Differences between revisions 39 and 40
Revision 39 as of 2004-11-29 14:00:27
Size: 24397
Editor: VinaySajip
Comment:
Revision 40 as of 2004-11-29 19:51:06
Size: 27435
Editor: VinaySajip
Comment:
Deletions are marked like this. Additions are marked like this.
Line 267: Line 267:
  I feel that the option "4) provide a mechanism for using optparse as is", as I have proposed, warrants further study because it lacks the drawbacks you have identified with the other 3 options. I'll try to address your concerns about how to make configuration-with-command-line-override work easily using the config module. As a first cut, I'll aim for a simple example of a "verbose" flag which can be placed in a configuration and overridden via a command-line option. If you think that's too simple as a proof of how it can work, please indicate a use case you'd like to see. -- VinaySajip   I feel that the option "4) provide a mechanism for using optparse as is", as I have proposed, warrants further study because it lacks the drawbacks you have identified with the other 3 options. I'll try to address your concerns about how to make configuration-with-command-line-override work easily using the config module. As a first cut, I'll aim for a simple example of a "verbose" flag which can be placed in a configuration and overridden via a command-line option. If you think that's too simple as a proof of how it can work, please indicate a use case you'd like to see. Otherwise, below is the 'verbose' example: -- VinaySajip

Assume that the config file is simply
{{{
#This file is config.cfg
verbose : False
}}}

and suppose you want to be able to override this using the command line. One possible program is:
{{{
#This is test.py
from config import Config, ConfigList
from optparse import OptionParser

def getMergedConfig(filename):
    filecfg = Config(filename)
    parser = OptionParser()
    parser.add_option('-v', '--verbose', action='store_true', dest='verbose', help='Produce verbose output')
    options, args = parser.parse_args()
    cfglist = ConfigList()
    optcfg = Config()
    for opt in options.__dict__:
        optval = options.__dict__[opt]
        if optval is not None:
            optcfg[opt] = optval
    cfglist.append(optcfg)
    cfglist.append(filecfg)
    return cfglist, args

def main():
    cfg, args = getMergedConfig('config.cfg')

    print "verbose output? %r" % cfg.getByPath('verbose')
    print "args: %s" % args

main()
}}}

If you run this with

{{{
python test.py an_arg
}}}

the output is
{{{
verbose output? False
args: ['an_arg']
}}}

whereas if you run this with

{{{
python test.py -v an_arg
}}}

the output is
{{{
verbose output? True
args: ['an_arg']
}}}

This admittedly simple example points to how simple the actual application code - i.e. main() could be. I'm assuming that getMergedConfig() is a utility module which could be shared across multiple scripts, but even so it is pretty simple, in my view. Notice that main() does not know where the configured value of 'verbose' came from.

To forestall the complaint that the optparse specification appears to require some duplication, you could change the config file to:
{{{
#This file is config.cfg
verbose : False
optparse:
[
  { name: verbose, short: '-v', long: '--verbose', action: 'store_true', default: None, help: 'Produce verbose output'}
]
}}}

and the code which sets the parser options to (indent appropriately):
{{{
for optspec in filecfg.optparse:
    parser.add_option(optspec.short, optspec.long, action=optspec.action, dest=optspec.name, help=optspec.help)
}}}

Remember, this is just one way of doing it - not the only way and perhaps not the best way for your needs, but a simple enough way.

-- VinaySajip
Line 275: Line 357:
  I'm not sure about the policy part, because people have different needs and I'm not sure it's good to be too prescriptive here. -- VinaySajip
Line 281: Line 365:
  I don't think this is a real showstopper. To get a setting using my module you can use cfg.setting, cfg['setting'] or even cfg.getByPath('setting'). The last of these is useful when you have a hierarchical path which you compute, as in:

{{{
path = sys.platform + ".database.connection"
connstr = cfg.getByPath(path)
}}}

  I'm assuming the config file could have a database connection setting for each platform, in the above example. If this is not what you meant by 'more than one instance of the setting', please clarify.

  -- VinaySajip

There has been some recent debate about [http://python.org/doc/current/lib/module-ConfigParser.html ConfigParser] on the Python mailing lists. For more, see these threads:

This page serves as a place to record alternatives, and discuss (in a semi-permanent way) the features such a library might have. A new configuration parsing library could go into the Python standard library, probably in addition to the current ConfigParser module (perhaps with that module being deprecated).

Broken Out Sections

  • ConfigParserGoals: what is ConfigParser really for? (But note that the Features section is still relevant.) Also note that there has been some confusion between "In memory storage of configuration data" and "Simple persistent storage of configuration data". Part of the problem is that almost every configuration storage system (including ConfigParser, optparse, and getopt) comes with its own in-memory API. There should be some uniform means of accessing data from configuration files and command line options parsed via optparse or getopt, including the ability to override options in configuration files with command line options. Ideally, the programmer API should not normally care whether an option was set in the configuration or the command line.

Implementations

Please list interesting implementations of config parsers here.

[http://www.voidspace.org.uk/atlantibots/configobj.html ConfigObj - A simple to use config file parser] Not really an alternative to ConfigParser, but a very easy to use config file parser. Dictionary like syntax with the ability to save modified config files. Preserves comments but not indentation.

M. Chermside's candidate

[http://www.mcherm.com/publish/2004-10-17/config.py code] and [http://www.mcherm.com/publish/2004-10-17/configTest.py test cases]. Currently allows files in either str or unicode, with sensible defaults. Allows dictionary or dotted-name access (though dotted-name can fail in some cases). Allows subsections of arbitrary length. For example,

    x.y = True 
    x.y.z = 47
    x.y.a = prime

Would allow x.y to be viewed as either a value ("True") or a section.

    x.y == "True"

but also

    x.y['a'] = 'prime'
    x.y['z'] = '47'

Note that keys and values are always strings or unicode -- no autoconversion to other types. Note that this focuses on storage and API -- reading and writing is left out at the moment, and might reasonably be in a separate module for each format supported.

.ini file parser and schema

Now contains two modules -- iniparser is a parser for .ini files. It needs to be subclassed to produce useful output. By subclassing the parser you can implement complex or simple config parsers, while allowing for context-sensitive error messages.

inischema allows schema definitions, with type specifications, attribute access, etc. It is still quite young, and missing some major features (like handling sections).

Available at:

Skip's Idea

In my use of INI files I've always been annoyed that I couldn't nest sections to an arbitrary depth and had to resort to baroque XML APIs to accomplish that sort of task. I also figured a structure defined by indentation would be a good way to go, though YAML always seemed too complex. I worked up a little [http://www.musi-cal.com/~skip/python/cfgparse.py config file parser] that reads and writes files like

empty section1:
level1 = new val
section1:
# this is a comment for section1.item1:
    item1 = item 1
          # this is another comment
    subsection:  
        item2 = item 2
section2:
   subsection:
       item3 = item 3
very last = 7

Dan Gass'

[https://sourceforge.net/projects/config-py/ config-py] allows either strings (safe) or evaluated code (powerful) with the same API. It needs python 2.3, or at least dict.pop()

Vinay Sajip's implementation

The [http://www.red-dove.com/python_config.html config] module allows a hierarchical configuration scheme with support for mappings and sequences, cross-references between one part of the configuration and another, the ability to flexibly access real Python objects without full-blown eval(), an include facility, simple expression evaluation and the ability to change, save, cascade and merge configurations. It has been developed on python 2.3 but should work on version 2.2 or greater.

A simple example - with the example configuration file:

messages:
[
  {
    stream : `sys.stderr`
    message: 'Welcome'
    name: 'Harry'
  }
  {
    stream : `sys.stdout`
    message: 'Welkom'
    name: 'Ruud'
  }
  {
    stream : $messages[0].stream
    message: 'Bienvenue'
    name: Yves
  }
]

a program to read the configuration would be::

from config import Config

f = file('simple.cfg')
cfg = Config(f)
for m in cfg.messages:
    s = '%s, %s' % (m.message, m.name)
    try:
        print >> m.stream, s
    except IOError, e:
        print e

which, when run, would yield the console output::

Welcome, Harry
Welkom, Ruud
Bienvenue, Yves

One problem I have with this implementation is the configuration file syntax. I respect the need for a syntax to handle dictionaries and lists but why invent yet another language? If one wants a Python like syntax make it the Python syntax. I don't think everyone is on board with a Python syntax for configuration files though. My biggest concern is to have a syntax that supports heirarchies and being able to construct Python objects for configuration settings. I lean toward using Python as the syntax (and even the parser) because of the flexibility offered but if there was another way that makes sense I'm ok with that. -- dan.gass@gmail.com

cfgparse

[http://www.cs.wisc.edu/~param/software/cfgparse/ cfgparse] - cfgparse is a Python module that provides mechanisms for managing configuration information. It is backward compatible with ConfigParser, in addition to having the following features:

  • Preserves structure of INI files: Order of sections & options, indentation (to some extent), comments, and blank lines are preserved when data is updated.

  • More convenient than ConfigParser: Values can be accessed using dotted notation, or using container syntax (cfg[key]).

  • Backward compatibility: Backward compatible implementations of ConfigParser, RawConfigParser, and SafeConfigParser are included that are API-compatible with the Python standard library. They pass all the unit tests in Python-2.3.4.

  • Extensible: It is possible to add other configuration formats, and to convert between different formats (as long as the data models are compatible).

ZConfig

[http://www.zope.org/Members/fdrake/zconfig/ ZConfig] - This Python package is a bit larger than some of the others, but provides for schema-based development of configuration structures. The schema language uses XML, but the configuration language is more like Apache's. Sections are typed and completely nestable.

Features

This is a list of features that should be taken into account. Certainly not all these features are required; maybe some aren't even desired.

  • Uses a human-editable source file format. Everyone seems fairly happy with INI files as a basis, but

    [http://cvs.zope.org/Zope3/doc/zcml/ ZCML] is an example of a different syntax.

  • Has a simple API, especially for the simplest case (read a bunch of key-value string pairs from a file).
  • Can be round-tripped, i.e. a program can modify the configuration, then write the new configuration out.
    • Bonus if you can keep comments intact when writing the file. Or at least keep track of the comments in some way, so it can be extended by someone else later (throwing away comments entirely puts up brick walls).
    • Keeps track of the order of items in the configuration file.
    • Is intelligent about default values; writes out a minimal config file.
  • Supports some notion of a config schema. You might define a certain key as being an integer, and it would be automatically converted when the file was loaded. This becomes a bit more difficult when you have more complex structures, multiple sections, etc. Defaults fit in here as well. Obviously the types available should not be fixed.
  • Supports repeating values in some fashion. Maybe a key can appear multiple times in the file. Maybe nested structures are allowed otherwise. Repeating values are common.
  • Allows multiple config files to be combined, e.g., a site config, user config, custom config. Consideration should be given to repeating values; sometimes a user config file may want to add to a site config without repeating all items in the site config.
  • Allows dynamic nesting. E.g., some configuration values may be modified only in a small context, or for a single request in a long-running server. The nesting needs to be undone later. Maybe this simply requires the config to be easily copiable (i.e., copy the config, modify that copy, throw it away when you are done); maybe something more sophisticated is possible.
  • Supports multiple syntaxes; e.g., an XML plist syntax, an INI syntax, a ZCML syntax.
  • Gives good error messages. Error messages should include file and line number. If we define types in a schema, invalid values should give good errors.
  • Supports (maybe through utility functions or otherwise) more advanced configuration directives, like including other files or string

    interpolation. But a lot of people don't like ConfigParser's current string interpolation.

  • Consider the general case of configuration settings, not just config.ini. While round-trip to human readable is important, it is also important that the in-memory version play well with other ways of setting parameters -- including some that set arbitrary python objects.

To be more explicit, it should work well with at least optparse (Optik), .ini files, .xml files, and computed-at-runtime values. The interface to the various storage mechanisms can be different, but developers shouldn't have to repeat information across the various formats; adding an option (and default value/help message/restrictions) should only need to be done once.

  • Backward compatible; at least to Python 2.2, best if portable to Python 2.1.
  • Implements a true parser that can be subclassed and specialized. If it only has a method that parses the config file into a dictionary, then any attempts to extend or specialize the parser won't have access to line information. The parser need not be a high-profile part of the module, so long as it is available.

Discussion

Discuss. Please sign your name.

What exactly is the goal? A new API to access configuration info? Or a specific file format itself? Or both? I don't have much problem with the current ConfigParser but ideally I would like use a 'simpler' API. This would allow attribute access (a.b.c) to values, provide default values, convert some types, and do some constraint checking (xyz is required) etc. It's very possible to get this functionality through a wrapper on top of ConfigParser. IMO that is the best approach, as long as there is a way to map the same API over a different underlying file format, such as XML. I think the 'dynamic nesting' point above is outside the scope of the config access API. -- Shalabh

Three features I want in a config parser are 1) keyed settings 2) pulling in settings from multiple configurition files, and 3) ability for user to pass in real python objects through the settings. The "keyed" settings can be thought of as namespaces and I need an arbitrary number of key nestings. For certain applications I EXPECT the user to pass in python objects that meet a specified API. This allows the user to customize a certain operation however they would like with the full power and flexibility of python. I then don't need my tool to be tailered with a switch statement having custom solutions for each user type. This would require the configuration file (or parts of it) to be able to be executed as a python script and I realize this is a security hole that would be unacceptable to many. What I would propose is the config parser module support two methodologies, both sharing the same API and configuration file syntax. One would parse the config file and prevent security issues, the other would either execute the whole config file or be a combination parse/execute but would support attaching real python objects to configuration settings. These features have been implemented in a configuration parser https://sourceforge.net/projects/config-py/ (needs python 2.3 unless the dict pops are reimplemented) and is available for experimentation/use. -- dan.gass@gmail.com 28oct04

I think it is reasonable to ask the setting code to create the object (possibly by running a random string); the config system just needs to accept objects that have already been made. A round-trip is useful, but I'm not sure source code is the best way to do that; editing will probably require an external tool anyhow. Maybe just use pickle to save arbitrary objects? (And avoid storing them *within* the config, as much as possible.) -- Jim J Jewett

ConfigParser, optparse Marriage

A marriage of the two would make a lot of sense to me. My thought is for the user (script) to instantiate a parser and add options to it much like optparse does today. When the parser processes the command line args (or args passed to it) it should also look at the configuration file (possibly using the long setting name) for the option settings. Its first priority would be command line args.

  • I don't believe that a configuration module should need to interface directly to command line options. Rather, an application should inspect command line arguments, select one or more configuration files to load, and then examine the configuration data together with command line options to determine behaviour. My [http://www.red-dove.com/python_config.html#integrating-with-command-line-options config module documentation] shows one way to easily interface with optparse. -- VinaySajip An easy way, maybe, but it doesn't seem very useful to me. What's the advantage of fetching a command-line option from cfg.cmdline_values.verbose instead of options.verbose? Also, we need to specify every option twice, once for optparse and once in the configuration file itself. For docutils, this would mean every config file would have 40(!) lines which are, essentially, implementation details.

    • It was only an example, and I accept it may not be the best example to showcase the functionality. Consider this, though: you can use a ConfigList to cascade configurations. If one of these is a set of options picked up from the command line, and another is from configuration files, then you can easily in your application just get e.g. 'cfglist.verbosity' which is picked from a command line option (if defined) or from the configuration file (otherwise). There's no need to define things twice, other than to define what happens when command line options are not provided. I'll work with you, if you like, to see how docutils configuration can be achieved using the module. -- VinaySajip

    These are probably fixable, but I think they exist because you're not solving the interesting part of optparse integration: every application still has to merge the command-line options with the options from the configuration file manually. I just grepped through my 'src' directory and found shtoom, docutils, roundup, logilab-common and spambayes all have modules which allow them to specify options only once. IPython, offlineimap, and trac are still merging without the aid of such a module.
    • The work required to merge can be as simple as setting up a ConfigList with a command-line-based configuration object first in the list, followed by one or more configurations read from files. Thereafter, the application resolves configuration data by asking the ConfigList. If a configuration item has been referenced in the command line, that value is returned; otherwise, a value defined in a configuration file is returned. You could even have e.g. project, user and application level configuration files, though there is no need for such complexity in most cases, and the API remains IMO fairly lean. -- VinaySajip

    Given that this is a problem which is being solved again and again, maybe the Python library should be providing mechanism *and* policy here. I don't think it would be too hard to draw up a reasonable default.

    -- JohannesGijsbers

    I whole heartedly agree with Johannes' comments. In my scripts I have no interest in knowing whether the setting came from the command line or from a configuration file. It just adds complexity to the script and the explanation how to use it to even allow the use of such knowledge. I don't know if I understood the ConfigList approach completely but it seems to me it adds an unnecessary layer and complexity. If we were to go down the road of a marriage I believe it would be one of three possibilities 1) add functionality to optparse 2) subclass optparse or 3) make a new module independent of optparse that duplicates alot of its functionality. I think we would need the input from the optparse implementors to help with the decision as the first option would require a great deal of work for them and the second option would require a commitment by them to an interface between us. The third wouldn't require any work but they may want to be involved. -- dan.gass@gmail.com

    I feel that the option "4) provide a mechanism for using optparse as is", as I have proposed, warrants further study because it lacks the drawbacks you have identified with the other 3 options. I'll try to address your concerns about how to make configuration-with-command-line-override work easily using the config module. As a first cut, I'll aim for a simple example of a "verbose" flag which can be placed in a configuration and overridden via a command-line option. If you think that's too simple as a proof of how it can work, please indicate a use case you'd like to see. Otherwise, below is the 'verbose' example: -- VinaySajip

Assume that the config file is simply

#This file is config.cfg
verbose : False

and suppose you want to be able to override this using the command line. One possible program is:

#This is test.py
from config import Config, ConfigList
from optparse import OptionParser

def getMergedConfig(filename):
    filecfg = Config(filename)
    parser = OptionParser()
    parser.add_option('-v', '--verbose', action='store_true', dest='verbose', help='Produce verbose output')
    options, args = parser.parse_args()
    cfglist = ConfigList()
    optcfg = Config()
    for opt in options.__dict__:
        optval = options.__dict__[opt]
        if optval is not None:
            optcfg[opt] = optval
    cfglist.append(optcfg)
    cfglist.append(filecfg)
    return cfglist, args

def main():
    cfg, args = getMergedConfig('config.cfg')

    print "verbose output? %r" % cfg.getByPath('verbose')
    print "args: %s" % args

main()

If you run this with

python test.py an_arg

the output is

verbose output? False
args: ['an_arg']

whereas if you run this with

python test.py -v an_arg

the output is

verbose output? True
args: ['an_arg']

This admittedly simple example points to how simple the actual application code - i.e. main() could be. I'm assuming that getMergedConfig() is a utility module which could be shared across multiple scripts, but even so it is pretty simple, in my view. Notice that main() does not know where the configured value of 'verbose' came from.

To forestall the complaint that the optparse specification appears to require some duplication, you could change the config file to:

#This file is config.cfg
verbose : False
optparse:
[
  { name: verbose, short: '-v', long: '--verbose', action: 'store_true', default: None, help: 'Produce verbose output'}
]

and the code which sets the parser options to (indent appropriately):

for optspec in filecfg.optparse:
    parser.add_option(optspec.short, optspec.long, action=optspec.action, dest=optspec.name, help=optspec.help)

Remember, this is just one way of doing it - not the only way and perhaps not the best way for your needs, but a simple enough way.

-- VinaySajip

A hierarchical “key” or “namespace” scheme should exist so that multiple settings may be stored in the user configuration file. I would propose a standard -k, –key option always present in the parser so the user can pass a list of keys to control the group of settings to use (I’d also like some of the other ways to control keys that I use in my config.py module). In addition I’d propose a -c, –config option that specifies a configuration file to use instead of the user’s default configuration file.

  • I (VinaySajip) agree that a hierarchical namespace is essential. I'm not sure I agree with removing "-k" and "-c" from the command line namespace for all applications by reserving these for use by particular library modules.

    Along the lines of Johannes' comment regarding "providing a mechanism *and* policy" (which I agree with), I think having a consistent command line option such as -k and -c is important. I don't really care what specifically they are, even if we required the use of long form --keys and --config (change the names if you like). So long as policy is set I'm happy. -- dan.gass@gmail.com

    I'm not sure about the policy part, because people have different needs and I'm not sure it's good to be too prescriptive here. -- VinaySajip

My last proposal for today is to change the “options” interface a bit from optparse. Instead of returning a static object of options it should be an object with a get() method so that additional keys can be used to get at settings in the configuration “on the fly”. One application where this is important is in test frameworks. “Test specification” input files may have additional “keys” to be used and aren’t known when the parser is instantiated.

  • Please take a look at [http://www.red-dove.com/python_config.html this config module] and let me know how you think it fails to meet the particular use case you have in mind. I don't see any need to change optparse to work well with a configuration module -- VinaySajip

    I looked and here is where it falls short from my perspecitve. To get a setting using your module you would use "cfg.setting". Instead a method should be offered so that keys may be passed "cfg.setting.get('key1')". Sometimes a script needs more than one instance of the setting. Other times the script processes an input file (after the configuration has already been read) which specifies the key. I'm sure there are other cases as well. -- dan.gass@gmail.com I don't think this is a real showstopper. To get a setting using my module you can use cfg.setting, cfg['setting'] or even cfg.getByPath('setting'). The last of these is useful when you have a hierarchical path which you compute, as in:

path = sys.platform + ".database.connection"
connstr = cfg.getByPath(path)
  • I'm assuming the config file could have a database connection setting for each platform, in the above example. If this is not what you meant by 'more than one instance of the setting', please clarify.

    -- VinaySajip

One benefit of this marriage is that the help messages available from the command line would also apply to what can be set in the configuration file. Also there is a lot of flexibility for the user with this scheme. For options they always use they can hard code them in their default configuration as settings. If they want to temporarily override them they can use command line options. For groups of settings that are always used together, they can be used by simply passing in a key with the -k option to select the group.

  • There are other ways in which help messages for configuration can be implemented. For example, with full hierarchical name support for configuration items, you could locate help for a configuration item in the configuration item itself - e.g. help on "a.config.item" might be found at "help.a.config.item". This approach can be used to store other meta-information, not just help text. -- VinaySajip

    It's my preference to have the application control the help like optparse does. IMO its better to have that information close to where the parser is being set up and not have to have or go to another file for that information. -- dan.gass@gmail.com

The part that isn’t addressed here that bothers me yet is how to weave in the ability to pass in real python objects as settings. It would be good to have both a secure (restricted to strings and simple settings) and a second less than secure parser for those applications that need the flexibility of python. I’d like to see the API’s of the same for both the parser and the configuration file. -- dan.gass@gmail.com

  • I (VinaySajip) believe that I have covered this in [http://www.red-dove.com/python_config.html this config module] - if you feel something fundamental is missing, please let me know.

    When I use python to set up objects to pass in as settings I often use multiple lines of code to do it. Sometimes I need to import modules so that I can instantiate a class. Other times I'm using the os module to figure out where the configuration file is located because I need the path information to figure out the absolute paths of other things that are located relative to it. Don't change yours on my account I'm holding tight with mine until a new and improved standard is available (so long as it works for me). Who is concerned about the security hole of executing Python and what restrictions do they want? -- dan.gass@gmail.com

Complexity killed the cat!

Look at what happened to urllib2 - the entry level for new users was greatly enhanced, but the added functionality was really little more than what you could already do more simply with urllib. With that in mind, I think the new ConfigParser should in it's simplest form act very closely like the original ConfigParser.

ConfigParserShootout (last edited 2014-04-02 18:08:02 by SkipMontanaro)

Unable to edit the page? See the FrontPage for instructions.