|Title:||Controlling Generation of Bytecode Files|
|Post-History:||27-Jan-2003, 31-Jan-2003, 17-Jun-2005|
This PEP outlines a mechanism for controlling the generation and location of compiled Python bytecode files. This idea originally arose as a patch request  and evolved into a discussion thread on the python-dev mailing list . The introduction of an environment variable will allow people installing Python or Python-based third-party packages to control whether or not bytecode files should be generated at installation time, and if so, where they should be written. It will also allow users to control whether or not bytecode files should be generated at application run-time, and if so, where they should be written.
Add a new environment variable, PYTHONBYTECODEBASE, to the mix of environment variables which Python understands. PYTHONBYTECODEBASE is interpreted as follows:
If not defined, Python bytecode is generated in exactly the same way as is currently done. sys.bytecodebase is set to the root directory (either / on Unix and Mac OSX or the root directory of the startup (installation???) drive -- typically C:\ -- on Windows).
If defined and it refers to an existing directory to which the user has write permission, sys.bytecodebase is set to that directory and bytecode files are written into a directory structure rooted at that location.
If defined but empty, sys.bytecodebase is set to None and generation of bytecode files is suppressed altogether.
If defined and one of the following is true:
- it does not refer to a directory,
- it refers to a directory, but not one for which the user has write permission
a warning is displayed, sys.bytecodebase is set to None and generation of bytecode files is suppressed altogether.
After startup initialization, all runtime references are to sys.bytecodebase, not the PYTHONBYTECODEBASE environment variable. sys.path is not modified.
From the above, we see sys.bytecodebase can only take on two valid types of values: None or a string referring to a valid directory on the system.
During import, this extension works as follows:
- The normal search for a module is conducted. The search order is roughly: dynamically loaded extension module, Python source file, Python bytecode file. The only time this mechanism comes into play is if a Python source file is found.
- Once we've found a source module, an attempt to read a byte-compiled file in the same directory is made. (This is the same as before.)
- If no byte-compiled file is found, an attempt to read a byte-compiled file from the augmented directory is made.
- If bytecode generation is required, the generated bytecode is wrtten to the augmented directory if possible.
Note that this PEP is explicitly not about providing module-by-module or directory-by-directory control over the disposition of bytecode files.
- "bytecode base" refers to the current setting of sys.bytecodebase.
- "augmented directory" refers to the directory formed from the bytecode base and the directory name of the source file.
- PYTHONBYTECODEBASE refers to the environment variable when necessary to distinguish it from "bytecode base".
When the interpreter is searching for a module, it will use sys.path as usual. However, when a possible bytecode file is considered, an extra probe for a bytecode file may be made. First, a check is made for the bytecode file using the directory in sys.path which holds the source file (the current behavior). If a valid bytecode file is not found there (either one does not exist or exists but is out-of-date) and the bytecode base is not None, a second probe is made using the directory in sys.path prefixed appropriately by the bytecode base.
When the bytecode base is not None, a new bytecode file is written to the appropriate augmented directory, never directly to a directory in sys.path.
Conceptually, the augmented directory for a bytecode file is the directory in which the source file exists prefixed by the bytecode base. In a Unix environment this would be:
pcb = os.path.abspath(sys.bytecodebase) if sourcefile == os.sep: sourcefile = sourcefile[1:] augdir = os.path.join(pcb, os.path.dirname(sourcefile))
On Windows, which does not have a single-rooted directory tree, the drive letter of the directory containing the source file is treated as a directory component after removing the trailing colon. The augmented directory is thus derived as
pcb = os.path.abspath(sys.bytecodebase) drive, base = os.path.splitdrive(os.path.dirname(sourcefile)) drive = drive[:-1] if base == "\\": base = base[1:] augdir = os.path.join(pcb, drive, base)
During program startup, the value of the PYTHONBYTECODEBASE environment variable is made absolute, checked for validity and added to the sys module, effectively:
pcb = os.path.abspath(os.environ["PYTHONBYTECODEBASE"]) probe = os.path.join(pcb, "foo") try: open(probe, "w") except IOError: sys.bytecodebase = None else: os.unlink(probe) sys.bytecodebase = pcb
This allows the user to specify the bytecode base as a relative path, but not have it subject to changes to the current working directory during program execution. (I can't imagine you'd want it to move around during program execution.)
There is nothing special about sys.bytecodebase. The user may change it at runtime if desired, but normally it will not be modified.
In many environments it is not possible for non-root users to write into directories containing Python source files. Most of the time, this is not a problem as Python source is generally byte compiled during installation. However, there are situations where bytecode files are either missing or need to be updated. If the directory containing the source file is not writable by the current user a performance penalty is incurred each time a program importing the module is run.  Warning messages may also be generated in certain circumstances. If the directory is writable, nearly simultaneous attempts to write the bytecode file by two separate processes may occur, resulting in file corruption. 
In environments with RAM disks available, it may be desirable for performance reasons to write bytecode files to a directory on such a disk. Similarly, in environments where Python source code resides on network file systems, it may be desirable to cache bytecode files on local disks.
The only other alternative proposed so far  seems to be to add a -R flag to the interpreter to disable writing bytecode files altogether. This proposal subsumes that. Adding a command-line option is certainly possible, but is probably not sufficient, as the interpreter's command line is not readily available during installation (early during program startup???).
- Interpretation of a module's __file__ attribute. I believe the __file__ attribute of a module should reflect the true location of the bytecode file. If people want to locate a module's source code, they should use imp.find_module(module).
- Security - What if root has PYTHONBYTECODEBASE set? Yes, this can present a security risk, but so can many other things the root user does. The root user should probably not set PYTHONBYTECODEBASE except possibly during installation. Still, perhaps this problem can be minimized. When running as root the interpreter should check to see if PYTHONBYTECODEBASE refers to a directory which is writable by anyone other than root. If so, it could raise an exception or warning and set sys.bytecodebase to None. Or, see the next item.
- More security - What if PYTHONBYTECODEBASE refers to a general directory (say, /tmp)? In this case, perhaps loading of a preexisting bytecode file should occur only if the file is owned by the current user or root. (Does this matter on Windows?)
- The interaction of this PEP with import hooks has not been considered yet. In fact, the best way to implement this idea might be as an import hook. See PEP 302. 
- In the current (pre-PEP 304) environment, it is safe to delete a source file after the corresponding bytecode file has been created, since they reside in the same directory. With PEP 304 as currently defined, this is not the case. A bytecode file in the augmented directory is only considered when the source file is present and it thus never considered when looking for module files ending in ".pyc". I think this behavior may have to change.
In the examples which follow, the urllib source code resides in /usr/lib/python2.3/urllib.py and /usr/lib/python2.3 is in sys.path but is not writable by the current user.
- The bytecode base is /tmp. /usr/lib/python2.3/urllib.pyc exists and is valid. When urllib is imported, the contents of /usr/lib/python2.3/urllib.pyc are used. The augmented directory is not consulted. No other bytecode file is generated.
- The bytecode base is /tmp. /usr/lib/python2.3/urllib.pyc exists, but is out-of-date. When urllib is imported, the generated bytecode file is written to urllib.pyc in the augmented directory which has the value /tmp/usr/lib/python2.3. Intermediate directories will be created as needed.
- The bytecode base is None. No urllib.pyc file is found. When urllib is imported, no bytecode file is written.
- The bytecode base is /tmp. No urllib.pyc file is found. When urllib is imported, the generated bytecode file is written to the augmented directory which has the value /tmp/usr/lib/python2.3. Intermediate directories will be created as needed.
- At startup, PYTHONBYTECODEBASE is /tmp/foobar, which does not exist. A warning is emitted, sys.bytecodebase is set to None and no bytecode files are written during program execution unless sys.bytecodebase is later changed to refer to a valid, writable directory.
- At startup, PYTHONBYTECODEBASE is set to /, which exists, but is not writable by the current user. A warning is emitted, sys.bytecodebase is set to None and no bytecode files are written during program execution unless sys.bytecodebase is later changed to refer to a valid, writable directory. Note that even though the augmented directory constructed for a particular bytecode file may be writable by the current user, what counts is that the bytecode base directory itself is writable.
- At startup PYTHONBYTECODEBASE is set to the empty string. sys.bytecodebase is set to None. No warning is generated, however. If no urllib.pyc file is found when urllib is imported, no bytecode file is written.
In the Windows examples which follow, the urllib source code resides in C:\PYTHON22\urllib.py. C:\PYTHON22 is in sys.path but is not writable by the current user.
- The bytecode base is set to C:\TEMP. C:\PYTHON22\urllib.pyc exists and is valid. When urllib is imported, the contents of C:\PYTHON22\urllib.pyc are used. The augmented directory is not consulted.
- The bytecode base is set to C:\TEMP. C:\PYTHON22\urllib.pyc exists, but is out-of-date. When urllib is imported, a new bytecode file is written to the augmented directory which has the value C:\TEMP\C\PYTHON22. Intermediate directories will be created as needed.
- At startup PYTHONBYTECODEBASE is set to TEMP and the current working directory at application startup is H:\NET. The potential bytecode base is thus H:\NET\TEMP. If this directory exists and is writable by the current user, sys.bytecodebase will be set to that value. If not, a warning will be emitted and sys.bytecodebase will be set to None.
- The bytecode base is C:\TEMP. No urllib.pyc file is found. When urllib is imported, the generated bytecode file is written to the augmented directory which has the value C:\TEMP\C\PYTHON22. Intermediate directories will be created as needed.
|||(1, 2) patch 602345, Option for not writing py.[co] files, Klose (http://www.python.org/sf/602345)|
|||python-dev thread, Disable writing .py[co], Norwitz (https://mail.python.org/pipermail/python-dev/2003-January/032270.html)|
|||Debian bug report, Mailman is writing to /usr in cron, Wegner (http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=96111)|
|||python-dev thread, Parallel pyc construction, Dubois (https://mail.python.org/pipermail/python-dev/2003-January/032060.html)|
|||PEP 302, New Import Hooks, van Rossum and Moore (http://www.python.org/dev/peps/pep-0302)|
|||patch 677103, PYTHONBYTECODEBASE patch (PEP 304), Montanaro (http://www.python.org/sf/677103)|
This document has been placed in the public domain.