[rrdtool.git] / doc / rrdcached.pod

=pod

=head1 NAME

rrdcached - Data caching daemon for rrdtool

=head1 SYNOPSIS

B<rrdcached>
[B<-P>E<nbsp>I<permissions>]
[B<-l>E<nbsp>I<address>]
[B<-w>E<nbsp>I<timeout>]
[B<-z>E<nbsp>I<delay>]
[B<-f>E<nbsp>I<timeout>]
[B<-p>E<nbsp>I<pid_file>]
[B<-t>E<nbsp>I<write_threads>]
[B<-j>E<nbsp>I<journal_dir>]
[-F]
[-g]
[B<-b>E<nbsp>I<base_dir>E<nbsp>[B<-B>]]

=head1 DESCRIPTION

B<rrdcached> is a daemon that receives updates to existing RRD files,
accumulates them and, if enough have been received or a defined time has
passed, writes the updates to the RRD file. A I<flush> command may be used to
force writing of values to disk, so that graphing facilities and similar can
work with up-to-date data.

The daemon was written with big setups in mind. Those setups usually run into
IOE<nbsp>related problems sooner or later for reasons that are beyond the scope
of this document. Check the wiki at the RRDTool homepage for details. Also
check L<SECURITY CONSIDERATIONS> below before using this daemon! A detailed
description of how the daemon operates can be found in the L<HOW IT WORKS>
section below.

=head1 OPTIONS

=over 4

=item B<-l> I<address>

Tells the daemon to bind to I<address> and accept incoming connections on that
socket. If I<address> begins with C<unix:>, everything following that prefix is
interpreted as the path to a UNIX domain socket. Otherwise the address or node
name are resolved using L<getaddrinfo>.

For network sockets, a port may be specified by using the form
C<B<[>I<address>B<]:>I<port>>. If the address is an IPv4 address or a fully
qualified domain name (i.E<nbsp>e. the address contains at least one dot
(C<.>)), the square brackets can be omitted, resulting in the (simpler)
C<I<address>B<:>I<port>> pattern. The default port is B<42217/udp>. If you
specify a network socket, it is mandatory to read the
L</"SECURITY CONSIDERATIONS"> section.

The following formats are accepted. Please note that the address of the UNIX
domain socket B<must> start with a slash in the second case!

   unix:</path/to/unix.sock>
   /<path/to/unix.sock>
   <hostname-or-ip>
   [<hostname-or-ip>]:<port>
   <hostname-or-ipv4>:<port>

If the B<-l> option is not specified the default address,
C<unix:/tmp/rrdcached.sock>, will be used.

=item B<-P> I<command>[,I<command>[,...]]

Specifies the commands accepted via a network socket. This allows
administrators of I<RRDCacheD> to control the actions accepted from various
sources.

The arguments given to the B<-P> option is a comma separated list of commands.
For example, to allow the C<FLUSH> and C<PENDING> commands one could specify:

  rrdcached -P FLUSH,PENDING $MORE_ARGUMENTS

The B<-P> option effects the I<following> socket addresses (the following B<-l>
options). In the following example, only the IPv4 network socket (address
C<10.0.0.1>) will be restricted to the C<FLUSH> and C<PENDING> commands:

  rrdcached -l unix:/some/path -P FLUSH,PENDING -l 10.0.0.1

A complete list of available commands can be found in the section
L</"Valid Commands"> below. There are two minor special exceptions:

=over 4

=item *

The C<HELP> and C<QUIT> commands are always allowed.

=item *

If the C<BATCH> command is accepted, the B<.>E<nbsp>command will automatically
be accepted, too.

=back

Please also read L</"SECURITY CONSIDERATIONS"> below.

=item B<-w> I<timeout>

Data is written to disk every I<timeout> seconds. If this option is not
specified the default interval of 300E<nbsp>seconds will be used.

=item B<-z> I<delay>

If specified, rrdcached will delay writing of each RRD for a random number
of seconds in the rangeE<nbsp>[0,I<delay>).  This will avoid too many
writes being queued simultaneously.  This value should be no greater than
the value specified in B<-w>.  By default, there is no delay.

=item B<-f> I<timeout>

Every I<timeout> seconds the entire cache is searched for old values which are
written to disk. This only concerns files to which updates have stopped, so
setting this to a high value, such as 3600E<nbsp>seconds, is acceptable in most
cases. This timeout defaults to 3600E<nbsp>seconds.

=item B<-p> I<file>

Sets the name and location of the PID-file. If not specified, the default,
C<I<$localststedir>/run/rrdcached.pid> will be used.

=item B<-t> I<write_threads>

Specifies the number of threads used for writing RRD files.  The default
isE<nbsp>4.  Increasing this number will allow rrdcached to have more
simultaneous I/O requests into the kernel.  This may allow the kernel to
re-order disk writes, resulting in better disk throughput.

=item B<-j> I<dir>

Write updates to a journal in I<dir>.  In the event of a program or system
crash, this will allow the daemon to write any updates that were pending
at the time of the crash.

On startup, the daemon will check for journal files in this directory.  If
found, all updates therein will be read into memory before the daemon
starts accepting new connections.

The journal will be rotated with the same frequency as the flush timer
given by B<-f>.

When journaling is enabled, the daemon will use a fast shutdown procedure.
Rather than flushing all files to disk, it will make sure the journal is
properly written and exit immediately.  Although the RRD data files are
not fully up-to-date, no information is lost; all pending updates will be
replayed from the journal next time the daemon starts up.

To disable fast shutdown, use the B<-F> option.

=item B<-F>

ALWAYS flush all updates to the RRD data files when the daemon is shut
down, regardless of journal setting.

=item B<-g>

Run in the foreground.  The daemon will not fork().

=item B<-b> I<dir>

The daemon will change into a specific directory at startup. All files passed
to the daemon, that are specified by a B<relative> path, will be interpreted
to be relative to this directory. If not given the default, C</tmp>, will be
used.

  +------------------------+------------------------+
  ! Command line           ! File updated           !
  +------------------------+------------------------+
  ! foo.rrd                ! /tmp/foo.rrd           !
  ! foo/bar.rrd            ! /tmp/foo/bar.rrd       !
  ! /var/lib/rrd/foo.rrd   ! /var/lib/rrd/foo.rrd   !
  +------------------------+------------------------+
  Paths given on the command  line and paths actually
  updated by the daemon,  assuming the base directory
  "/tmp".

B<WARNING:> The paths up to and including the base directory B<MUST NOT BE>
symbolic links.  In other words, if the base directory is
specified as:

    -b /base/dir/somewhere

... then B<NONE> of the following should be symbolic links:

    /base
    /base/dir
    /base/dir/somewhere

=item B<-B>

Only permit writes into the base directory specified in B<-b> (and any
sub-directories).  This does B<NOT> detect symbolic links.  Paths
containing C<../> will also be blocked.

=back

=head1 AFFECTED RRDTOOL COMMANDS

The following commands may be made aware of the B<rrdcached> using the command
line argument B<--daemon> or the environment variable B<RRDCACHED_ADDRESS>:

=over 4

=item B<dump>

=item B<fetch>

=item B<flush>

=item B<graph>

=item B<graphv>

=item B<info>

=item B<last>

=item B<lastupdate>

=item B<update>

=item B<xport>

=back

The B<update> command can send values to the daemon instead of writing them to
the disk itself. All other commands can send a B<FLUSH> command (see below) to
the daemon before accessing the files, so they work with up-to-date data even
if the cache timeout is large.

=head1 ERROR REPORTING

The daemon reports errors in one of two ways: During startup, error messages
are printed to C<STDERR>. One of the steps when starting up is to fork to the
background and closing C<STDERR> - after this writing directly to the user is
no longer possible. Once this has happened, the daemon will send log messages
to the system logging daemon using L<syslog(3)>. The facility used is
C<LOG_DAEMON>.

=head1 HOW IT WORKS

When receiving an update, B<rrdcached> does not write to disk but looks for an
entry for that file in its internal tree. If not found, an entry is created
including the current time (called "First" in the diagram below). This time is
B<not> the time specified on the command line but the time the operating system
considers to be "now". The value and time of the value (called "Time" in the
diagram below) are appended to the tree node.

When appending a value to a tree node, it is checked whether it's time to write
the values to disk. Values are written to disk if
S<C<now() - First E<gt>= timeout>>, where C<timeout> is the timeout specified
using the B<-w> option, see L<OPTIONS>. If the values are "old enough" they
will be enqueued in the "update queue", i.E<nbsp>e. they will be appended to
the linked list shown below.  Because the tree nodes and the elements of the
linked list are the same data structures in memory, any update to a file that
has already been enqueued will be written with the next write to the RRD file,
too.

A separate "update thread" constantly dequeues the first element in the update
queue and writes all its values to the appropriate file. So as long as the
update queue is not empty files are written at the highest possible rate.

Since the timeout of files is checked only when new values are added to the
file, "dead" files, i.E<nbsp>e. files that are not updated anymore, would never
be written to disk. Therefore, every now and then, controlled by the B<-f>
option, the entire tree is walked and all "old" values are enqueued. Since this
only affects "dead" files and walking the tree is relatively expensive, you
should set the "flush interval" to a reasonably high value. The default is
3600E<nbsp>seconds (one hour).

The downside of caching values is that they won't show up in graphs generated
from the RRDE<nbsp>files. To get around this, the daemon provides the "flush
command" to flush specific files. This means that the file is inserted at the
B<head> of the update queue or moved there if it is already enqueued. The flush
command will return only after the file's pending updates have been written
to disk.

 +------+   +------+                               +------+
 ! head !   ! root !                               ! tail !
 +---+--+   +---+--+                               +---+--+
     !         /\                                      !
     !        /  \                                     !
     !       /\  /\                                    !
     !      /\/\ \ `----------------- ... --------,    !
     V     /      `-------,                       !    V
 +---+----+---+    +------+-----+             +---+----+---+
 ! File:  foo !    ! File:  bar !             ! File:  qux !
 ! First: 101 !    ! First: 119 !             ! First: 180 !
 ! Next:&bar -+--->! Next:&... -+---> ... --->! Next:NULL  !
 | Prev:NULL  !<---+-Prev:&foo  !<--- ... ----+-Prev: &... !
 +============+    +============+             +============+
 ! Time:  100 !    ! Time:  120 !             ! Time:  180 !
 ! Value:  10 !    ! Value: 0.1 !             ! Value: 2,2 !
 +------------+    +------------+             +------------+
 ! Time:  110 !    ! Time:  130 !             ! Time:  190 !
 ! Value:  26 !    ! Value: 0.1 !             ! Value: 7,3 !
 +------------+    +------------+             +------------+
 :            :    :            :             :            :
 +------------+    +------------+             +------------+
 ! Time:  230 !    ! Time:  250 !             ! Time:  310 !
 ! Value:  42 !    ! Value: 0.2 !             ! Value: 1,2 !
 +------------+    +------------+             +------------+

The above diagram demonstrates:

=over

=item *

Files/values are stored in a (balanced) tree.

=item *

Tree nodes and entries in the update queue are the same data structure.

=item *

The local time ("First") and the time specified in updates ("Time") may differ.  

=item *

Timed out values are inserted at the "tail".

=item *

Explicitly flushed values are inserted at the "head".

=item *

ASCII art rocks.

=back

=head1 SECURITY CONSIDERATIONS

=head2 Authentication

There is no authentication.

The client/server protocol does not yet have any authentication mechanism. It
is likely that authentication and encryption will be added in a future version,
but for the time being it is the administrator's responsibility to secure the
traffic from/to the daemon!

It is highly recommended to install a packet filter or similar mechanism to
prevent unauthorized connections. Unless you have a dedicated VLAN or VPN for
this, using network sockets is probably a bad idea!

=head2 Authorization

There is minimal per-socket authorization.

Authorization is currently done on a per-socket basis. That means each socket
has a list of commands it will accept and it will accept. It will accept only
those commands explicitly listed but it will (currently) accept these commands
from anyone reaching the socket.

If the networking sockets are to be used, it is necessary to restrict the
accepted commands to those needed by external clients. If, for example,
external clients want to draw graphs of the cached data, they should only be
allowed to use the C<FLUSH> command.

=head2 Encryption

There is no encryption.

Again, this may be added in the future, but for the time being it is your job
to keep your private data private. Install a VPN or an encrypted tunnel if you
statistics are confidential!

=head2 Sanity checking

There is no sanity checking.

The daemon will blindly write to any file it gets told, so you really should
create a separate user just for this daemon. Also it does not do any sanity
checks, so if it gets told to write values for a time far in the future, your
files will be messed up good!

=head2 Conclusion

=over 4

=item *

Security is the job of the administrator.

=item *

We recommend to allow write access via UNIX domain sockets only.

=item *

You have been warned.

=back

=head1 PROTOCOL

The daemon communicates with clients using a line based ASCII protocol which is
easy to read and easy to type. This makes it easy for scripts to implement the
protocol and possible for users to use L<telnet> to connect to the daemon
and test stuff "by hand".

The protocol is line based, this means that each record consists of one or more
lines. A line is terminated by the line feed character C<0x0A>, commonly
written as C<\n>. In the examples below, this character will be written as
C<E<lt>LFE<gt>> ("line feed").

After the connection has been established, the client is expected to send a
"command". A command consists of the command keyword, possibly some arguments,
and a terminating newline character. For a list of commands, see
L<Valid Commands> below.

Example:

  FLUSH /tmp/foo.rrd<LF>

The daemon answers with a line consisting of a status code and a short status
message, separated by one or more space characters. A negative status code
signals an error, a positive status code or zero signal success. If the status
code is greater than zero, it indicates the number of lines that follow the
status line.

Examples:

 0 Success<LF>

 2 Two lines follow<LF>
 This is the first line<LF>
 And this is the second line<LF>

=head2 Valid Commands

The following commands are understood by the daemon:

=over 4

=item B<FLUSH> I<filename>

Causes the daemon to put I<filename> to the B<head> of the update queue
(possibly moving it there if the node is already enqueued). The answer will be
sent B<after> the node has been dequeued.

=item B<FLUSHALL>

Causes the daemon to start flushing ALL pending values to disk.  This
returns immediately, even though the writes may take a long time.

=item B<PENDING> I<filename>

Shows any "pending" updates for a file, in order.  The updates shown have
not yet been written to the underlying RRD file.

=item B<FORGET> I<filename>

Removes I<filename> from the cache.  Any pending updates B<WILL BE LOST>.

=item B<QUEUE>

Shows the files that are on the output queue.  Returns zero or more lines
in the following format, where E<lt>num_valsE<gt> is the number of values
to be written for the E<lt>fileE<gt>:

    <num_vals> <file>

=item B<HELP> [I<command>]

Returns a short usage message. If no command is given, or I<command> is
B<HELP>, a list of commands supported by the daemon is returned. Otherwise a
short description, possibly containing a pointer to a manual page, is returned.
Obviously, this is meant for interactive usage and the format in which the
commands and usage summaries are returned is not well defined.

=item B<STATS>

Returns a list of metrics which can be used to measure the daemons performance
and check its status. For a description of the values returned, see
L<Performance Values> below.

The format in which the values are returned is similar to many other line based
protocols: Each value is printed on a separate line, each consisting of the
name of the value, a colon, one or more spaces and the actual value.

Example:

 9 Statistics follow
 QueueLength: 0
 UpdatesReceived: 30
 FlushesReceived: 2
 UpdatesWritten: 13
 DataSetsWritten: 390
 TreeNodesNumber: 13
 TreeDepth: 4
 JournalBytes: 190
 JournalRotate: 0

=item B<UPDATE> I<filename> I<values> [I<values> ...]

Adds more data to a filename. This is B<the> operation the daemon was designed
for, so describing the mechanism again is unnecessary. Read L<HOW IT WORKS>
above for a detailed explanation.

Note that rrdcached only accepts absolute timestamps in the update values.
Updates strings like "N:1:2:3" are automatically converted to absolute
time by the RRD client library before sending to rrdcached.

=item B<WROTE> I<filename>

This command is written to the journal after a file is successfully
written out to disk.  It is used during journal replay to determine which
updates have already been applied.  It is I<only> valid in the journal; it
is not accepted from the other command channels.

=item B<BATCH>

This command initiates the bulk load of multiple commands.  This is
designed for installations with extremely high update rates, since it
permits more than one command to be issued per read() and write().

All commands are executed just as they would be if given individually,
except for output to the user.  Messages indicating success are
suppressed, and error messages are delayed until the client is finished.

Command processing is finished when the client sends a dot (".") on its
own line.  After the client has finished, the server responds with an
error count and the list of error messages (if any).  Each error messages
indicates the number of the command to which it corresponds, and the error
message itself.  The first user command after B<BATCH> is command number one.

    client:  BATCH
    server:  0 Go ahead.  End with dot '.' on its own line.
    client:  UPDATE x.rrd 1223661439:1:2:3            <--- command #1
    client:  UPDATE y.rrd 1223661440:3:4:5            <--- command #2
    client:  and so on...
    client:  .
    server:  2 Errors
    server:  1 message for command 1
    server:  12 message for command 12

=item B<QUIT>

Disconnect from rrdcached.

=back

=head2 Performance Values

The following counters are returned by the B<STATS> command:

=over 4

=item B<QueueLength> I<(unsigned 64bit integer)>

Number of nodes currently enqueued in the update queue.

=item B<UpdatesReceived> I<(unsigned 64bit integer)>

Number of UPDATE commands received.

=item B<FlushesReceived> I<(unsigned 64bit integer)>

Number of FLUSH commands received.

=item B<UpdatesWritten> I<(unsigned 64bit integer)>

Total number of updates, i.E<nbsp>e. calls to C<rrd_update_r>, since the
daemon was started.

=item B<DataSetsWritten> I<(unsigned 64bit integer)>

Total number of "data sets" written to disk since the daemon was
started. A data set is one or more values passed to the B<UPDATE>
command. For example: C<1223661439:123:456> is one data set with two
values. The term "data set" is used to prevent confusion whether
individual values or groups of values are counted.

=item B<TreeNodesNumber> I<(unsigned 64bit integer)>

Number of nodes in the cache.

=item B<TreeDepth> I<(unsigned 64bit integer)>

Depth of the tree used for fast key lookup.

=item B<JournalBytes> I<(unsigned 64bit integer)>

Total number of bytes written to the journal since startup.

=item B<JournalRotate> I<(unsigned 64bit integer)>

Number of times the journal has been rotated since startup.

=back

=head1 SIGNALS

=over 4

=item SIGINT and SIGTERM

The daemon exits normally on receipt of either of these signals.  Pending
updates are handled in accordance with the B<-j> and B<-F> options.

=item SIGUSR1

The daemon exits AFTER flushing all updates out to disk.  This may take a
while.

=item SIGUSR2

The daemon exits immediately, without flushing updates out to disk.
Pending updates will be replayed from the journal when the daemon starts
up again.  B<WARNING: if journaling (-j) is NOT enabled, any pending
updates WILL BE LOST>.

=back

=head1 BUGS

No known bugs at the moment.

=head1 SEE ALSO

L<rrdtool>, L<rrdgraph>

=head1 AUTHOR

B<rrdcached> and this manual page have been written by Florian Forster
E<lt>octoE<nbsp>atE<nbsp>verplant.orgE<gt>.

=head1 CONTRIBUTORS

kevin brintnall E<lt>kbrint@rufus.netE<gt>

=cut