summaryrefslogtreecommitdiffstats
path: root/docs/manual/mod/mod_unique_id.xml
blob: df76ae7b8fee186cb9f4e766dfa33dc08709a8bc (plain)
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
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
<?xml version="1.0"?>
<!DOCTYPE modulesynopsis SYSTEM "../style/modulesynopsis.dtd">
<?xml-stylesheet type="text/xsl" href="../style/manual.en.xsl"?>
<!-- $LastChangedRevision$ -->

<!--
 Licensed to the Apache Software Foundation (ASF) under one or more
 contributor license agreements.  See the NOTICE file distributed with
 this work for additional information regarding copyright ownership.
 The ASF licenses this file to You under the Apache License, Version 2.0
 (the "License"); you may not use this file except in compliance with
 the License.  You may obtain a copy of the License at

     http://www.apache.org/licenses/LICENSE-2.0

 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
-->

<modulesynopsis metafile="mod_unique_id.xml.meta">

<name>mod_unique_id</name>
<description>Provides an environment variable with a unique
identifier for each request</description>
<status>Extension</status>
<sourcefile>mod_unique_id.c</sourcefile>
<identifier>unique_id_module</identifier>

<summary>

    <p>This module provides a magic token for each request which is
    guaranteed to be unique across "all" requests under very
    specific conditions. The unique identifier is even unique
    across multiple machines in a properly configured cluster of
    machines. The environment variable <code>UNIQUE_ID</code> is
    set to the identifier for each request. Unique identifiers are
    useful for various reasons which are beyond the scope of this
    document.</p>
</summary>

<section id="theory">
    <title>Theory</title>

    <p>First a brief recap of how the Apache server works on Unix
    machines. This feature currently isn't supported on Windows NT.
    On Unix machines, Apache creates several children, the children
    process requests one at a time. Each child can serve multiple
    requests in its lifetime. For the purpose of this discussion,
    the children don't share any data with each other. We'll refer
    to the children as <dfn>httpd processes</dfn>.</p>

    <p>Your website has one or more machines under your
    administrative control, together we'll call them a cluster of
    machines. Each machine can possibly run multiple instances of
    Apache. All of these collectively are considered "the
    universe", and with certain assumptions we'll show that in this
    universe we can generate unique identifiers for each request,
    without extensive communication between machines in the
    cluster.</p>

    <p>The machines in your cluster should satisfy these
    requirements. (Even if you have only one machine you should
    synchronize its clock with NTP.)</p>

    <ul>
      <li>The machines' times are synchronized via NTP or other
      network time protocol.</li>

      <li>The machines' hostnames all differ, such that the module
      can do a hostname lookup on the hostname and receive a
      different IP address for each machine in the cluster.</li>
    </ul>

    <p>As far as operating system assumptions go, we assume that
    pids (process ids) fit in 32-bits. If the operating system uses
    more than 32-bits for a pid, the fix is trivial but must be
    performed in the code.</p>

    <p>Given those assumptions, at a single point in time we can
    identify any httpd process on any machine in the cluster from
    all other httpd processes. The machine's IP address and the pid
    of the httpd process are sufficient to do this. A httpd process
    can handle multiple requests simultaneously if you use a
    multi-threaded MPM. In order to identify threads, we use a thread
    index Apache httpd uses internally. So in order to
    generate unique identifiers for requests we need only
    distinguish between different points in time.</p>

    <p>To distinguish time we will use a Unix timestamp (seconds
    since January 1, 1970 UTC), and a 16-bit counter. The timestamp
    has only one second granularity, so the counter is used to
    represent up to 65536 values during a single second. The
    quadruple <em>( ip_addr, pid, time_stamp, counter )</em> is
    sufficient to enumerate 65536 requests per second per httpd
    process. There are issues however with pid reuse over time, and
    the counter is used to alleviate this issue.</p>

    <p>When an httpd child is created, the counter is initialized
    with ( current microseconds divided by 10 ) modulo 65536 (this
    formula was chosen to eliminate some variance problems with the
    low order bits of the microsecond timers on some systems). When
    a unique identifier is generated, the time stamp used is the
    time the request arrived at the web server. The counter is
    incremented every time an identifier is generated (and allowed
    to roll over).</p>

    <p>The kernel generates a pid for each process as it forks the
    process, and pids are allowed to roll over (they're 16-bits on
    many Unixes, but newer systems have expanded to 32-bits). So
    over time the same pid will be reused. However unless it is
    reused within the same second, it does not destroy the
    uniqueness of our quadruple. That is, we assume the system does
    not spawn 65536 processes in a one second interval (it may even
    be 32768 processes on some Unixes, but even this isn't likely
    to happen).</p>

    <p>Suppose that time repeats itself for some reason. That is,
    suppose that the system's clock is screwed up and it revisits a
    past time (or it is too far forward, is reset correctly, and
    then revisits the future time). In this case we can easily show
    that we can get pid and time stamp reuse. The choice of
    initializer for the counter is intended to help defeat this.
    Note that we really want a random number to initialize the
    counter, but there aren't any readily available numbers on most
    systems (<em>i.e.</em>, you can't use rand() because you need
    to seed the generator, and can't seed it with the time because
    time, at least at one second resolution, has repeated itself).
    This is not a perfect defense.</p>

    <p>How good a defense is it? Suppose that one of your machines
    serves at most 500 requests per second (which is a very
    reasonable upper bound at this writing, because systems
    generally do more than just shovel out static files). To do
    that it will require a number of children which depends on how
    many concurrent clients you have. But we'll be pessimistic and
    suppose that a single child is able to serve 500 requests per
    second. There are 1000 possible starting counter values such
    that two sequences of 500 requests overlap. So there is a 1.5%
    chance that if time (at one second resolution) repeats itself
    this child will repeat a counter value, and uniqueness will be
    broken. This was a very pessimistic example, and with real
    world values it's even less likely to occur. If your system is
    such that it's still likely to occur, then perhaps you should
    make the counter 32 bits (by editing the code).</p>

    <p>You may be concerned about the clock being "set back" during
    summer daylight savings. However this isn't an issue because
    the times used here are UTC, which "always" go forward. Note
    that x86 based Unixes may need proper configuration for this to
    be true -- they should be configured to assume that the
    motherboard clock is on UTC and compensate appropriately. But
    even still, if you're running NTP then your UTC time will be
    correct very shortly after reboot.</p>

    <!-- FIXME: thread_index is unsigned int, so not always 32bit.-->
    <p>The <code>UNIQUE_ID</code> environment variable is
    constructed by encoding the 144-bit (32-bit IP address, 32 bit
    pid, 32 bit time stamp, 16 bit counter, 32 bit thread index)
    quadruple using the
    alphabet <code>[A-Za-z0-9@-]</code> in a manner similar to MIME
    base64 encoding, producing 24 characters. The MIME base64
    alphabet is actually <code>[A-Za-z0-9+/]</code> however
    <code>+</code> and <code>/</code> need to be specially encoded
    in URLs, which makes them less desirable. All values are
    encoded in network byte ordering so that the encoding is
    comparable across architectures of different byte ordering. The
    actual ordering of the encoding is: time stamp, IP address,
    pid, counter. This ordering has a purpose, but it should be
    emphasized that applications should not dissect the encoding.
    Applications should treat the entire encoded
    <code>UNIQUE_ID</code> as an opaque token, which can be
    compared against other <code>UNIQUE_ID</code>s for equality
    only.</p>

    <p>The ordering was chosen such that it's possible to change
    the encoding in the future without worrying about collision
    with an existing database of <code>UNIQUE_ID</code>s. The new
    encodings should also keep the time stamp as the first element,
    and can otherwise use the same alphabet and bit length. Since
    the time stamps are essentially an increasing sequence, it's
    sufficient to have a <em>flag second</em> in which all machines
    in the cluster stop serving any request, and stop using the old
    encoding format. Afterwards they can resume requests and begin
    issuing the new encodings.</p>

    <p>This we believe is a relatively portable solution to this
    problem. The identifiers
    generated have essentially an infinite life-time because future
    identifiers can be made longer as required. Essentially no
    communication is required between machines in the cluster (only
    NTP synchronization is required, which is low overhead), and no
    communication between httpd processes is required (the
    communication is implicit in the pid value assigned by the
    kernel). In very specific situations the identifier can be
    shortened, but more information needs to be assumed (for
    example the 32-bit IP address is overkill for any site, but
    there is no portable shorter replacement for it). </p>
</section>


</modulesynopsis>