Hi,

I have the following in my program:

try:
logging.config.fileConfig(args.config_file)
config = configparser.ConfigParser()
config.read(args.config_file)
if args.verbose:
print(f"Configuration file: {args.config_file}")
except FileNotFoundError:
print(f"Error: configuration file {args.config_file} not found. Exiting.")
sys.exit(0)

and when I ran the program I got the error

Error: configuration file /usr/local/etc/sc_mailer not found. Exiting.

However, this file *does* exist and *can* be read. By checking the
'filename' attribute of the exception I discovered that the problem was
the log file defined *in* the config file, namely

[handler_fileHandler]
class=FileHandler
level=DEBUG
formatter=defaultFormatter
args=('/var/log/my_prog.log', 'a')

This log file did not exist. The exception is thrown by

logging.config.fileConfig(args.config_file)

My questions are:

1. Should I be surprised by this behaviour?
2. In terms of generating a helpful error message, how should one
distinguish between the config file not existing and the log file not
existing?

Cheers,

Loris

--
This signature is currently under constuction.

Poor error reporting is a very common problem in programming. Python
is not anything special in this case. Of course, it would've been
better if the error reported what file wasn't found. But, usually
these problems are stacking, like in your code. Unfortunately, it's
your duty, as the language user, to anticipate those problems and act
accordingly. Now you've learned that the one file you believe that
could be the source for the error isn't the only one--well, adjust
your code to differentiate between those two (and potentially other?)
cases. There's very little else you can do beside that.

NB. On the system level, the error has no information about what file
wasn't found. It simply returns some numeric value (the famous
ENOENT) in case when the system call to open a file fails. Python
could've been more helpful by figuring out what path caused the
problem and printing that in the error message, but it doesn't...
That's why I, myself, never use the vanilla FileNotFoundError, I
always re-rise it with a customized version that incorporates the
information about the missing file in the error message.

NB2. It's always a bad idea to print logs to files. Any sysadmin /
ops / infra person worth their salt will tell you that. The only
place the logs should go to is the standard error. There are true and
tried tools that can pick up logs from that point on, and do with them
whatever your heart desires. That is, of course, unless you are
creating system tools for universal log management (in which case, I'd
question the choice of Python as a suitable language for such a task).
Unfortunately, even though this has been common knowledge for decades,
it's still elusive in the world of application development :|

On Mon, Nov 11, 2024 at 4:00 PM Loris Bennett via Python-list
<python-list@python.org> wrote:
>
> Hi,
>
> I have the following in my program:
>
> try:
> logging.config.fileConfig(args.config_file)
> config = configparser.ConfigParser()
> config.read(args.config_file)
> if args.verbose:
> print(f"Configuration file: {args.config_file}")
> except FileNotFoundError:
> print(f"Error: configuration file {args.config_file} not found. Exiting.")
> sys.exit(0)
>
> and when I ran the program I got the error
>
> Error: configuration file /usr/local/etc/sc_mailer not found. Exiting.
>
> However, this file *does* exist and *can* be read. By checking the
> 'filename' attribute of the exception I discovered that the problem was
> the log file defined *in* the config file, namely
>
> [handler_fileHandler]
> class=FileHandler
> level=DEBUG
> formatter=defaultFormatter
> args=('/var/log/my_prog.log', 'a')
>
> This log file did not exist. The exception is thrown by
>
> logging.config.fileConfig(args.config_file)
>
> My questions are:
>
> 1. Should I be surprised by this behaviour?
> 2. In terms of generating a helpful error message, how should one
> distinguish between the config file not existing and the log file not
> existing?
>
> Cheers,
>
> Loris
>
> --
> This signature is currently under constuction.
> --
> https://mail.python.org/mailman/listinfo/python-list

Loris Bennett wrote at 2024-11-11 15:05 +0100:
>I have the following in my program:
> try:
> logging.config.fileConfig(args.config_file)
> config = configparser.ConfigParser()
> config.read(args.config_file)
> if args.verbose:
> print(f"Configuration file: {args.config_file}")
> except FileNotFoundError:
> print(f"Error: configuration file {args.config_file} not found. Exiting.")

Do not replace full error information (including a traceback)
with your own reduced error message.
If you omit your "try ... except FileNotFoundError`
(or start the `except` clause with a `raise`), you
will learn where in the code the exception has been raised
and likely as well what was not found (Python is quite good
with such error details).

> ...
>My questions are:
>
>1. Should I be surprised by this behaviour?

Your code contains a major weakness (see above); thus surprises
are not unlikely.

>2. In terms of generating a helpful error message, how should one
> distinguish between the config file not existing and the log file not
> existing?

You look at the error information provided by Python
(and its library) rather than hiding it.

On Mon, 11 Nov 2024 15:05:56 +0100, Loris Bennett wrote:

> except FileNotFoundError:
> print(f"Error: configuration file {args.config_file} not found.
> Exiting.")
> sys.exit(0)
>
> and when I ran the program I got the error
>
> Error: configuration file /usr/local/etc/sc_mailer not found.
> Exiting.
>
> However, this file *does* exist and *can* be read.

This is your own fault for intercepting the exception and printing out
your own misleading error message. If you had left the exception uncaught,
it would have printed out the right file name.

On Tue, 12 Nov 2024 at 01:59, Loris Bennett via Python-list
<python-list@python.org> wrote:
> 2. In terms of generating a helpful error message, how should one
> distinguish between the config file not existing and the log file not
> existing?

By looking at the exception's attributes rather than assuming and
hard-coding the path in your message? Or, even better, just let the
exception bubble.

ChrisA

Chris Angelico <rosuav@gmail.com> writes:

> On Tue, 12 Nov 2024 at 01:59, Loris Bennett via Python-list
> <python-list@python.org> wrote:
>> 2. In terms of generating a helpful error message, how should one
>> distinguish between the config file not existing and the log file not
>> existing?
>
> By looking at the exception's attributes rather than assuming and
> hard-coding the path in your message? Or, even better, just let the
> exception bubble.

As Dieter also pointed out, I obviously made a mistake in assuming that
I knew what 'FileNotFound' was referring to.

However, it strikes me as not immediately obvious that the logging file
must exist at this point. I can imagine a situation in which I want to
configure a default log file and create it if it missing.

Cheers,

Loris

--
This signature is currently under constuction.

Chris Angelico <rosuav@gmail.com> writes:

> On Tue, 12 Nov 2024 at 01:59, Loris Bennett via Python-list
> <python-list@python.org> wrote:
>> 2. In terms of generating a helpful error message, how should one
>> distinguish between the config file not existing and the log file not
>> existing?
>
> By looking at the exception's attributes rather than assuming and
> hard-coding the path in your message? Or, even better, just let the
> exception bubble.

I didn't consider letting the exception bubble as is the top of the
code for a CLI program. I was hoping to just catch what I thought might
be a common error of the config file being missing.

--
This signature is currently under constuction.

Left Right <olegsivokon@gmail.com> writes:

> Poor error reporting is a very common problem in programming. Python
> is not anything special in this case. Of course, it would've been
> better if the error reported what file wasn't found. But, usually
> these problems are stacking, like in your code. Unfortunately, it's
> your duty, as the language user, to anticipate those problems and act
> accordingly. Now you've learned that the one file you believe that
> could be the source for the error isn't the only one--well, adjust
> your code to differentiate between those two (and potentially other?)
> cases. There's very little else you can do beside that.
>
> NB. On the system level, the error has no information about what file
> wasn't found. It simply returns some numeric value (the famous
> ENOENT) in case when the system call to open a file fails. Python
> could've been more helpful by figuring out what path caused the
> problem and printing that in the error message, but it doesn't...
> That's why I, myself, never use the vanilla FileNotFoundError, I
> always re-rise it with a customized version that incorporates the
> information about the missing file in the error message.

That sounds like a good idea.

> NB2. It's always a bad idea to print logs to files. Any sysadmin /
> ops / infra person worth their salt will tell you that. The only
> place the logs should go to is the standard error. There are true and
> tried tools that can pick up logs from that point on, and do with them
> whatever your heart desires. That is, of course, unless you are
> creating system tools for universal log management (in which case, I'd
> question the choice of Python as a suitable language for such a task).
> Unfortunately, even though this has been common knowledge for decades,
> it's still elusive in the world of application development :|

I am not entirely convinced by NB2. I am, in fact, a sort of sysadmin
person and most of my programs write to a log file. The programs are
also moderately complex, so a single program might access a database,
query an LDAP server, send email etc., so potentially quite a lot can go
wrong. They are also not programs whose output I would pipe to another
command. What would be the advantage of logging to stderr? Quite apart
from that, I find having a log file a useful for debugging when I am
developing.

Cheers,

Loris

> On Mon, Nov 11, 2024 at 4:00 PM Loris Bennett via Python-list
> <python-list@python.org> wrote:
>>
>> Hi,
>>
>> I have the following in my program:
>>
>> try:
>> logging.config.fileConfig(args.config_file)
>> config = configparser.ConfigParser()
>> config.read(args.config_file)
>> if args.verbose:
>> print(f"Configuration file: {args.config_file}")
>> except FileNotFoundError:
>> print(f"Error: configuration file {args.config_file} not found. Exiting.")
>> sys.exit(0)
>>
>> and when I ran the program I got the error
>>
>> Error: configuration file /usr/local/etc/sc_mailer not found. Exiting.
>>
>> However, this file *does* exist and *can* be read. By checking the
>> 'filename' attribute of the exception I discovered that the problem was
>> the log file defined *in* the config file, namely
>>
>> [handler_fileHandler]
>> class=FileHandler
>> level=DEBUG
>> formatter=defaultFormatter
>> args=('/var/log/my_prog.log', 'a')
>>
>> This log file did not exist. The exception is thrown by
>>
>> logging.config.fileConfig(args.config_file)
>>
>> My questions are:
>>
>> 1. Should I be surprised by this behaviour?
>> 2. In terms of generating a helpful error message, how should one
>> distinguish between the config file not existing and the log file not
>> existing?
>>
>> Cheers,
>>
>> Loris
>>
>> --
>> This signature is currently under constuction.
>> --
>> https://mail.python.org/mailman/listinfo/python-list
--
Dr. Loris Bennett (Herr/Mr)
FUB-IT, Freie Universität Berlin

> I am not entirely convinced by NB2. I am, in fact, a sort of sysadmin
> person and most of my programs write to a log file. The programs are
> also moderately complex, so a single program might access a database,
> query an LDAP server, send email etc., so potentially quite a lot can go
> wrong. They are also not programs whose output I would pipe to another
> command. What would be the advantage of logging to stderr? Quite apart
> from that, I find having a log file a useful for debugging when I am
> developing.

First, the problem with writing to files is that there is no way to
make these logs reliable. This is what I mean by saying these are
unreliable: since logs are designed to grow indefinitely, the natural
response to this design property is log rotation. But, it's
impossible to reliably rotate a log file. There's always a chance
that during the rotation some log entries will be written to the file
past the point of rotation, but prior to the point where the next logs
volume starts.

There are similar reliability problems with writing to Unix or
Internet sockets, databases etc. For different reasons, but at the
end of the day, whoever wants logs, they want them to be reliable.
Both simplicity and convention selected for stderr as the only and the
best source of logging output.

Programs that write their output to log files will always irritate
their users because users will have to do some detective work to
figure out where those files are, and in some cases they will have to
do administrative works to make sure that the location where the
program wants to store the log files is accessible, has enough free
space, is speedy enough etc. So, from the ops perspective, whenever I
come across a program that tries to write logs to anything other than
stderr, I make an earnest effort to throw that program into the gutter
and never touch it again. It's too much headache to babysit every
such program, to remember the location of the log files of every such
program, the required permissions, to provision storage. If you are
in that line of work, you just want all logs to go to the same place
(journal), where you can later filter / aggregate / correlate and
perform other BI tasks as your heart desires.

Of course, if you only administer your own computer, and you have low
single digits programs to run, and their behavior doesn't change
frequently, and you don't care to drop some records every now and
then... it's OK to log to files directly from a program. But then you
aren't really in the sysadmin / infra / ops category, as you are more
of a hobby enthusiast.

Finally, if you want your logs to go to a file, and currently, your
only option is stderr, your shell gives you a really, really simple
way of redirecting stderr to a file. So, really, there aren't any
excuses to do that.

On 11/12/24 12:10, Left Right via Python-list wrote:

> Finally, if you want your logs to go to a file, and currently, your
> only option is stderr, your shell gives you a really, really simple
> way of redirecting stderr to a file. So, really, there aren't any
> excuses to do that.

an awful lot of the programs that need to do extensive logging don't run
under control of a shell, and thus "shell redirection" helps not at all.

On Wed, 13 Nov 2024 at 07:29, Mats Wichmann via Python-list
<python-list@python.org> wrote:
>
> On 11/12/24 12:10, Left Right via Python-list wrote:
>
> > Finally, if you want your logs to go to a file, and currently, your
> > only option is stderr, your shell gives you a really, really simple
> > way of redirecting stderr to a file. So, really, there aren't any
> > excuses to do that.
>
> an awful lot of the programs that need to do extensive logging don't run
> under control of a shell, and thus "shell redirection" helps not at all.
>

Redirection's still a thing. "Shell redirection" is just the shell
syntax to request redirection.

ChrisA

On 13/11/24 8:10 am, Left Right wrote:
> since logs are designed to grow indefinitely, the natural
> response to this design property is log rotation.

I don't see how writing logs to stderr solves that problem in any way.
Whatever stderr is sent to still has a potentially unlimited amount
of data to deal with.

> But, it's
> impossible to reliably rotate a log file. There's always a chance
> that during the rotation some log entries will be written to the file
> past the point of rotation, but prior to the point where the next logs
> volume starts.

Not sure I follow you there. You seem to be thinking of a particular
way of rotating log files, where an external process tries to swap
the program's log file out from under it without its knowledge. That
could be vulnerable to race conditions. But if the program doing the
logging handles the rotation itself, there's no reason it has to
lose data.

--
Greg

On Wed, 13 Nov 2024 14:04:33 +1300, Greg Ewing wrote:

> You seem to be thinking of a particular way
> of rotating log files, where an external process tries to swap the
> program's log file out from under it without its knowledge. That could
> be vulnerable to race conditions.

But if you just use standard system facilities, like syslog or the systemd
journal, that is automatically handled for you.

Left Right <olegsivokon@gmail.com> writes:

>> I am not entirely convinced by NB2. I am, in fact, a sort of sysadmin
>> person and most of my programs write to a log file. The programs are
>> also moderately complex, so a single program might access a database,
>> query an LDAP server, send email etc., so potentially quite a lot can go
>> wrong. They are also not programs whose output I would pipe to another
>> command. What would be the advantage of logging to stderr? Quite apart
>> from that, I find having a log file a useful for debugging when I am
>> developing.
>
> First, the problem with writing to files is that there is no way to
> make these logs reliable. This is what I mean by saying these are
> unreliable: since logs are designed to grow indefinitely, the natural
> response to this design property is log rotation. But, it's
> impossible to reliably rotate a log file. There's always a chance
> that during the rotation some log entries will be written to the file
> past the point of rotation, but prior to the point where the next logs
> volume starts.

> There are similar reliability problems with writing to Unix or
> Internet sockets, databases etc. For different reasons, but at the
> end of the day, whoever wants logs, they want them to be reliable.
> Both simplicity and convention selected for stderr as the only and the
> best source of logging output.

If I understand correctly you are not saying that logrotate is less
reliable that the other methods mentioned above. But in that case,
reliability seems no more of a reason not to log to files than it is a
reason not to write to a socket or to a database.

> Programs that write their output to log files will always irritate
> their users because users will have to do some detective work to
> figure out where those files are, and in some cases they will have to
> do administrative works to make sure that the location where the
> program wants to store the log files is accessible, has enough free
> space, is speedy enough etc.

All your points regarding the work involved are valid, but many
programs, such as MariaDB, OpenLDAP or SSSD, do write to a log file (and
it is usually under /var/log or /var/log/something. So it seems like a
common approach.

Besides, I define the location of the logfile in the config file for the
program (the problem in my original question arose from this filename in
the config file not existing). So finding the location is not an issue.
You have to find the config file, of course, but I think /etc or
/usr/local/etc are fairly standard and my programs generally have an
option '--config-file' anyway.

> So, from the ops perspective, whenever I
> come across a program that tries to write logs to anything other than
> stderr, I make an earnest effort to throw that program into the gutter
> and never touch it again. It's too much headache to babysit every
> such program, to remember the location of the log files of every such
> program, the required permissions, to provision storage. If you are
> in that line of work, you just want all logs to go to the same place
> (journal), where you can later filter / aggregate / correlate and
> perform other BI tasks as your heart desires.

That may be true in many cases, but those I am dealing with don't
require much filtering beyond 'grep' and also don't require aggregation
or correlation.

> Of course, if you only administer your own computer, and you have low
> single digits programs to run, and their behavior doesn't change
> frequently, and you don't care to drop some records every now and
> then... it's OK to log to files directly from a program. But then you
> aren't really in the sysadmin / infra / ops category, as you are more
> of a hobby enthusiast.

What I do is indeed a bit of a niche, but I do get paid for this, so I
would not consider myself a 'hobby enthusiast'.

> Finally, if you want your logs to go to a file, and currently, your
> only option is stderr, your shell gives you a really, really simple
> way of redirecting stderr to a file. So, really, there aren't any
> excuses to do that.

I don't quite understand what your suggestion is. Do you mean that I
should log to stderr and then run my program as

my_program ... 2>&1 | logger

?

Cheers,

Loris

--
This signature is currently under constuction.

Op 12/11/2024 om 20:10 schreef Left Right via Python-list:
> > I am not entirely convinced by NB2. I am, in fact, a sort of sysadmin
> > person and most of my programs write to a log file. The programs are
> > also moderately complex, so a single program might access a database,
> > query an LDAP server, send email etc., so potentially quite a lot can go
> > wrong. They are also not programs whose output I would pipe to another
> > command. What would be the advantage of logging to stderr? Quite apart
> > from that, I find having a log file a useful for debugging when I am
> > developing.
>
> First, the problem with writing to files is that there is no way to
> make these logs reliable. This is what I mean by saying these are
> unreliable: since logs are designed to grow indefinitely, the natural
> response to this design property is log rotation. But, it's
> impossible to reliably rotate a log file. There's always a chance
> that during the rotation some log entries will be written to the file
> past the point of rotation, but prior to the point where the next logs
> volume starts.

What I most often do is use one logfile per day, with the date in the
filename. Then simply delete all files older than 7 days, or 30 days, or
whatever is useful for the task at hand. Not only does that sidestep any
issues with rotating logs, but I also find it's very useful to have the
date in the filename.
> Of course, if you only administer your own computer, and you have low
> single digits programs to run, and their behavior doesn't change
> frequently, and you don't care to drop some records every now and
> then... it's OK to log to files directly from a program. But then you
> aren't really in the sysadmin / infra / ops category, as you are more
> of a hobby enthusiast.
I would not use my scheme for something released to a wider audience.
For in-house software though, I like that I can easily put each
application's logs next to its other data files, and that I don't have
to figure out how to get the system's own log infrastructure to work is
I want it to.
> Finally, if you want your logs to go to a file, and currently, your
> only option is stderr, your shell gives you a really, really simple
> way of redirecting stderr to a file.
I feel this is the worst of both worlds. Now your program doesn't have
any control over filename or log expiration, and neither does your
system's logging infrastructure. You just get one indefinitely growing
log file.

--
"You can fool some of the people all the time, and all of the people some
of the time, but you cannot fool all of the people all of the time."
-- Abraham Lincoln
"You can fool too many of the people too much of the time."
-- James Thurber

Loris Bennett wrote at 2024-11-12 10:00 +0100:
> ...
>However, it strikes me as not immediately obvious that the logging file
>must exist at this point. I can imagine a situation in which I want to
>configure a default log file and create it if it missing.

This is what happens usually:
if you open a file with mode `a` or `w`, the file is created
if it does not yet exist.

Thus, a missing log file should not give you the `FileNotFound`
exception.
Look at the exception details: they should tell you what really
was not found (maybe the directory for the logfile).

On Wed, Nov 13 2024 at 07:36:04 PM, dieter.maurer@online.de wrote:
> Loris Bennett wrote at 2024-11-12 10:00 +0100:
>> ...
>>However, it strikes me as not immediately obvious that the logging file
>>must exist at this point. I can imagine a situation in which I want to
>>configure a default log file and create it if it missing.
>
> This is what happens usually:
> if you open a file with mode `a` or `w`, the file is created
> if it does not yet exist.
>
> Thus, a missing log file should not give you the `FileNotFound`
> exception.
> Look at the exception details: they should tell you what really
> was not found (maybe the directory for the logfile).

It is possible a directory along the path does not exist.

--
regards,
kushal

On 11/12/24 12:10 PM, Left Right via Python-list wrote:
> But, it's
> impossible to reliably rotate a log file. There's always a chance
> that during the rotation some log entries will be written to the file
> past the point of rotation, but prior to the point where the next logs
> volume starts.

On any Unix system this is untrue. Rotating a log file is quite simple:
simply rename the log file, then send a signal to the process to close
the log file handle and open a new one. After that perhaps compress the
rotated log file. Nothing is lost. This is standard practice in Unix.
It is reliable.

Perhaps the scenario you posit would happen on Windows.

> On any Unix system this is untrue. Rotating a log file is quite simple:

I realized I posted this without cc'ing the list:
http://jdebp.info/FGA/do-not-use-logrotate.html .

The link above gives a more detailed description of why log rotation
on the Unix system is not only not simple, but is, in fact,
unreliable.

NB. Also, it really rubs me the wrong way when the word "standard" is
used to mean "common" (instead of "as described in a standard
document"). And when it comes to popular tools, oftentimes "common"
is wrong because commonly the tool is used by amateurs rather than
experts. In other words, you only reinforced what I wrote initially:
plenty of application developers don't know how to do logging well.
It also appears that they would lecture infra / ops people on how to
do something that they aren't experts on, while the latter are :)

On Thu, 14 Nov 2024 at 18:05, Left Right via Python-list
<python-list@python.org> wrote:
>
> > On any Unix system this is untrue. Rotating a log file is quite simple:
>
> I realized I posted this without cc'ing the list:
> http://jdebp.info/FGA/do-not-use-logrotate.html .
>
> The link above gives a more detailed description of why log rotation
> on the Unix system is not only not simple, but is, in fact,
> unreliable.

You're assuming a very specific tool here. Log rotation isn't
necessarily performed by that one tool. There are many ways to do it.

Log to stderr. That puts the power in the hands of the sysadmin,
rather than forcing trickery like setting the log file name to be
/proc/self/fd/2 to get around it.

ChrisA

On 11/13/24 02:12, Roel Schroeven via Python-list wrote:
> What I most often do is use one logfile per day, with the date in the
> filename. Then simply delete all files older than 7 days, or 30 days, or
> whatever is useful for the task at hand. Not only does that sidestep any
> issues with rotating logs, but I also find it's very useful to have the
> date in the filename.

I do something similar for my system logs except that I let the system
use the usual names and, at midnight, I rename the file appending the
previous day's date to it and restarting services.

--
D'Arcy J.M. Cain
Vybe Networks Inc.
http://www.VybeNetworks.com/
IM:darcy@VybeNetworks.com VoIP: sip:darcy@VybeNetworks.com

> On 14 Nov 2024, at 14:07, Loris Bennett via Python-list <python-list@python.org> wrote:
>
> I don't quite understand what your suggestion is. Do you mean that I
> should log to stderr and then run my program as
>
> my_program ... 2>&1 | logger

On almost all Linux distros you would run a long running program as a systemd service and let it put logs into the journal. I wonder if that was what was being hinted at?

Barry

On 11/13/24 23:03, Left Right via Python-list wrote:

>> On any Unix system this is untrue. Rotating a log file is quite simple:
>
> I realized I posted this without cc'ing the list:
> http://jdebp.info/FGA/do-not-use-logrotate.html .
>
> The link above gives a more detailed description of why log rotation
> on the Unix system is not only not simple, but is, in fact,
> unreliable.

Having read the linked article, I see it is not relevant to Python, as Python's logging tool is
the writer/rotator program, thus no window for lost entries exists.

> NB. Also, it really rubs me the wrong way when the word "standard" is
> used to mean "common" (instead of "as described in a standard
> document").

Yes, that is irritating.

> And when it comes to popular tools, oftentimes "common"
> is wrong because commonly the tool is used by amateurs rather than
> experts. In other words, you only reinforced what I wrote initially:
> plenty of application developers don't know how to do logging well.
> It also appears that they would lecture infra / ops people on how to
> do something that they aren't experts on, while the latter are :)

Well, since this is a Python list, perhaps you could make sure your advice is also Python appropriate. I appreciate
diversions into general areas and learning new things, but your general claims were untrue when it comes to Python
specifically, and that was unclear until I read your linked post.

--
~Ethan~

On 11/14/24 12:03 AM, Left Right wrote:
>> On any Unix system this is untrue. Rotating a log file is quite simple:
>
> I realized I posted this without cc'ing the list:
> http://jdebp.info/FGA/do-not-use-logrotate.html .
>
> The link above gives a more detailed description of why log rotation
> on the Unix system is not only not simple, but is, in fact,
> unreliable.

Nothing in that article contradicts what I said about renaming log
files. His argument is that renaming log files messes with tail -F, and
therefore broken and unreliable. Which a pretty strange argument. tail
-F might not see some data during the rotation, but the log files
themselves don't miss anything, which was my contention. In all my
years of sysadmin-ing I have never once worried about problems GNU tail
might have with a file that gets rotated out from under you. Not sure
why the author is so fixated on it.

There are actual legitimate issues at play, such as a mechanism for
informing the process to close the file (rotate usually requires
processes to respond to SIGHUP). And of course the disk can fill up
causing a denial of service of one kind or another. The latter is the
biggest source of problems.

Of course you could just log using the standard libc syslog facility.
Or better yet, start your process from a systemd unit file and let the
journal automatically log stderr. In both cases that would satisfy the
technical objections of the author of that little treatise.

On 2024-11-14, Michael Torrie <torriem@gmail.com> wrote:
> On 11/14/24 12:03 AM, Left Right wrote:
>>> On any Unix system this is untrue. Rotating a log file is quite simple:
>>
>> I realized I posted this without cc'ing the list:
>> http://jdebp.info/FGA/do-not-use-logrotate.html .
>>
>> The link above gives a more detailed description of why log rotation
>> on the Unix system is not only not simple, but is, in fact,
>> unreliable.
>
> Nothing in that article contradicts what I said about renaming log
> files. His argument is that renaming log files messes with tail -F, and
> therefore broken and unreliable. Which a pretty strange argument. tail
> -F might not see some data during the rotation, but the log files
> themselves don't miss anything, which was my contention. In all my
> years of sysadmin-ing I have never once worried about problems GNU tail
> might have with a file that gets rotated out from under you. Not sure
> why the author is so fixated on it.

I really wouldn't worry about anything Jonathan de Boyne Pollard says.