Thank you for raising this topic! I think that vast majority of the message(MESS_ERROR, ...) calls should result in non-zero exit code. The few exceptions could be fixed to either return non-zero code, or by replacing MESS_ERROR by a different level. We could probably introduce MESS_WARNING to print warnings. Then we can easily modify the message() function to always print the error: prefix for MESS_ERROR and warning: for MESS_WARNING, instead of duplicating the prefixes all over the code. Does it make sense?

It does make sense. New push coming shortly.

This pull request becomes messy as there are a lot of commits reverting each other. The resulting diff just adds the debug/info/verbose/error/fatal/unknown prefix, as I understand it. Note that the debug prefix does not work well for example with rotateLogSet(), where it prints something like this:

debug: reading config file lr.conf
debug: Reading state from file: lr.stat
debug: Allocating hash table for state file, size 64 entries
debug: Creating new state
debug: 
Handling 1 logs
debug: 
rotating pattern: /tmp/xxx  debug: 1048576 bytes debug: (1 rotations)
debug: empty log files are rotated, debug: old logs are removed
debug: considering log /tmp/xxx
debug:   Now: 2019-11-04 13:26
debug:   Last rotated at 2019-11-04 13:00
debug:   log does not need rotating (log size is below the 'size' threshold)
debug: set default create context to unconfined_u:object_r:user_home_t:s0

@kdudka I apologize for the messy commit history, I’ve been working in Go mostly these days, and didn’t know how to best build and run test locally. I ended up having a couple of script locally.

I hadn’t considered how it would pollute the output. I hadn’t been concerned with that since I’m used to writing syslog formatted messages so the log level being on each line feels normal but I understand the need to minimize behavior changes. Perhaps an additional flag for whether or not to pretend with debug: then entire blocks would look more like

debug: reading config file lr.conf
       Reading state from file: lr.stat
       Allocating hash table for state file, size 64 entries
       Creating new state
       
Handling 1 logs
       
rotating pattern: /tmp/xxx  debug: 1048576 bytes debug: (1 rotations)
       empty log files are rotated, debug: old logs are removed
      considering log /tmp/xxx
         Now: 2019-11-04 13:26
         Last rotated at 2019-11-04 13:00
         log does not need rotating (log size is below the 'size' threshold)
       set default create context to unconfined_u:object_r:user_home_t:s0 

That way each time a series of debug level logs comes through, it is only pretended once?

Can't we just use the prefix for error messages only? Those affect logrotate's exit code after all...

Not sure if this is related place to post, but it seems that destination /my/rotated/file/new-name already exists, skipping rotation and log does not need rotating (log size is below the 'size' threshold) messages both goes to the stderr.

First type will trigger RC 1, while second type will return RC 0. Perhaps it would be possible to either:

Place output to stdout which does not trigger RC 0, and/or separate error messages in stderr to determine which ones does trigger RC 1.

Not sure if this is related place to post, but it seems that destination /my/rotated/file/new-name already exists, skipping rotation and log does not need rotating (log size is below the 'size' threshold) messages both goes to the stderr.

First type will trigger RC 1, while second type will return RC 0. Perhaps it would be possible to either:

Place output to stdout which does not trigger RC 0, and/or separate error messages in stderr to determine which ones does trigger RC 1.

Currently all messages are printed to stderr