Revision history for Perl extension Cache::Memcached::Fast.
0.19 2010-04-24
- revert SIGPIPE suppression change. Upgrade if you couldn't
build 0.18.
Changes since 0.18:
Turned out some orthodox systems do not have sigtimedwait().
Since advanced systems that have sigtimedwait() also have
MSG_NOSIGNAL and thus do not use SIGPIPE suppression code in
question, and other systems have neither MSG_NOSIGNAL nor
sigtimedwait(), the only option is to revert to the previous
suppression code. It works correctly unless your program is
multi-threaded.
0.18 2010-04-06
- make module thread-safe, and fix several bugs. No need to
upgrade unless you experienced limitations mentioned below.
Changes since 0.17:
Fix RT#56142: handle Perl magic. Tied scalars, arrays and
hashes are supported now, as well as Readonly variables and
other magic stuff.
Fix t/command.t failure on some Solaris distributions.
Make module thread-safe with Perl >= 5.7.2.
0.17 2009-11-02
- remove delete expiration aka delay parameter. Upgrade is
required for use with memcached >= 1.3.2.
Changes since 0.16:
Warn about ignoring non-zero delete expiration time.
Expiration time for "delete" command has been optional in
older versions of memcached, and was dropped in memcached
1.3.2 (in an incompatible manner). Not knowing server version
it's impossible to tell whether the server will handle the
parameter or not, so we simply do not send it at all.
Explicit zero passed to delete() or delete_multi() is silently
ignored, and a warning is given if you pass non-zero delete
expiration time. Furthermore the syntax delete($key, $delay)
and delete_multi([$key, $delay]) is obsoleted and no more
documented.
0.16 2009-06-15
- add missing files. This is a fix for the previous release.
Changes since 0.15:
Add src/addrinfo_hostent.c and src/addrinfo_hostent.h to
MANIFEST. Oops :(.
0.15 2009-06-15
- no need to upgrade for existing users. This release fixes
issues that new users may encounter (lack of getaddrinfo() on
some Windows systems; failing test case due to behaviour
change in memcached 1.3.3 that would break automatic
installation from CPAN).
Changes since 0.14:
Previous release (0.14) featured the change in the Ketama
algorithm. While that was a necessary move to improve key
distribution across servers, it introduced incompatibility
with earlier versions, and provided no means for smooth
transition for existing users of Ketama algorithm (i.e.,
without complete loss of cached data). The following
workaround was suggested by Masahiro Nagano---thanks!:
- apply the patch at http://limilic.com/entry/ljlt0sksbiqi16p3
- create two instances of C::M::F, one with enabled
old_ketama_algo => 1.
- for some time populate caches via both client instances,
while serving gets through old_ketama_algo. For this you
may use Cache::Migrate (http://gist.github.com/110981).
- after some time (depends on cache refresh rate) you may
drop old_ketama_algo, as new data is now distributed
according to the new Ketama implementation as well (this
doubles memory use though).
I do not include the named patch in the distribution, because
it's use is limited for one-time transition for existing
Ketama users only.
On systems lacking getaddrinfo() implement it with
gethostbyname(). In particular this should fix the build on
Win32 (Cygwin, Mingw32). Problem report (for Windows 2000 and
below) and initial patch are by Yasuhiro Matsumoto.
Fix t/commands.t: starting with memcached 1.3.3 incr/decr
commands expect numeric value in the cache. Patch by Jason
M. Mills (RT#46883).
0.14 2009-02-07
- improve Ketama distribution. The change is incompatible
with Ketama implementation in previous versions.
Changes since 0.13:
Change calculation of Ketama points for a given server to get
mapping ratios closer to server weight ratios. New Ketama
mode is compatible with nginx's module memcached_hash
(http://openhack.ru/nginx-patched/wiki/MemcachedHash) version
0.4, which has the same change, and is incompatible with
earlier versions of both named and this module.
Add script/ketama-distr.pl to compute server's share of
continuum. Real-world example:
script/ketama-distr.pl --ketama_points=150 \
--server=10.0.143.4:11211 \
--server=10.0.143.6:11211 \
--server=10.0.143.7:11211 \
--server=10.0.143.8:11211
outputs:
Old:
server 1: 10.0.143.4:11211 total = 671585356 (15.64%)
server 2: 10.0.143.6:11211 total = 601117590 (14.00%)
server 3: 10.0.143.7:11211 total = 1771239512 (41.24%)
server 4: 10.0.143.8:11211 total = 1251024837 (29.13%)
New:
server 1: 10.0.143.4:11211 total = 1057134262 (24.61%)
server 2: 10.0.143.6:11211 total = 1111432463 (25.88%)
server 3: 10.0.143.7:11211 total = 1017280856 (23.69%)
server 4: 10.0.143.8:11211 total = 1109119714 (25.82%)
Continuum array size = 4800 bytes
I.e. "Old" is what we were getting before the change, and
"New" is what we are getting now. "Continuum array size" is
how much memory is used to store the continuum array. Try
this script on your servers to see your distribution, and
possibly to tweak ketama_points and/or server weights to get a
better one.
Fix RT#41077 (http://rt.cpan.org/Ticket/Display.html?id=41077):
implement disconnect_all(). Patch is by norbi.nix.hu, with
minor corrections---thanks!.
0.13 2008-10-13
- introduce Win32 support (based on the patch by Yasuhiro
Matsumoto---arigatou!), and use poll() instead of select()
(suggested by Vladimir Timofeev).
Changes since 0.12:
Add support for Win32. I personally don't have the system nor
compiler, so I can't even test the build. Win32 port is
expected to be supported by community. Note: on Win32 below
Windows Vista max number of memcached servers is 64. See
comment on FD_SETSIZE in src/socket_win32.h to learn how to
increase this value if you need to connect to more than 64
servers.
Use poll() system call instead of select(). The latter has
the limit on the file descriptor value. I.e. even when the
number of memcached servers is low, but your application opens
lots of other files, then after some point socket() returns
fd value larger that select() can handle. poll() doesn't have
this limitation. On a side note, we don't have to use
advanced calls like epoll()/kqueue(), because number of
memcached servers is normally not very large (and single
request touches even a smaller subset).
Add bench make target ('make bench'). You may change initial
parameters at the top of script/benchmark.pl to benchmark
different types of loads.
Do not install libclient.a. Thanks to Vladimir Timofeev for
finding out how to prevent the installation.
0.12 2008-07-21
- this release fixes compilation on Solaris broken in 0.11.
If you were able to compile 0.11, then you weren't affected.
Changes since 0.11:
Fix compilation on Solaris. Ironically, 0.11 was supposed to
fix unrecognized format problem on Solaris, but instead broke
the compilation. Back in May the line
return XSRETURN_EMPTY;
has been introduced to the code, whereas it should be
XSRETURN_EMPTY;
It went unnoticed because expansion produces the code
permitted by compilers like GCC that allow blocks to return a
value. But Perl on Solaris has a different definition of this
macro that is not an expression. Hence the problem.
0.11 2008-07-17
- this release fixes %zu problem on Solaris.
Changes since 0.10:
Replace size_t/%zu with unsigned long/%lu for benefit of
systems that do not recognize the former format (Solaris 9).
Thanks to Michael Vychizhanin!
Add new parameter max_size (suggested by Alex Kapranoff, see
RT#35588, http://rt.cpan.org/Ticket/Display.html?id=35588).
Fix RT#36074 (http://rt.cpan.org/Ticket/Display.html?id=36074):
ExtUtils::MakeMaker version 6.17 and below generated invalid
Makefile.
0.10 2008-05-01
- this is a bugfix release fixing key distribution in
compatible mode (reported by Anton Yuzhaninov---thanks!), and
minor fixes in Ketama mode for collision case (very rare).
Changes since 0.09:
Fix key distribution bug in compatible mode. Because of
accumulated rounding error some keys were mapped to the
different server than with Cache::Memcached.
Remove explicit OPTIMIZE setting from the makefiles. It ought
to be set explicitly by the user. You can run 'perl -V:optimize'
to learn the default. It is safe to increase optimization
level to -O3.
Fix hash_namespace parameter: it didn't work for the namespace
specified in the constructor.
Minor fixes in Ketama mode for rare collision case.
0.09 2008-03-06
- this release makes 'noreply' mode official, and improves
latency for some command patterns by managing TCP_NODELAY flag.
Changes since 0.08:
Enable TCP_NODELAY to reduce latency of commands that wait for
the reply, and disable it for throughput of commands that don't
need any reply.
t/encoding.t was removed from distribution. It doesn't work
with Perl 5.6, which has no 'encoding' pragma, and I don't
know how to fix it (beautifully, there's a number of ugly
ways).
'noreply' mode is documented as the official feature of
memcached 1.2.5. script/benchmark.pl uses it automatically
when available.
Add parameter 'hash_namespace' (see docs).
Add method 'namespace' to get and/or set the namespace prefix.
This is handy when you "lease" C::M::F object for different
parts of your application that do unrelated things.
0.08 2008-01-24
- this is mostly a cleanup release, that also improves
decompression speed.
Changes since 0.07:
compress_alog was replaced with more generic compress_methods.
It turned out that IO::Uncompress::Gunzip is much (~4x) slower
than Compress::Zlib, so the latter is used as the default for
compress_methods. Besides, previous mechanism didn't scale
well.
Implemented the check of constructor arguments to catch
various typos (enabled by default), and new 'check_args'
parameter to disable it.
Workaround "cast from integer to pointer of different size"
warning on 64-bit platforms.
Removed 127.0.0.x from test cases since not all hosts resolve
such addresses when x is not 1.
Use plain malloc()/free() instead of Newx()/Safefree() in XS.
This should fix the build with Perl versions 5.6.0--5.8.6.
0.07 2008-01-18
- this is a major release that features a significant speed
improvement and introduces multi update commands. It also
fixes a bug in nowait mode that was introduced in 0.06.
Upgrade is recommended if you are using nowait => 1, or want
to employ new features and speed.
Changes since 0.06:
For performance reasons all Perl code was converted to XS. If
you performed any benchmarks, you may want to repeat them ;).
This change has a drawback that now you can't derive from
C::M::F easily.
Every update command now has its multi equivalent: set_multi,
add_multi, replace_multi, append_multi, prepend_multi,
incr_multi, decr_multi and delete_multi. By using multi
commands you will reduce request latency: several requests
would be sent in one packet, and the client would talk to all
servers simultaneously.
Improved results of update commands: commands that previously
returned boolean value now return true for positive server
reply, false for negative server reply, and undef on error.
Similar to DBI, decr returns "0E0" for positive zero reply,
which evaluates to true in a boolean context.
flush_all now returns a hash 'server_name' => 'bool_result'.
New commands 'server_versions' and 'nowait_push'.
New 'utf8' and 'serialize_methods' parameters for constructor.
script/compare.pl was removed. Instead script/benchmark.pl
takes an optional last argument "compare".
t/03-flush_all.t was removed. We use PID as a part of a
namespace prefix now, so we don't need clear cache any more.
And there are good reasons not to have flush_all: the user may
flush a valuable cache by mistake, and "flush_all 0" doesn't
work on memcached 1.2.2, so the outcome was uncertain anyway.
Without flush_all it's possible to execute in parallel more
than one 'make test'.
Fixed bug in 'nowait' mode that could cause false negative
replies to subsequent requests that you are waiting for.
Workaround the loss of the last 'noreply' request on the
server. 'noreply' is still an undocumented feature.
To support multi commands the API of internal client
implementation was changed. If you were using it as a
standalone library---sorry :). After API will freeze C client
library will likely be released on its own.
0.06 2007-12-20
- this release fixes a serious bug and features the 'nowait'
mode. Upgrade is recommended.
Changes since 0.05:
Fix a subtle race bug present since 0.02 that could mix
several values into one when you have more than one memcached
server and executed get_multi or gets_multi. If you have seen
the following in your logs or on the console:
Attempt to free unreferenced scalar: SV 0x9e89d34 at
/tmp/Cache-Memcached-Fast/blib/lib/Cache/Memcached/Fast.pm line 688.
that was it.
Added 'nowait' mode that allows the client to not wait the
reply, and thus reduces the wallclock time. Run
script/benchmark.pl and see how it affects wallclock time,
especially over the wire.
Made undocumented 'noreply' parameter a per-server setting.
0.05 2007-12-18
- more fixes to build and pass tests on different platforms.
There's no functional changes again.
Changes since 0.04:
At least on Perl 5.6.2 SVREF typemap entry has the meaning of
the class based on blessed SV. So instead of SVREF we use
custom Ref_SV and custom typemap entry. Found by CPAN
testers.
Do not use AUTOLOAD for normal commands, only for undocumented
ones.
Add incr() and decr() to script/compare.pl and
script/benchmark.pl.
Test case is split into several files and cleaned a bit.
Rename 'sun' -> 's_unix' in connect.c, because 'sun' is
reserved on... guesses? Right, Sun systems! Found by CPAN
testers.
Fix test case for FreeBSD and derivatives: do not use
localhost.localdomain address, which FreeBSD doesn't resolve
by default.
Added docs section "UTF-8 and tainted data".
0.04 2007-12-16
- CPAN testers found more issues on different platforms,
fixing these.
Changes since 0.03:
const char * is missing from typemap in Perl 5.6.2, fixed by
using plain char * for class name in new().
Use AI_ADDRCONFIG conditionally, some systems do not have it
(NetBSD 3.1).
0.03 2007-12-16
- this release has no new functionality compared to 0.02. It
only fixes some test and documentation problems that were
discovered, as it always happens, after the upload to CPAN.
Changes since 0.02:
Fixed internal documentation references.
Added missing docs for incr and decr.
Fixed test case to test server version and use
cas/gets/append/prepend only for 1.2.4 and up.
Actually there's new _undocumented_ command server_versions,
is has broken interface, you can't say which version
corresponds to which server when some of them do not respond.
This command is subject to change.
0.02 2007-12-15
- first public release.
Changes since 0.01:
Fully restartable code was replaced with classic
fill-the-buffer-first approach: it turned out that
restartability doesn't add much advantage. Lots of beautiful
and ugly code gone away.
Added support for multiple servers (select machine, CRC32,
timeouts).
Added script/compare.pl to compare this module to the original
Cache::Memcached, and also to measure speed in general.
Added Unix socket support, SIGPIPE handling (proper ignoring
actually ;)).
Added support for serialization of Perl data structures, and
for compression.
New commands: incr, decr, gets, gets_multi, cas,
enable_compression, remove (alias to delete).
New client parameters: connect_timeout, io_timeout (aka
select_timeout), compress_threshold, compress_ratio,
compress_algo, max_failure, failure_timeout.
Finally added documentation!
Added support for server weights.
Added the Ketama consistent hashing algorithm.
Added default tests as generated with Module::Starter. Main
test case is still messy, should be split into several tests.
There's also support for 'noreply'-enabled memcached, but
those patches are not (yet) accepted to mainline, so 'noreply'
is not an official feature and is not documented.
0.01 2007-11-26
- first internal release.
Supported only one server connection (TCP). Commands are
set(), add(), replace(), prepend(), append(), delete(),
flush_all(), get(), get_multi(). Client parameters are
'servers', 'namespace', 'close_on_error'. There's no
documentation. Test case is messy (but does its job).
0.00 2007-11-19
- original version; created by h2xs 1.23 with options
-O --omit-autoload --compat-version=5.6.0 \
--use-new-tests --name=Cache::Memcached::Fast