APR's Version Numbering
This document covers how the APR projects are versioned. Since the APR
projects are libraries, it is very important to define a stable API for users
of the libraries. However, we also need to move the libraries forward,
technologically. To balance these two needs, a strict policy of versioning is
required, which users can rely upon to understand the limitations,
restrictions, and the changes that can occur from one release of APR to the
Versions are denoted using a standard triplet of integers:
MAJOR.MINOR.PATCH. The basic intent is that
MAJOR versions are incompatible, large-scale upgrades
of the API. MINOR versions retain source and binary
compatibility with older minor versions, and changes in the
PATCH level are perfectly compatible, forwards and
It is important to note that a library that has not reached 1.0.0 is
not subject to the guidelines described in this document.
Before a 1.0 release (version 0.x.y), the API can and will be
changing freely, without regard to the restrictions detailed below.
We define "source compatible" to mean that an application will continue
to build without error, and that the semantics will remain unchanged.
Applications that write against a particular version will remain
source-compatible against later versions, until the major number changes.
However, if an application uses an API which has become available in a
particular minor version, it (obviously) will no longer build or operate
against previous minor versions.
We define "binary compatible" to mean that a compiled application can
be linked (possibly dynamically) against the library and continue to function
Similar to source compatibility, an application that has been compiled
against a particular version will continue to be linkable against later
versions (unless the major number changes). It is possible that an application
will not be able to successfully link against a previous minor version.
Here are some examples to demonstrate the compatibility:
Compatibility across patch versions is guaranteed.
Compatibility across patch versions is guaranteed.
Compatibility with later minor versions is guaranteed.
Compatibility with prior minor versions is not guaranteed.
Compatibility with different major versions is not guaranteed.
Compatibility with different major versions is not guaranteed.
Note: while some of the cells say "no", it is possible that
the versions may be compatible, depending very precisely upon the particular
APIs used by the application.
This section details how we will build the code to meet the above
requirements and guidelines.
To retain perfect source and binary compatibility, a patch release can only
change function implementations. Changes to the API, to the signatures of
public functions, or to the interpretation of function parameters is
not allowed. Effectively, these releases are pure bug fix
Minor releases can introduce new functions, new symbolic and enumerated
constants, and deprecate existing functions.
- New functions
An application coded against an older minor release will still have
all of its functions available with their original signatures.
Once an application begins to use a new function, however, they
will be unable to work against older minor versions.
It is tempting to say that introducing new functions might
create incompatibility across minor releases. If an
application takes advantage of an API that was introduced in
version 2.3 of a library, then it is not going to work
against version 2.2. However, we have stated that an any
application built against version 2.2 will continue to work
for all 2.x releases. Thus, an application that states
"requires 2.3 or later" is perfectly acceptable -- the user
or administrator simply upgrades the installed library to
2.3. This is a safe operation and will not break any other
application that was using the 2.2 library.
In other words, yes an incompatibility arises by mandating
that a specific version needs to be installed. But in
practice, this will not be a problem since upgrading to
newer versions is always safe.
- New constants
- Similar to functions, all of the original (old) constants will be
available to an application. An application can then choose to use
new constants to pick up new semantics and features.
- Replacing functions
- This gets a bit trickier. The original function
must remain available at the link-level so that an
application compiled against a minor version will continue to work
with later minor versions. Further, if an application is
designed to work with an earlier minor version, then we
don't want to suddenly change the requirements for that application.
This means that the headers cannot silently map an old function
into a newer function, as that would turn an application, say,
based on 1.2 into an application requiring the 1.4 or later release.
This means that functions cannot truly be replaced. The new,
alternate function can be made available in the header and
applications can choose to use it (and become dependent upon
the minor release where the function appears).
It is possible to design a set of headers where a macro will
always refer to the "latest" function available. Of course, if an
application chooses to use this macro, then the resulting
compiled-binary will be dependent upon whatever version it was
compiled against. This strategy adds the new functionality for
applications, yet retains the necessary source and binary
compatibility for applications designed or built against previous
Constants (enumerated values and preprocessor macros) are
not allowed to change since an older application
will still be using them. Similarly, function signatures at the
link-level may not change, so that support for older, compiled
applications is maintained.
- Deprecating functions
- Since a function must remain available for applications coded
against a previous minor release, it is only possible to
"deprecate" a function. It cannot be
removed from the headers (so that source compatibility is retained)
and it cannot be removed from the library (so that binary
compatibility is retained).
If you deprecate a function in APR, please mark it as such in the
function documentation, using the doxygen "
tag. Deprecated functions can only be removed in major releases.
A deprecated function should remain available through
the original header. The function prototype should remain in the
same header, or if moved to a "deprecated functions" header, then
the alternate header should be included by the original header. This
requirement is to ensure that source compatibility is retained.
Finally, if you are deprecating a function so that you can change
the name of the function, please use the method described above
under "Replacing functions", so that projects which use APR can
retain binary compatibility.
Note that all deprecated functions will be removed at the next
major version bump.
Any kind of change can be made during a major version release. Particular
types of changes that might occur:
- remove or change constants
- remove (deprecated) functions
- fold together macro-ized function replacements
In many cases, the user of a library will need to check the version that
they are compiling against, or that is being used at runtime. Because of the
strict rules of source and binary compatibility, these checks can be simpler
and more complicated depending on what is needed.
Libraries should make their version number available as compile-time
constants. For example:
#define FOO_MAJOR_VERSION 1
#define FOO_MINOR_VERSION 4
#define FOO_PATCH_VERSION 0
The above symbols are the minimum required for this specification.
An application that desires, at compile-time, to decide on whether and how
to use a particular library feature needs to only check two values: the major
and the minor version. Since, by definition, there are no API changes across
patch versions, that symbol can be safely ignored. Note that any kind of a
check for a minimum version will then pin that application to at least that
version. The application's installation mechanism should then ensure that that
minimal version has been installed (for example, using RPM dependency checks).
If the feature changes across minor versions are source compatible, but
are (say) simply different choices of values to pass into the library, then an
application can support a wider variety of installed libraries if it avoids
A library meeting this specification should support a way for an
application to determine the library's version at run-time. This will
usually be emboded as a simple function which returns the MAJOR,
MINOR, and PATCH triplet in some form.
Run-time checks are preferable in all cases. This type of check enables an
application to run against a wider variety of minor releases of a library (the
application is "less coupled" to a particular library release). Of
course, if an application requires a function that was introduced in a later,
minor release, then the application will require that, at least, that release
is installed on the target system.
Run-time checks are particurly important if the application is trying to
determine if the library has a particular bug that may need to be worked
around, but has been fixed in a later release. If the bug is fixed in a patch
release, then the only avenue for an application is to perform a runtime check.
This is because an application cannot require a specific patch level of the
library to be installed -- those libraries are perfectly forward and backwards
compatible, and the administrator is free to choose any patch release, knowing
that all applications will continue to function properly. If the bug was fixed
in a minor release, then it is possible to use a compile-time check, but that
would create a tighter coupling to the library.
Parallel installation refers to the ability to install multiple
versions of a library simultaneously -- they exist in parallel. This document
will not discuss the full rationale for why this is important, but will instead
detail how this versioning specification maps onto those concepts. Please refer
to Havoc Pennington's
"Parallel Installation" essay for futher details and the rationale behind this form of parallel
On Unix-ish platforms, the library name should include the MAJOR
This strategy allows an application to explicitly state which version
of the library that it wants to link against. If the application was built for
version 2 of the API, then it can link against libFOO-2.so. If another
application was built against version 3 of the API, then it links against
libFOO-3.so. Since both libraries can reside on the system at the same
time, both applications' needs can be satisfied.
Typically, shared libraries on Unix-ish platforms will set up symlinks from
the .so library to specific versions of that library. For example:
libFOO-MAJOR.so -> libFOO-MAJOR.so.0
libFOO-MAJOR.so.0 -> libFOO-MAJOR.so.0.MINOR.PATCH
In this configuration, applications will be bound to the .so.0
library. The minor version does not come into play here because we want
applications to dynamically load and link to the new library when a new minor
version is installed. Thus, the MINOR and the PATCH values
are relegated to the library name after the .so.0 portion.
The implication here is that build systems for libraries should arrange
to generate .so libraries matching the above pattern.
The default installation directory for a library's include files should
specify the MAJOR version number, and should normally be installed as
a subdirectory in some standard location. For example:
An application can place the FOO-MAJOR directory on its
include path and include the files normally:
Depending upon the API that the application is designed to work against, it
can simply include different versions of the include directory.
NOTE: There is no recommendation at this time for the
best and proper handling of, say, FOO-config types of files. Or
non-code types of files (e.g. things that typically get installed into areas
Further thought and exploration is needed here.
It is expected that other libraries, besides those in the APR project, will
want to use the above definitions of versioning. This is quite fine, and those
libraries can simply reference this document. Its canonical location is: