On 11-10-14 12:27 AM, Patricia Shanahan wrote:
Arne Vajh?j wrote:
On 10/13/2011 6:16 AM, Arved Sandstrom wrote:
On 11-10-12 09:23 PM, Stanimir Stamenkov wrote:
Wed, 12 Oct 2011 17:04:50 -0700, /Roedy Green/:
It would be nice for debugging to include the line number of where the
code is when printing out the error message. Is there a simple way to
get it, or do you need to create a Throwable then analyse the
stacktrace?
I guess one could also use Thread.currentThread().getStackTrace():
http://download.oracle.com/javase/6/docs/api/java/lang/Thread.html#getStackTrace%28%29
But how's the current code line useful in such a log? If you see the
exact message you probably already know (or could easily find) where in
the source it is produced. If the same message is logged from
different
locations, probably the message should be revised to be more
specific as
appropriate?
[ SNIP ]
I agree 100%. In production Java applications that I help maintain or
write, line numbers only feature in the error logs, and that's because
we dump the full stack trace into the error log if there was a serious
enough exception. For logging at any other level (info, warn, debug etc)
the message format we have devised, and simply having a descriptive
enough message, always pinpoints the location in the source. At worst
you might have to do an indirect search in a properties file for a
message key, then grep the source for that key, but that's like 10 extra
seconds.
We (I) happen to usually use log4j, and in the API docs for log4j
PatternLayout, there are warnings about using the conversion specifiers
for class name, file name, method name and line number, because they are
slow or extremely slow. Since you don't actually need to compute any of
these if you do what Stanimir suggests, why take the hit?
This observation - about making the debugging message descriptive
enough, without using computed source location information, to identify
the location of the message - is applicable not just to use of a logger
like log4j, but also quick and dirty println's or printf's. For example,
if doing some throwaway printf's, I usually include a string of form
"classname.methodname: " in the message.
The problem with that approach is that it is designing the
logging based on an assumption that developers will do what they
are supposed to do. That assumption is not always true.
Another way of looking at this is as a matter of where to put the burden
of making sure the message can always be tracked back to the code that
logged it, on the programmer or on the logging infrastructure.
I understand the point you and Arne are making, but I see this issue
differently. Stanimir, and I, have advanced the notion that if the
message is sufficiently descriptive that one can locate the source of it
without difficulty. To answer your criticism, I'd simply say that if one
is so unsure about the quality of their programmers that they can't rely
on the messages being sufficiently descriptive for this purpose, all
that you've really got with the logging infrastructure supplying
location information is _tracing_.
Personally, if I wanted tracing I'd use dtrace if I had it, or use
aspects, or both.
Sure there is a burden on the programmer. Unless one assumes that the
single or main purpose of the log statements _is_ to trace, that burden
is to ensure that the log statements are descriptive and useful. That is
the prime directive. I actually don't see the point in assuming that the
log statements might not be sufficiently descriptive: sure, you can put
in the location specifiers but then you've simply got useless log
statements that you can accurately source.
In general, the more of the routine work that has to be done all over
the place the infrastructure can manage the better. At a minimum, it
saves programmer time and thinking for things the infrastructure cannot
handle. It is especially important for code that may not be heavily
tested - some log messages result from conditions that should never happen.
But in this case all we are asking the programmers to do is to put in a
sufficiently descriptive log message. That's it. Not requiring them to
at least do that is to defeat the purpose of logging.
I've had the opportunity to gather requirements for, design, and
implement major logging initiatives for a number of large J2EE/JEE
applications. In fact logging improvement, along with error handling
revamp, was the purpose of these mid-life projects. Requirements input
came from both operations support staff (who typically use Splunk to see
logs), and from production support developers (who get tasked with
fixing production defects and who work with the app logs also). I still
do other maintenance and new projects for these clients, so over the
past few years I've gotten useful feedback on how the new logging is
working. It's been quite positive: not only are the logs much better for
their intended purpose, but all the developers like the new system.
_They_ don't think it's a burden.