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 httpd processes.
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.
The machines in your cluster should satisfy these
requirements. (Even if you have only one machine you should
synchronize its clock with NTP.)
- The machines' times are synchronized via NTP or other
network time protocol.
- 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.
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.
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. So in order to
generate unique identifiers for requests we need only
distinguish between different points in time.
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 ( ip_addr, pid, time_stamp, counter ) 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.
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).
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).
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 (i.e., 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.
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).
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.
The UNIQUE_ID environment variable is
constructed by encoding the 112-bit (32-bit IP address, 32 bit
pid, 32 bit time stamp, 16 bit counter) quadruple using the
alphabet [A-Za-z0-9@-] in a manner similar to MIME
base64 encoding, producing 19 characters. The MIME base64
alphabet is actually [A-Za-z0-9+/] however
+ and / 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
UNIQUE_ID as an opaque token, which can be
compared against other UNIQUE_IDs for equality
only.
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 UNIQUE_IDs. 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 flag second in which all machines
in the cluster stop serving and request, and stop using the old
encoding format. Afterwards they can resume requests and begin
issuing the new encodings.
This we believe is a relatively portable solution to this
problem. It can be extended to multithreaded systems like
Windows NT, and can grow with future needs. 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).
