view release on metacpan or  search on metacpan

lib/Cache/Memcached/Fast.pm  view on Meta::CPAN

because hash values of I<'key'> and I<'prefix/key'> may be different.

However sometimes is it necessary to hash the namespace prefix, for
instance for interoperability with other clients that do not have the
notion of the namespace.  When I<hash_namespace> is enabled, both
examples above will use the same server, the one that I<'prefix/key'>
is mapped to.  Note that there's no performance penalty then, as
namespace prefix is hashed only once.  See L</namespace>.

=item I<nowait>

  nowait => 1
  (default: disabled)

The value is a boolean which enables (true) or disables (false)
I<nowait> mode.  If enabled, when you call a method that only returns
its success status (like L</set>), B<I<in a void context>>, it sends
the request to the server and returns immediately, not waiting the
reply.  This avoids the round-trip latency at a cost of uncertain
command outcome.

Internally there is a counter of how many outstanding replies there
should be, and on any command the client reads and discards any
replies that have already arrived.  When you later execute some method
in a non-void context, all outstanding replies will be waited for, and
then the reply for this command will be read and returned.

=item I<connect_timeout>

  connect_timeout => 0.7
  (default: 0.25 seconds)

The value is a non-negative rational number of seconds to wait for
connection to establish.  Applies only to network connections.  Zero
disables timeout, but keep in mind that operating systems have their
own heuristic connect timeout.

Note that network connect process consists of several steps:
destination host address lookup, which may return several addresses in
general case (especially for IPv6, see
L<http://people.redhat.com/drepper/linux-rfc3484.html> and
L<http://people.redhat.com/drepper/userapi-ipv6.html>), then the
attempt to connect to one of those addresses.  I<connect_timeout>
applies only to one such connect, i.e. to one I<connect(2)>
call.  Thus overall connect process may take longer than
I<connect_timeout> seconds, but this is unavoidable.

=item I<io_timeout> (or deprecated I<select_timeout>)

  io_timeout => 0.5
  (default: 1.0 seconds)

The value is a non-negative rational number of seconds to wait before
giving up on communicating with the server(s).  Zero disables timeout.

Note that for commands that communicate with more than one server
(like L</get_multi>) the timeout applies per server set, not per each
server.  Thus it won't expire if one server is quick enough to
communicate, even if others are silent.  But if some servers are dead
those alive will finish communication, and then dead servers would
timeout.

=item I<close_on_error>

  close_on_error => 0
  (default: enabled)

The value is a boolean which enables (true) or disables (false)
I<close_on_error> mode.  When enabled, any error response from the
B<memcached> server would make client close the connection.  Note that
such "error response" is different from "negative response".  The
latter means the server processed the command and yield negative
result.  The former means the server failed to process the command for
some reason.  I<close_on_error> is enabled by default for safety.
Consider the following scenario:

=over

=item 1 Client want to set some value, but mistakenly sends malformed
        command (this can't happen with current module of course ;)):

  set key 10\r\n
  value_data\r\n

=item 2 Memcached server reads first line, 'set key 10', and can't
        parse it, because there's wrong number of tokens in it.  So it
        sends

  ERROR\r\n

=item 3 Then the server reads 'value_data' while it is in
        accept-command state!  It can't parse it either (hopefully),
        and sends another

  ERROR\r\n

=back

But the client expects one reply per command, so after sending the
next command it will think that the second 'ERROR' is a reply for this
new command.  This means that all replies will shift, including
replies for L</get> commands!  By closing the connection we eliminate
such possibility.

When connection dies, or the client receives the reply that it can't
understand, it closes the socket regardless the I<close_on_error>
setting.

=item I<compress_threshold>

  compress_threshold => 10_000
  (default: -1)

The value is an integer.  When positive it denotes the threshold size
in bytes: data with the size equal or larger than this should be
compressed.  See L</compress_ratio> and L</compress_methods> below.

Negative value disables compression.

=item I<compress_ratio>