PyCalVer: Automatic Calendar Versioning

PyCalVer is a CLI-tool to search and replace version strings in your project files. This project follows the pattern conventions from calver.org.

Project/Repo:

Code Quality/CI:

Name	role	since	until
Manuel Barkhau (mbarkhau@gmail.com)	author/maintainer	2018-09	-

• Usage
  • Configuration
  • Pattern Search and Replacement
  • Examples
  • Version State
  • The Current Version
  • Bump It Up
• The PyCalVer Format
  • Parsing
  • Incrementing Behaviour
  • Lexical Ids
• Semantics of PyCalVer
  • Intentional Breaking Changes
  • Costs and Benefits
  • Unintentional Breaking Changes
  • Pinning is not a Panacea
  • Zeno's 1.0 and The Eternal Beta

Usage

Configuration

The fastest way to setup a project is to use pycalver init.

$ pip install pycalver
...
Installing collected packages: click pathlib2 typing toml pycalver
Successfully installed pycalver-202010.1038b0

$ cd myproject
~/myproject/
$ pycalver init --dry
WARNING - File not found: pycalver.toml
Exiting because of '--dry'. Would have written to pycalver.toml:

    [pycalver]
    current_version = "v202010.1001-alpha"
    version_pattern = "vYYYY0M.BUILD[-RELEASE]"
    commit_message = "bump version to {new_version}"
    commit = true
    tag = true
    push = true

    [pycalver.file_patterns]
    "README.md" = [
        "{version}",
        "{pep440_version}",
    ]
    "pycalver.toml" = [
        'current_version = "{version}"',
    ]

If you already have a setup.cfg file, the init sub-command will write to that instead.

~/myproject
$ ls
README.md  setup.cfg  setup.py

~/myproject
$ pycalver init
WARNING - Couldn't parse setup.cfg: Missing [pycalver] section.
Updated setup.cfg

This will add the something like the following to your setup.cfg (depending on what files already exist in your project):

# setup.cfg
[pycalver]
current_version = "v201902.1001-alpha"
version_pattern = "vYYYY0M.BUILD[-RELEASE]"
commit_message = "bump version to {new_version}"
commit = True
tag = True
push = True

[pycalver:file_patterns]
setup.cfg =
    current_version = {version}
setup.py =
    "{version}",
    "{pep440_version}",
README.md =
    {version}
    {pep440_version}

This probably won't cover every version number used in your project and you will have to manually add entries to pycalver:file_patterns. Something like the following may illustrate additional changes you might need to make.

[pycalver:file_patterns]
setup.cfg =
    current_version = {version}
setup.py =
    version="{pep440_version}"
src/mymodule_v*/__init__.py =
    __version__ = "{version}"
README.md =
    [CalVer {version}]
    img.shields.io/static/v1.svg?label=CalVer&message={version}&color=blue

To see if a pattern is found, you can use pycalver bump --dry, which will leave your project files untouched and only show you a diff of the changes it would have made.

$ pycalver bump --dry --no-fetch
INFO    - Old Version: v201901.1001-beta
INFO    - New Version: v201902.1002-beta
--- README.md
+++ README.md
@@ -11,7 +11,7 @@

 [![Supported Python Versions][pyversions_img]][pyversions_ref]
-[![Version v201901.1001-beta][version_img]][version_ref]
+[![Version v201902.1002-beta][version_img]][version_ref]
 [![PyPI Releases][pypi_img]][pypi_ref]

--- src/mymodule_v1/__init__.py
+++ src/mymodule_v1/__init__.py
@@ -1,1 +1,1 @@
-__version__ = "v201901.1001-beta"
+__version__ = "v201902.1002-beta"

--- src/mymodule_v2/__init__.py
+++ src/mymodule_v2/__init__.py
@@ -1,1 +1,1 @@
-__version__ = "v201901.1001-beta"
+__version__ = "v201902.1002-beta"

--- setup.py
+++ setup.py
@@ -44,7 +44,7 @@
     name="myproject",
-    version="201901.1001b0",
+    version="201902.1002b0",
     license="MIT",

If there is no match for a pattern, bump will report an error.

$ pycalver bump --dry --no-fetch
INFO    - Old Version: v201901.1001-beta
INFO    - New Version: v201902.1002-beta
ERROR   - No match for pattern 'img.shields.io/static/v1.svg?label=PyCalVer&message={pycalver}&color=blue'
ERROR   - Pattern compiles to regex 'img\.shields\.io/static/v1\.svg\?label=PyCalVer&message=(?P<pycalver>v(?P<year>\d{4})(?P<month>(?:0[0-9]|1[0-2]))\.(?P<bid>\d{4,})(?:-(?P
<tag>(?:alpha|beta|dev|rc|post|final)))?)&color=blue'

The internally used regular expression is also shown, which you can use to debug the issue, for example on regex101.com.

Pattern Search and Replacement

The neat thing about PyCalVer is that you don't have to separately declare the search pattern and the replacement template. You declare one pattern that is used to derive both.

for valid placeholders is treated as literal text. Available placeholders are:

You define your pattern in the pycalver:version_pattern option of your config.

You can also define custom patterns in the items of the pycalver:file_patterns section of your configuration, but usually it will be easier to just use the special {version} and {pep440_version} patterns, which are derived from what you configure as your version_pattern. is used both to search and also to replace version strings in your projects files. Everything except for valid placeholders is treated as literal text. Available placeholders are:

These patterns are closely based on https://calver.org/

placeholder	range / example(s)	comment
YYYY	2019, 2020...	%Y
YY	18, 19..99, 1, 2	int(%y)
0Y	18, 19..99, 01, 02	%y
Q	1, 2, 3, 4	quarter
MM	9, 10, 11, 12	int(%m)
0M	09, 10, 11, 12	%m
DD	1, 2, 3..31	int(%d)
0D	01, 02, 03..31	%d
JJJ	1,2,3..366	int(%j)
00J	001, 002..366	%j
BUILD	0011, 1001, 1002, ..	build number (lexid)
BLD	11, 1001, 1002, ..	zero truncated BUILD
MAJOR	0..9, 10..99, 100..	--major
MINOR	0..9, 10..99, 100..	-m/--minor
PATCH	0..9, 10..99, 100..	-p/--patch
NUM	0, 1, 2...	-r/--release-num
RELEASE	alpha, beta, rc	--release=<tag>
PYTAG	a, b, rc	--release=<tag>

Week Numbering

Week numbering is a bit special, as it depends on your definition of "week":

• Does it start on a Monday or a Sunday?
• Range from 0-52 or 1-53 ?
• At the beginning/end of the year, do you have partial weeks or do you have a week that span mutliple years?
• If a week spans multiple years, what is the year number?

placeholder	range / example(s)	comment
WW	0, 1, 2..52	int(%W)
0W	00, 01, 02..52	%W
UU	0, 1, 2..52	int(%U) us_week
0U	00, 01, 02..52	%U us_week
VV	1, 2..53	int(%V) iso week
0V	01, 02..53	%V iso_week
GGGG	2019, 2020...	strftime("%G") ISO 8601 week-based year
GG	19, 20...99, 0, 1	Short ISO 8601 week-based year
0G	19, 20...99, 00, 01	Zero-padded ISO 8601 week-based year

Normalization Caveats

Since other tools parse your version numbers, they may not care about your choice of formatting. In the case of Python, the packaging tools (such as pypi.org) follow PEP440 normalization rules.

According to these rules:

• Any non-numerical prefix (such as v) is removed
• Leading zeros in parts are truncated XX.08 -> XX.8
• Tags are converted to a short form (-alpha -> a0)

For example:

• Pattern: vYY.0M.0D[-RELEASE]
• Version: v20.08.02-beta
• PEP440 : 20.8.2b0

It may be confusing to your users to see versions displayed in two different forms. It is not immediately obvious that v20.08.02-beta is the same 20.8.2b0 on pypi. If you wish to avoid this, you should usa a pattern which is as close as possible to the normalized form of your version.

It may also be confusing to your users if they a list of version numbers, sorted lexiographically by some tool (e.g. a list of git tags) and a newer version is listed after older versions like this:

3.9.1
3.8.1
3.8.0
3.10.0

If you wish to avoid this, you should use a pattern which maintains lexiographical ordering.

pattern	example	lexio.	PEP440	lexio.
YYYY0M.BUILD[-RELEASE]		yes		yes
YYYY.BUILD[-RELEASE]		yes		yes
YYYY0M.MINOR[-RELEASE]		yes²		yes
YY0M.BUILD[-RELEASE]		yes¹		yes¹
YYYY.MM.MINOR[-RELEASE]		no		no
YYYY.0M.MINOR[-RELEASE]		yes²		no
YYYY.WW.MINOR[-RELEASE]		no		no
YYYY.0W.MINOR[-RELEASE]		yes²		no
YYYY.0M.0D		yes		no
YYYY.MM.DD		no		no
vYYYY.0W		yes		no
vYYYY.WW		no		no
YYYY.0M		yes		no
YYYY.MM		no		no

• ¹ Until 2099. If your project has new releases after 2099, future maintainers can change YY/0Y -> YYYY so that they don't release 00.xx.
• ² As long as MINOR <= 9

Legacy Patterns

These patterns use curly braces {} and were the initial implementation. They are still supported and still follow their original semantics.

The pycalver:file_patterns section of the configuration uses a different set of placeholders and does not use curly braces to mark placeholders. It is still supported, but we don't recomend you use it.

Available placeholders are:

placeholder	range / example(s)	comment
{year}	2019...	%Y
{yy}	18, 19..99, 01, 02	%y
{quarter}	1, 2, 3, 4
{month}	09, 10, 11, 12	%m
{iso_week}	00..53	%W
{us_week}	00..53	%U
{dom}	01..31	%d
{doy}	001..366	%j
{build}	.1023	lexical id
{build_no}	1023, 20345	...
{release}	-alpha, -beta, -rc	--release=
{release_tag}	alpha, beta, rc	...

placeholder	range / example(s)	comment
{pycalver}	v201902.1001-beta
{pep440_pycalver}	201902.1b0
{semver}	1.2.3

Pattern Usage

There are some limitations to keep in mind:

1. A version string cannot span multiple lines.
2. Characters generated by a placeholder cannot be escaped.
3. The timezone is always UTC.

The lack of escaping may for example be an issue with badge URLs. You may want to put the following text in your README.md (note that shields.io parses the two "-" dashes before beta as one literal "-"):

https://img.shields.io/badge/myproject-v202010.1116--beta-blue.svg

While you could use the following pattern, which will work fine for a while:

README.md =
    /badge/myproject-{vYYYY0M.BUILD[--RELEASE]}-blue.svg

Eventually this will break, when you do a final release, at which point the following will be put in your README.md:

https://img.shields.io/badge/myproject-v202010.1117--final-blue.svg

When what you probably wanted was this (with the --final tag omitted):

https://img.shields.io/badge/myproject-v202010.1117-blue.svg

Examples

The easiest way to test a pattern is with the pycalver test sub-command.

$ pycalver test 'v18w01' 'vYYw0W'
New Version: v19w06
PEP440     : v19w06

# TODO (mb 2020-09-24): Update regexp pattern

$ pycalver test 'v18.01' 'vYYw0W'
ERROR   - Invalid version string 'v18.01' for pattern
  'vYYw0W'/'v(?P<YY>\d{2})w(?P<0W>(?:[0-4]\d|5[0-2]))'
ERROR   - Invalid version 'v18.01' and/or pattern 'vYYw0W'.

As you can see, each pattern is internally translated to a regular expression. All version strings in your project must match either this regular expression or the corresponding regular expression for the PEP440 version string.

The pycalver test sub-command accepts the same cli flags as pycalver bump to update the components that are not updated automatically (eg. based on the calendar).

$ pycalver test 'v18.1.1' 'vYY.MINOR.PATCH'
New Version: v19.1.1
PEP440     : 19.1.1

$ pycalver test 'v18.1.1' 'vYY.MINOR.PATCH' --patch
New Version: v19.1.2
PEP440     : 19.1.2

$ pycalver test 'v18.1.2' 'vYY.MINOR.PATCH' --minor
New Version: v19.2.0
PEP440     : 19.2.0

$ pycalver test 'v201811.1051-beta' 'vYYYYMM.BUILD[-RELEASE]'
New Version: v201902.1052-beta
PEP440     : 201902.1052b0

$ pycalver test 'v201811.0051-beta' 'vYYYYMM.BUILD[-RELEASE]' --release rc
New Version: v201902.1052-rc
PEP440     : 201902.1052rc0

$ pycalver test 'v201811.0051-beta' 'vYYYYMM.BUILD[-RELEASE]' --release final
New Version: v201902.1052
PEP440     : 201902.1052

Note that pypi/setuptools/pip will normalize version strings to a format defined in PEP440. You can use a format that deviates from this, just be aware that version strings processed by these tools will look different.

Version State

The "current version" is considered global state that needs to be stored somewhere. Typically this might be stored in a VERSION file, or some other file which is part of the repository. This creates the risk that parallel branches can have different states. If the "current version" were defined only by files in the local checkout, the same version might be generated for different commits.

To avoid this issue, pycalver treats VCS tags as the canonical / SSOT for the most recent version and attempts to change this state in the most atomic way possible. This is why some actions of the pycalver command can take a while, as it is synchronizing with the remote repository to get the most recent versions and to push any new version tags as soon as possible.

The Current Version

The current version that will be bumped is defined either as

• Typically: The lexically largest git/mercurial tag which matches the version_pattern from your config.
• Initially: Before any tags have been created (or you're not using a supported VCS), the value of pycalver.current_version in setup.cfg / pyproject.toml / pycalver.toml.

As part of doing pycalver bump and pycalver show, your local VCS index is updated using git fetch --tags/hg pull.

$ time pycalver show --verbose
INFO    - fetching tags from remote (to turn off use: -n / --no-fetch)
INFO    - Working dir version        : v202010.1018
INFO    - Latest version from git tag: v202010.1019-beta
Current Version: v202010.1019-beta
PEP440         : 202010.1019b0

real    0m4,254s

$ time pycalver show --verbose --no-fetch
...
real    0m0,840s

Here we see that:

• The VCS had a newer version than we had locally.
• It took 4 seconds to fetch the tags from the remote repository.

This approach reduces the risk that new tags are unknown locally and makes it less likely that the same version string is generated for different commits, which would result in an ambiguous version tag. This can happen if multiple maintainers produce a release at the same time or if a build system is triggered multiple times and multiple builds run concurrently to each other.

For a small project (with only one maintainer and no build system) this is a non-issue and you can always use -n/--no-fetch to skip updating the tags.

Bump It Up

To increment the current version and publish a new version, you can use the pycalver bump sub-command. bump is configured in the pycalver config section:

[pycalver]
current_version = "v202010.1006-beta"
version_pattern = "vYYYY0M.BUILD[-RELEASE]"
commit_message = "bump version to {new_version}"
commit = True
tag = True
push = True

This configuration is appropriate to create a commit which

1. contains the changes to the version strings,
2. contains no other changes (unrelated to bumping the version),
3. is tagged with the new version,
4. has a version tag that is unique in the repository.

In order to make sure only changes to version strings are in the commit, you need to make sure you have a clean VCS checkout when you invoke pycalver bump.

The steps performed by bump are:

0. Check that your repo doesn't have any local changes.
1. Fetch the most recent global VCS tags from origin (-n/--no-fetch to disable).
2. Generate a new version, incremented from the current version.
3. Update version strings in all files configured in file_patterns.
4. Commit the updated version strings.
5. Tag the new commit.
6. Push the new commit and tag.

Again, you can use --dry to inspect the changes first.

$ pycalver bump --dry
--- setup.cfg
+++ setup.cfg
@@ -65,7 +65,7 @@

 [pycalver]
-current_version = v202010.1005-beta
+current_version = v202010.1006-beta
 version_pattern = "vYYYY0M.BUILD[-RELEASE]"
 commit_message = "bump version to {new_version}"
 commit = True
...

If everything looks OK, you can do pycalver bump.

$ pycalver bump --verbose
INFO    - fetching tags from remote (to turn off use: -n / --no-fetch)
INFO    - Old Version: v202010.1005-beta
INFO    - New Version: v202010.1006-beta
INFO    - git commit --message 'bump version to v202010.1006-beta'
INFO    - git tag --annotate v202010.1006-beta --message v202010.1006-beta
INFO    - git push origin v202010.1006-beta

Config Parameters

TODO: Descriptions

Config Parameter	Type	Description
current_version	string
version_pattern	string
commit_message	string	¹Template fro commit message
commit	boolean
tag	boolean
push	boolean

• ¹ Available placeholders:
  • {new_version}
  • {old_version}
  • {new_version_pep440}
  • {old_version_pep440}

CLI Arguments

TODO: Descriptions

CLI Argument	Description
--major
-m --minor
-p --patch
-r --release-num
--pin-date
--no-fetch
--dry
--allow-dirty

The PyCalVer Format

The PyCalVer format for version strings has three parts:

    o Year and Month of Release
    |       o Sequential Build Number
    |       |      o Release Tag (optional)
    |       |      |
 ---+---  --+--  --+--
 v202008  .0123  -beta

Some examples:

v201711.0001-alpha
v201712.0027-beta
v201801.0031
v201801.0032-post
...
v202207.18133
v202207.18134

This slightly verbose format was chosen in part to be distinctive from others, so that users of your package can see at a glance that your project will strive to maintain the one semantic that really matters: newer == better.

To convince you of the merits of not breaking things, here are some resources which PyCalVer was inspired by:

• "Speculation" talk by Rich Hicky
• Designing a Version by Mahmoud Hashemi
• calver.org
• "The cargo cult of versioning" by Kartik Agaram
• The bumpversion project, upon which PyCalVer is partially based.
• "Our Software Dependency Problem" by Russ Cox

Parsing

These version strings can be parsed with the following regular expression:

import re

# https://regex101.com/r/fnj60p/10
PYCALVER_PATTERN = r"""
\b
(?P<pycalver>
    (?P<vYYYYMM>
       v                    # "v" version prefix
       (?P<year>\d{4})
       (?P<month>\d{2})
    )
    (?P<build>
        \.                  # "." build nr prefix
        (?P<build_no>\d{4,})
    )
    (?P<release>
        \-                  # "-" release prefix
        (?P<release_tag>alpha|beta|dev|rc|post)
    )?
)(?:\s|$)
"""
PYCALVER_REGEX = re.compile(PYCALVER_PATTERN, flags=re.VERBOSE)

version_str = "v201712.0001-alpha"
version_match = PYCALVER_REGEX.match(version_str)

assert version_match.groupdict() == {
    "pycalver"   : "v201712.0001-alpha",
    "vYYYYMM"    : "v201712",
    "year"       : "2017",
    "month"      : "12",
    "build"      : ".0001",
    "build_no"   : "0001",
    "release"    : "-alpha",
    "release_tag": "alpha",
}

version_str = "v201712.0033"
version_match = PYCALVER_REGEX.match(version_str)

assert version_match.groupdict() == {
    "pycalver"   : "v201712.0033",
    "vYYYYMM"    : "v201712",
    "year"       : "2017",
    "month"      : "12",
    "build"      : ".0033",
    "build_no"   : "0033",
    "release"    : None,
    "release_tag": None,
}

Incrementing Behaviour

To see how version strings are incremented, we can use pycalver test:

$ pycalver test v201801.0033-beta
New Version: v201902.0034-beta
PEP440     : 201902.34b0

This is the simple case:

• The calendar component is updated to the current year and month.
• The build number is incremented by 1.
• The optional release tag is preserved as is.

You can explicitly update the release tag by using the --release=<tag> argument:

$ pycalver test v201801.0033-alpha --release=beta
New Version: v201902.0034-beta
PEP440     : 201902.34b0

$ pycalver test v201902.0034-beta --release=final
New Version: v201902.0035
PEP440     : 201902.35

To maintain lexical ordering of version numbers, the version number is padded with extra zeros (see Lexical Ids ).

Lexical Ids

VCS tags.

This sorting even works correctly in JavaScript!

Semantics of PyCalVer

Disclaimer: This section can of course only be aspirational. There is nothing to prevent package maintainers from publishing packages with different semantics than what is presented here.

PyCalVer places a greater burden on package maintainers than SemVer. Backward incompatibility is not encoded in the version string, because maintainers should not intentionally introduce breaking changes. This is great for users of a package, who can worry a bit less about an update causing their project to break. A paranoid user can of course still pin to a known good version, and freezing dependencies for deployments is still a good practice, but for development, users ideally shouldn't need any version specifiers in their requirements.txt. This way they always get the newest bug fixes and features.

Part of the reason for the distinctive PyCalVer version string, is for users to be able to recognize, just from looking at the version string, that a package comes with the promise (or at least aspiration) that it won't break, that it is safe for users to update. Compare this to a SemVer version string, where maintainers explicitly state that an update might break their program and that they may have to do extra work after updating and even if it hasn't in the past, the package maintainers anticipate that they might make such breaking changes in the future.

In other words, the onus is on the user of a package to update their software, if they want to update to the latest version of a package. With PyCalVer the onus is on package maintainer to maintain backward compatibility.

Ideally users can trust the promise of a maintainer that the following semantics will always be true:

• Newer is compatible.
• Newer has fewer bugs.
• Newer has more features.
• Newer has equal or better performance.

Alas, the world is not ideal. So how do users and maintainers deal with changes that violate these promises?

Intentional Breaking Changes

Namespaces are a honking great idea

• let's do more of those!

• The Zen of Python

If you must make a breaking change to a package, instead of incrementing a number, the recommended approach with PyCalVer is to create a whole new namespace. Put differently, the major version becomes part of the name of the module or even of the package. Typically you might add a numerical suffix, eg. mypkg -> mypkg2.

In the case of python distributions, you can include multiple module packages like this.

# setup.py
setuptools.setup(
    name="my-package",
    license="MIT",
    packages=["mypkg", "mypkg2"],
    package_dir={"": "src"},
    ...
)

In other words, you can ship older versions side by side with newer ones, and users can import whichever one they need. Alternatively you can publish a new package distribution, with new namespace, but please consider also renaming the module.

# setup.py
setuptools.setup(
    name="my-package-v2",
    license="MIT",
    packages=["mypkg2"],
    package_dir={"": "src"},
    ...
)

Users will have an easier time working with your package if import mypkg2 is enough to determine which version of your project they are using. A further benefit of creating multiple modules is that users can import both old and new modules in the same environment and can use some packages which depend on the old version as well as some that depend on the new version. The downside for users, is that they may have to do minimal changes to their code, even if the breaking change did not affect them.

- import mypkg
+ import mypkg2

  def usage_code():
-     mypkg.myfun()
+     mypkg2.myfun()

Costs and Benefits

If this seems like overkill because it's a lot of work for you as a maintainer, consider first investing some time in your tools, so you minimize future work required to create new packages. I've done this for my personal projects, but you may find other approaches to be more appropriate for your use.

If this seems like overkill because you're not convinced that imposing a very small burden on users is such a big deal, consider that your own projects may indirectly depend on dozens of libraries which you've never even heard of. If every maintainer introduced breaking changes only once per year, users who depend on only a dozen libraries would be dealing with packaging issues every month! In other words: Breaking things is a big deal. A bit of extra effort for a few maintainers seems like a fair trade to lower the effort imposed on many users, who would be perfectly happy to continue using the old code until they decide when to upgrade.

Unintentional Breaking Changes

The other kind of breaking change is the non-intentional kind, otherwise known as a "bug" or "regression". Realize first of all, that it is impossible for any versioning system to encode that this has happened: Since the maintainer isn't knowingly introducing a bug they naturally can't update their version numbers to reflect what they don't know about. Instead we have to deal with these issues after the fact.

The first thing a package maintainer can do is to minimize the chance of inflicting buggy software on users. After any non-trivial (potentially breaking) change, it is a good practice to first create an -alpha/-beta/-rc release. These so called --pre releases are intended to be downloaded only by the few and the brave: Those who are willing to participate in testing. After any issues are ironed out with the --pre releases, a final release can be made for the wider public.

Note that the default behaviour of pip install <package> (without any version specifier) is to download the latest final release. It will download a --pre release only if

1. no final release is available
2. the --pre flag is explicitly used, or
3. if the requirement specifier explicitly includes the version number of a pre release, eg. pip install mypkg==v202009.1007-alpha.

Should a release include a bug (heaven forbid and despite all precautions), then the maintainer should publish a new release which either fixes the bug or reverts the change. If users previously downloaded a version of the package which included the bug, they only have to do pip install --upgrade <package> and the issue will be resolved.

Perhaps a timeline will illustrate more clearly:

v202008.1665       # last stable release
v202008.1666-beta  # pre release for testers
v201901.1667       # final release after testing

# bug is discovered which effects v202008.1666-beta and v201901.1667

v201901.1668-beta  # fix is issued for testers
v201901.1669       # fix is issued everybody

# Alternatively, revert before fixing

v201901.1668       # same as v202008.1665
v201901.1669-beta  # reintroduce change from v202008.1666-beta + fix
v201901.1670       # final release after testing

In the absolute worst case, a change is discovered to break backward compatibility, but the change is nonetheless considered to be desirable. At that point, a new release should be made to revert the change.

This allows 1. users who were exposed to the breaking change to update to the latest release and get the old (working) code again, and 2. users who were not exposed to the breaking change to never even know anything was broken.

Remember that the goal is to always make things easy for users who have your package as a dependency. If there is any issue whatsoever, all they should have to do is pip install --update. If this doesn't work, they may have to temporarily pin to a known good version, until a fixed release has been published.

After this immediate fire has been extinguished, if the breaking change is worth keeping, then create a new module or even a new package. This package will perhaps have 99% overlap to the previous one and the old one may eventually be abandoned.

mypkg  v202008.1665    # last stable release
mypkg  v202008.1666-rc # pre release for testers
mypkg  v201901.1667    # final release after testing period

# bug is discovered in v202008.1666-beta and v201901.1667

mypkg  v201901.1668    # same as v202008.1665

# new package is created with compatibility breaking code

mypkg2 v201901.1669    # same as v201901.1667
mypkg  v201901.1669    # updated readme, declaring support
                       # level for mypkg, pointing to mypgk2
                       # and documenting how to upgrade.

Pinning is not a Panacea

Freezing your dependencies by using pip freeze to create a file with packages pinned to specific version numbers is great to get a stable and repeatable deployment.

The main problem with pinning is that it is another burden imposed on users, and it is a burden which in practice only some can bear. The vast majority of users either 1) pin their dependencies and update them without determining what changed or if it is safe for them to update, or 2) pin their dependencies and forget about them. In case 1 the only benefit is that users might at least be aware of when an update happened, so they can perhaps correlate that a new bug in their software might be related to a recent update. Other than that, keeping tabs on dependencies and updating without diligence is hardly better than not having pinned at all. In case 2, an insurmountable debt will pile up and the dependencies of a project are essentially frozen in the past.

Yes, it is true that users will be better off if they have sufficient test coverage to determine for themselves that their code is not broken even after their dependencies are updated. It is also true however, that a package maintainer is usually in a better position to judge if a change might cause something to break.

Zeno's 1.0 and The Eternal Beta

There are two opposite approaches to backward compatibility which find a reflection in the version numbers they use. In the case of SemVer, if a project has a commitment to backward compatibility, it may end up never incriminating the major version, leading to the Zeno 1.0 paradox. On the other end are projects that avoid any commitment to backward compatibility and forever keep the "beta" label.

Of course an unpaid Open Source developer does not owe anybody a commitment to backward compatibility. Especially when a project is young and going through major changes, such a commitment may not make any sense. For these cases you can still use PyCalVer, just so long as there is a big fat warning at the top of your README, that your project is not ready for production yet.

Note that there is a difference between software that is considered to be in a "beta" state and individual releases which have a -beta tag. These do not mean the same thing. In the case of releases of python packages, the release tag (-alpha, -beta, -rc) says something about the stability of a particular release. This is similar (perhaps identical) to the meaning of release tags used by the CPython interpreter. A release tag is not a statement about the general stability of the software as a whole, it is metadata about a particular release artifact of a package, eg. a .whl file.