PEP: 552 Title: Deterministic pycs Version: $Revision$ Last-Modified: $Date$ Author: Benjamin Peterson Status: Final Type: Standards Track Content-Type: text/x-rst Created: 04-Sep-2017 Python-Version: 3.7 Post-History: 2017-09-07 Resolution: https://mail.python.org/pipermail/python-dev/2017-September/149649.html Abstract ======== This PEP proposes an extension to the pyc format to make it more deterministic. Rationale ========= A `reproducible build`_ is one where the same byte-for-byte output is generated every time the same sources are built—even across different machines (naturally subject to the requirement that they have rather similar environments set up). Reproducibility is important for security. It is also a key concept in content-based build systems such as Bazel_, which are most effective when the output files’ contents are a deterministic function of the input files’ contents. The current Python pyc format is the marshaled code object of the module prefixed by a `magic number`_, the source timestamp, and the source file size. The presence of a source timestamp means that a pyc is not a deterministic function of the input file’s contents—it also depends on volatile metadata, the mtime of the source. Thus, pycs are a barrier to proper reproducibility. Distributors of Python code are currently stuck with the options of 1. not distributing pycs and losing the caching advantages 2. distributing pycs and losing reproducibility 3. carefully giving all Python source files a deterministic timestamp (see, for example, https://github.com/python/cpython/pull/296) 4. doing a complicated mixture of 1. and 2. like generating pycs at installation time None of these options are very attractive. This PEP proposes allowing the timestamp to be replaced with a deterministic hash. The current timestamp invalidation method will remain the default, though. Despite its nondeterminism, timestamp invalidation works well for many workflows and usecases. The hash-based pyc format can impose the cost of reading and hashing every source file, which is more expensive than simply checking timestamps. Thus, for now, we expect it to be used mainly by distributors and power use cases. (Note there are other problems [#frozensets]_ [#interning]_ we do not address here that can make pycs non-deterministic.) Specification ============= The pyc header currently consists of 3 32-bit words. We will expand it to 4. The first word will continue to be the magic number, versioning the bytecode and pyc format. The second word, conceptually the new word, will be a bit field. The interpretation of the rest of the header and invalidation behavior of the pyc depends on the contents of the bit field. If the bit field is 0, the pyc is a traditional timestamp-based pyc. I.e., the third and forth words will be the timestamp and file size respectively, and invalidation will be done by comparing the metadata of the source file with that in the header. If the lowest bit of the bit field is set, the pyc is a hash-based pyc. We call the second lowest bit the ``check_source`` flag. Following the bit field is a 64-bit hash of the source file. We will use a SipHash_ with a hardcoded key of the contents of the source file. Another fast hash like MD5 or BLAKE2_ would also work. We choose SipHash because Python already has a builtin implementation of it from :pep:`456`, although an interface that allows picking the SipHash key must be exposed to Python. Security of the hash is not a concern, though we pass over completely-broken hashes like MD5 to ease auditing of Python in controlled environments. When Python encounters a hash-based pyc, its behavior depends on the setting of the ``check_source`` flag. If the ``check_source`` flag is set, Python will determine the validity of the pyc by hashing the source file and comparing the hash with the expected hash in the pyc. If the pyc needs to be regenerated, it will be regenerated as a hash-based pyc again with the ``check_source`` flag set. For hash-based pycs with the ``check_source`` unset, Python will simply load the pyc without checking the hash of the source file. The expectation in this case is that some external system (e.g., the local Linux distribution’s package manager) is responsible for keeping pycs up to date, so Python itself doesn’t have to check. Even when validation is disabled, the hash field should be set correctly, so out-of-band consistency checkers can verify the up-to-dateness of the pyc. Note also that the :pep:`3147` edict that pycs without corresponding source files not be loaded will still be enforced for hash-based pycs. The programmatic APIs of ``py_compile`` and ``compileall`` will support generation of hash-based pycs. Principally, ``py_compile`` will define a new enumeration corresponding to all the available pyc invalidation modules:: class PycInvalidationMode(Enum): TIMESTAMP CHECKED_HASH UNCHECKED_HASH ``py_compile.compile``, ``compileall.compile_dir``, and ``compileall.compile_file`` will all gain an ``invalidation_mode`` parameter, which accepts a value of the ``PycInvalidationMode`` enumeration. The ``compileall`` tool will be extended with a command new option, ``--invalidation-mode`` to generate hash-based pycs with and without the ``check_source`` bit set. ``--invalidation-mode`` will be a tristate option taking values ``timestamp`` (the default), ``checked-hash``, and ``unchecked-hash`` corresponding to the values of ``PycInvalidationMode``. ``importlib.util`` will be extended with a ``source_hash(source)`` function that computes the hash used by the pyc writing code for a bytestring **source**. Runtime configuration of hash-based pyc invalidation will be facilitated by a new ``--check-hash-based-pycs`` interpreter option. This is a tristate option, which may take 3 values: ``default``, ``always``, and ``never``. The default value, ``default``, means the ``check_source`` flag in hash-based pycs determines invalidation as described above. ``always`` causes the interpreter to hash the source file for invalidation regardless of value of ``check_source`` bit. ``never`` causes the interpreter to always assume hash-based pycs are valid. When ``--check-hash-based-pycs=never`` is in effect, unchecked hash-based pycs will be regenerated as unchecked hash-based pycs. Timestamp-based pycs are unaffected by ``--check-hash-based-pycs``. References ========== .. _reproducible build: https://reproducible-builds.org/ .. _Bazel: https://bazel.build/ .. _BLAKE2: https://blake2.net/ .. _SipHash: https://131002.net/siphash/ .. [#frozensets] http://benno.id.au/blog/2013/01/15/python-determinism .. [#interning] http://bugzilla.opensuse.org/show_bug.cgi?id=1049186 .. _magic number: https://docs.python.org/3/library/importlib.html#importlib.util.MAGIC_NUMBER Credits ======= The author would like to thank Gregory P. Smith, Christian Heimes, and Steve Dower for useful conversations on the topic of this PEP. Copyright ========= This document has been placed in the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 80 coding: utf-8 End: