[boost] Boost.Logging

From: Richard Glanmark (Richard.Glanmark_at_[hidden])
Date: 2005-12-15 06:57:11

Some thoughts and ideas about some of the topics that has been discussed
regarding the proposed Boost.Logging library.

1. Enabled/disabled logger, logging macros and efficiency

The logging macro BOOST_LOG should accept two arguments instead of one.
The logger and the log message. The syntax

(1)	BOOST_LOG(some_log) << "my message " << foo() << bar();

should be replaced with

(2)	BOOST_LOG(some_log,  "my message " << foo() << bar());

and the macro being defined something like this

(3)	#define BOOST_LOG(logger, msg) \
                if(logger.isEnabled()) { /* code that logs msg */ ... }

Thus, disabled loggers will never take longer than simple if-statement,
as long as the isEnabled() method is fast, which IMO should take no
longer than comparing booleans.

In the original solution (1) the method foo() will always be called,
which is inefficient when the logger is disabled, in solution (2), foo()
will never be called when the logger is disabled.

Further more, the macro could insert extra information such as __LINE__,
__FILE__ etc. If needed or wanted, the macro could of course be
redefined to leave no trace of the logging. I could also envision a few
other macros that for example creates a logger object on the stack that
logs upon creation and destruction, a special macro to aid developers
that are always compiled out of release code. Perhaps such macros is
best left to the user of the library although I do think that there are
common set of macros that is useful for most applications and situations
and thus are good candidates for being defined in the library.

2. Filtering

There has been some discussion about the ability of a library to be able
to do more filtering than simple log levels, notably by Gennadiy
Rozental.

Gennadiy writes:
> At the bare minimum it should support:
>
> entry level - levels set is ordered set of values indication
> importance of information provided. Filtering is based on threshold.
> Examples: DEBUG, INFO, MAJOR

> entry category - categories set is a set of unique values indicating
> kind of information provided Filtering is based on masking.
> Examples: ARGS, RETURN_VALUE, ERROR_LOG, DATA_FLOW, PROG_FLOW

> entry keyword - keyword set is set of user defined keywords (most
> frequently strings) identifying area of the program. Filtering is
based on match of keywords.
> Keywords usually are used to mark specific part of application.

I agree that it is very important for a logging library to support this.
However, I do not think that the solution is for the logging library to
be aware of such special values. The library should be as simplistic as
possible, with no or few assumptions about its use, at least at the
basic "framework level". Only then IMO could the library accodomate
various neeeds. Let me explain. If special values are introduced in the
library, there will always be applications and users that lack some
other special value, perhaps specific to their domain, but even more
likely, to their taste and style.

My proposed solution is to basically let loggers be somewhat entry
(Gennadiy's definition above) unaware. Instead, let the developer be
entry aware. By that I mean that the developer should log messages that
belong to a specific entry category/level to a specific logger. For
example, if there is a need to log and filter a RETUN_VALUE category,
then a specific logger, perhaphs named RETURN_VALUE_LOGGER (or
whatever), and let the user send log messages to that logger. This could
look something like this:

  BOOST_LOG( RETURN_VALUE_LOGGER, "The return value is: " << value );

This solution also means that you have a different logger for each
level, level filtering perhaps being the most common filter. For
example:

  BOOST_LOG( debug, "some debug output " << a << foo() );
  BOOST_LOG( error, "Error: " << e.what() );

where 'debug' and 'error' are loggers. Now, the drawback with this
approach is that several loggers needs to be defined and configured, but
as long as this is simple this should be ok. There should also be
default definitions and configurations available from the library, which
"sits" on top of the basic library framework.

As for entry keywords (see above), I see two interpretations of
Gennadiy's idea. (1) A log statement is marked with a keyword. For
example:

   BOOST_LOGX( logger, "user defined keyword", "my log msg " << value );

Or (2), logging is filtered on what strings are contained in the log
message. That is, log statements are written in the normal way by the
programmer but is filtered by some mechanism where the log message is
matched against certain keywords, like the grep command.

For interpretation (1): My proposed solution should solve the problem.
Declare a specific logger that the programmer send log statements to for
the specified keyword.

For interpretation (2 ): IMO the filtering mechanism should either sit
on top of loggers or possibly in the appenders, or the tools that are
used to view logs (e.g. grep and such). The "on top" solution would be
to use specicial functions that utilizes the "lower" layer of simple
loggers. It should be an extension to the log library and not be part of
the core functionality (but could still be provided by a logging
library). For example:

   BOOST_LOGX( logger, Predicate( ... ), "my log msg" << value );

Thus, the approach to filtering, based on Gennadiy's idea of various
criteria is to to enable/disable loggers depending on
system/users/developers needs'. Hence, enabling/disabling must be
simple. Using John Torjo's solution of manipulate_log-functions this
ought be relative easy.

The programmer should be able to enable/disable loggers on various
critiera. One solution for this would be to attach user, as well as
common/default library, supplied name-value pairs to loggers and let the
manipulate_log functions enable/disable loggers that meet certain
criteria for these name-value pairs. For example, when defining which
loggers that are available to the programmer (by himself or some other
developer), a log level name-value object is attached to the loggers.
The logger named 'debug' is attached with a log level of 'debug' (a
literal integer value chosen carefully e.g. 0). Later on in the program,
the programmer can issue the statement, using the
manipulate_log-functions to enable all loggers that have a property name
log level with a value greater than 'debug' and so on.

Additional benefits of using this clean approach (maybe possible with
Johns solution as well (?)) to loggers is that it is quite possible to
have debug loggers log in one subsystem while normal logging is done in
the rest of the program. Any combination is possible.

The major drawback with this clean approach to loggers as I see it is
the additional complexity, cognitive burden, for setting up loggers.
Part of this problem should be alleviated by the library by providing a
default set of loggers for common/simple/I just wanna get started
-situations.

Cheers,

Richard Glanmark

The content of this e-mail is intended only for the confidential use of the person(s) to whom it is addressed. If the reader of this message is not such a person, you are hereby notified that you have received this communication in error and that reading it, copying it, or in any way disseminating its content to any other person, is strictly prohibited. If you have received this message in error, please notify the author by replying to the e-mail immediately. NeoNet AB and its subsidiaries (NeoNet Securities AB, NeoNet Securities Inc. and NeoNet Technology AB) are unable to exercise control over the content of information contained in transmissions made via the Internet and hereby excludes any warranty as to the quality or accuracy of any information contained in this message and any liability of any kind for the information contained in it, or for its transmission, reception, storage or use in any way whatsoever.