1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
|
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html><head><title>Compatibility notes for rsyslog v4</title>
</head>
<body>
<h1>Compatibility Notes for rsyslog v4</h1>
<p><small><i>Written by <a href="http://www.gerhards.net/rainer">Rainer Gerhards</a>
(2009-07-15)</i></small></p>
<p>The changes introduced in rsyslog v4 are numerous, but not very intrusive.
This document describes things to keep in mind when moving from v3 to v4. It
does not list enhancements nor does it talk about compatibility concerns introduced
by v3 (for this, see the <a href="v3compatibility.html">rsyslog v3 compatibility notes</a>).
<h2>HUP processing</h2>
<p>With v3 and below, rsyslog used the traditional HUP behaviour. That meant that
all output files are closed and the configuration file is re-read and the new configuration
applied.
<p>With a program as simple and static as sysklogd, this was not much of an issue. The
most important config settings (like udp reception) of a traditional syslogd can not be
modified via the configuration file. So a config file reload only meant setting up a new set of filters. It also didn't account as problem that while doing so messages may be lost - without
any threading and queuing model, a traditional syslogd will potentially always loose
messages, so it is irrelevant if this happens, too, during the short config re-read
phase.
<p>In rsyslog, things are quite different: the program is more or less a framework into
which loadable modules are loaded as needed for a particular configuration. The software
that will acutally be running is taylored via the config file. Thus, a re-read of
the config file requires a full, very heavy restart, because the software acutally
running with the new config can be totally different from what ran with the old config.
<p>Consequently, the traditional HUP is a very heavy operation and may even cause some
data loss because queues must be shut down, listeners stopped and so on. Some of these
operations (depending on their configuration) involve intentional message loss. The operation
also takes up a lot of system resources and needs quite some time (maybe seconds) to be
completed. During this restart period, the syslog subsytem is not fully available.
<p>From the software developer's point of view, the full restart done by a HUP is rather complex,
especially if user-timeout limits set on action completion are taken into consideration (for
those in the know: at the extreme ends this means we need to cancel threads as a last resort,
but than we need to make sure that such cancellation does not happen at points where it
would be fatal for a restart). A regular restart, where the process is actually terminated, is
much less complex, because the operating system does a full cleanup after process termination,
so rsyslogd does not need to take care for exotic cleanup cases and leave that to the OS.
In the end result, restart-type HUPs clutter the code, increase complexity (read: add bugs)
and cost performance.
<p>On the contrary, a HUP is typically needed for log rotation, and the real desire is
to close files. This is a non-disruptive and very lightweigth operation.
<p>Many people have said that they are used to HUP the syslogd to apply configuration
changes. This is true, but it is questionable if that really justifies all the cost that
comes with it. After all, it is the difference between typing
<pre>
$ kill -HUP `cat /var/run/rsyslogd.pid`
</pre>
versus
<pre>
$ /etc/init.d/rsyslog restart
</pre>
Semantically, both is mostly the same thing. The only difference is that with the restart
command rsyslogd can spit config error message to stderr, so that the user is able to see
any problems and fix them. With a HUP, we do not have access to stderr and thus can log
error messages only to their configured destinations; exprience tells that most users
will never find them there. What, by the way, is another strong argument against
restarting rsyslogd by HUPing it.
<p>So a restart via HUP is not strictly necessary
and most other deamons require that a restart command is typed in if a restart is required.
<p>Rsyslog will follow this paradigm in the next versions, resulting in many benefits. In v4,
we provide some support for the old-style semantics. We introduced a setting $HUPisRestart
which may be set to "on" (tradional, heavy operationg)
or "off" (new, lightweight "file close only" operation).
The initial versions had the default set to traditional behavior, but starting with 4.5.1
we are now using the new behavior as the default.
<p>Most importantly, <b>this may break some scripts</b>, but my sincere belief is that
there are very few scripts that automatically <b>change</b> rsyslog's config and then do a
HUP to reload it. Anyhow, if you have some of these, it may be a good idea to change
them now instead of turning restart-type HUPs on. Other than that, one mainly needs
to change the habit of how to restart rsyslog after a configuration change.
<p><b>Please note that restart-type HUP is depricated and will go away in rsyslog v5.</b>
So it is a good idea to become ready for the new version now and also enjoy some of the
benefits of the "real restart", like the better error-reporting capability.
<p>Note that code complexity reduction (and thus performance improvement) needs the restart-type
HUP code to be removed, so these changes can (and will) only happen in version 5.
</body></html>
|
