So, just to finish off what I’ve learned about newsyslog on Mac OS X…
I’m running standard Apache 2.2 that comes with Mac OS X 10.6. I like to keep all my virtual hosts in one place, so in this example they are all in /Users/web/. Each host gets a directory structure with it’s own name:
# ls -l www.example.org/ drwxrwxr-x 4 _unknown _unknown 136 Jan 27 16:47 htdocs drwxrwxr-x 80 _www _www 2720 Mar 29 00:33 logs -rw-r--r-- 1 adoyle web 1881 Feb 20 15:44 www.example.org.conf
The logs for each virtual host go into the logs directory for that host (access_log, error_log, rewrite_log). Ownership on the log files turns out to be important. I’ve found it works best for me if they are owned by the www user and group (or _www, they are essentially the same – something I need to understand the reason for someday).
In /etc/newsyslog.d/local.conf, the following lines deal with rotating logs for three virtual hosts. Using the ‘G’ flag lets you use ‘*’ and other shell wildcards in the file names. I think I could probably have collapsed these into a single line if I had used /Users/web/*/logs/*log instead. In this case, folding things up too much makes it less readable, I think.
The _www:_www takes care of preserving the file ownership after the logs are rotated. A count of 30 means keep around up to 30 old logs. $D0 means rotate daily at 0:00. The ‘B’ flag prevents the “Log file was rotated” message. Apache keeps a pidfile in /var/run/httpd.pid. If you send a kill -30 to the pid in that file, it will cause the equivalent of an ‘apachectl graceful’.
If you don’t provide the pidfile and proper signal number, the logs will rotate, new log files get created, but Apache won’t write to them because it’s still trying to write to the old ones.
# logfilename [owner:group] mode count size when flags [/pid_file] [sig_num] /Users/web/lists.example.org/logs/*log _www:_www 664 30 * $D0 GBJ /var/run/httpd.pid 30 /Users/web/foo.example.org/logs/*log _www:_www 664 30 * $D0 GBJ /var/run/httpd.pid 30 /Users/web/www.example.org/logs/*log _www:_www 664 30 * $D0 GBJ /var/run/httpd.pid 30 #
With Mailman I had a slight problem. Mailman doesn’t use nice .log or _log names. It just uses names like bounce, error, post, qrunner, etc. I could have made an entry in the local.conf file for each one, but that seemed error-prone. What if later there’s a new version of Mailman that generates different log files?
My initial assumption was that I could use …/logs/* and newsyslog wouldn’t try to rotate logs it had already rotated. Guess again. After two days, my disk had nearly filled up with files ending in .bz2, .bz2.bz2, .bz2.bz2.b2z…, you get the picture. Luckily I noticed it before the disk did fill up. I got suspicious when my backups on the third day were 15GB bigger than the ones on the first day. I have no idea how many files actually got created. ‘ls’ was unable to produce a listing in the amount of time I was willing to wait. Luckily rm -rf did work. It took several hours to delete all the files.
I’m assuming newsyslog had gotten into a recursive loop right away when it ran at midnight and never stopped churning out files until I killed it.
So anyway, the moral of the story is, if you’re going to use wildcards, make sure they don’t match the rotated logs.
# logfilename [owner:group] mode count size when flags [/pid_file] [sig_num] /Users/mailman/logs/*[a-z] mailman:_www 664 30 * $D0 GBJ /Users/mailman/data/master-qrunner.pid 1 #
Mailman wants to be hit with a kill -1, so that’s what I used.
Things have been noodling along for a few weeks with my setup, so I think I have the kinks ironed out.