
to make a safe guess about what to pick.
A VERSION WHICH ONLY CONTAINS A SELECTION OF FIXES IS WAY MORE
DANGEROUS AND LESS STABLE THAN ONE WITHOUT ANY OF THESE FIXES.
Branches up to 1.8 are all designated as "long-term supported" ("LTS" for
short), which means that they are maintained for several years after the
release. These branches were emitted at a pace of one per year since 1.5 in
2014. As of 2019, 1.5 is still supported and widely used, eventhough it very
rarely receives updates. After a few years these LTS branches enter a
"critical fixes only" status, which means that they will rarely receive a fix
but if that a critital issue affects them, a release will be made, with or
without any other fix. Once a version is not supported anymore, it will not
receive any fix at all and it will really be time for you to upgrade to a more
recent branch. Please note that even when an upgrade is needed, a great care is
given to backwards compatibility so that most configs written for version 1.1
still work with little to no modification 16 years later on version 2.0.
Since 1.9, the release pacing has increased to match faster moving feature sets
and a faster stabilization of the technical foundations. The principle is now
the following:
- one release is emitted between October and December, with an odd version
number (such as "1.9"). This version heavily focuses on risky changes that
are considered necessary to develop new features. It can for example bring
nice performance improvements as well as invisible changes that will serve
later ; these versions will only be emitted for developers and highly
skilled users. They will not be maintained for a long time, they will
receive updates for 12 to 18 months only after which they will be marked
End-Of-Life ("EOL" for short). They may receive delicate fixes during their
maintenance cycle so users have to be prepared to see some breakage once in
a while as fixes are stabilizing. THESE VERSIONS MUST ABSOLUTELY NOT BE
PACKAGED BY OPERATING SYSTEM VENDORS.
- one release is emitted between May and June, with an even version number
(such as "2.0"). This version mostly relies on the technical foundations
brought by the previous release and tries hard not to apply risky changes.
Instead it will bring new user-visible features. Such versions will be
long-term supported and may be packaged by operating system vendors.
This development model provides better stability for end users and better
feedback for developers:
- regular users stick to LTS versions which rely on the same foundations
as the previous releases that had 6 months to stabilize. In terms of
stability it really means that the point zero version already accumulated
6 months of fixes and that it is much safer to use even just after it is
released.
- for developers, given that the odd versions are solely used by highly
skilled users, it's easier to get advanced traces and captures, and there
is less pressure during bug reports because there is no doubt the user is
autonomous and knows how to work around the issue or roll back to the last
working version.
Thus the release cycle from 1.8 to 2.2 should look like this:
1.8.0 1.9.0 2.0.0 2.1.0 2.2.0
--+---------------+---------------+--------------+--------------+----> master
\ \ \ \ \
\ \ \ \ `--> 2.2 LTS
\ \ \ `--+--+--+---+---> 2.1
\ \ `----+-----+------+-------+----> 2.0 LTS
\ `--+-+-+--+---+------+--------+-----| EOL 1.9
`---+---+---+-----+-------+-----------+---------------+------> 1.8 LTS
In short the non-LTS odd releases can be seen as technological previews of the
next feature release, and will be terminated much ealier. The plan is to barely
let them overlap with the next non-LTS release, allowing advanced users to
always have the choice between the last two major releases.
With all this in mind, what version should you use ? It's quite simple:
- if you're a first-time HAProxy user, just use the version provided by your
operating system. Just take a look at the "known bugs" section on the
haproxy.org web site to verify that it's not affected by bugs that could
have an impact for you.
- if you don't want or cannot use the version shipped with your operating
system, it is possible that other people (including the package maintainer)
provide alternate versions. This is the case for Debian and Ubuntu for
example, where you can choose your distribution and pick the branch you
need here: https://haproxy.debian.net/
- if you want to build with specific options, apply some patches, you'll
have to build from sources. If you have little experience or are not
certain to devote regular time to perform this task, take an "old" branch
(i.e. 1-2 years old max, for example 1.8 when 2.0 is emitted). You'll avoid
most bugs and will not have to work too often to update your local version.
- if you need a fresh version for application development, or to benefit from
latest improvements, take the most recent version of the most recent branch
and keep it up to date. You may even want to use the Git version or nightly
snapshots.
- if you want to develop on HAProxy, use the master from the Git tree.
- if you want to follow HAProxy's development by doing some tests without
the burden of entering too much into the development process, just use the
-dev versions of the master branch. At some point you'll feel the urge to
switch to the Git version anyway as it will ultimately simplify your work.
- if you're installing it on unmanaged servers with little to no hostile
exposure, or your home router, you should pick the latest version in one
of the oldest supported branches. While it doesn't guarantee that you will
never have to upgrade it, at least as long as you don't use too complex a
setup, it's unlikely that you will need to update it often.
And as a general rule, do not put a non-LTS version on a server unless you are
absolutely certain you are going to keep it up to date yourself and already
plan to replace it once the following LTS version is issued. If you are not
going to manage updates yourself, use pre-packaged versions exclusively and do
not expect someone else to have to deal with the burden of building from
sources.