maxclients configuration is now implemented dealing with the actual process rlimits. Setting maxclients to 0 no longer makes sense and is now invalid, the new default is 10000.
See issue #162 for more information.

Remove the write handler only if there are no longer objects in the output queue AND if the static buffer is empty. This bug was the cause of a possible server-stop-responding-to-client bug under some specific work load. Thanks to Pieter Noordhuis for spotting and fixing it.

Take a pointer to the relevant entry of the command table in the client structure. This is generally a more sounding design, simplifies a few functions prototype, and as a side effect fixes a bug related to the conversion of EXPIRE -1 to DEL: before of this fix Redis tried to convert it into an EXPIREAT in the AOF code, regardless of our rewrite of the command.

Make sure error and status replies emitted by Lua scripts can never have more than a newline, otherwise it is a protocol violation and clients will desync.

when creating not connected clients do not add them into the clients list, otherwise they will be subject to timeouts and other stuff

Correctly glue the reply buffer. For now returned as it is to Lua, but will be converted into Lua native type later.

Lua call of Redis command work in progress: sorry I have to go to the cinema to watch the Source Code movie

when Redis fails accepting a new connection reports the error at WARNING and not VERBOSE error level. Thanks to offby1 for proposing this in the Redis mailing list. #backport-candidate

Fixed a bug with replication where SLAVEOF NO ONE caused a slave to close the connection with its slaves

removed check for zmalloc return NULL in createClient(). The check was misplaced, and zmalloc never returns NULL.

fixed memory leak introduced with the previous commit. Many thanks to Pieter Noordhuis for spotting it in no time

Fixed issue #503. MONITOR + QUIT could crash the server, there are actually other interactions that could have the same effect (for instance Pub/Sub).

Fixed issue #435 and at the same time introduced explicit ping in the master-slave channel that will detect a blocked master or a broken even if apparently connected TCP link.

implemented a different approach to IO scheduling, so object->storage is no longer used, instead there is a queue and hash table of IO tasks to process, and it is always possible to know what are the scheduled and acrtive IO operations against every single key.

LRANGE converted into a COW friendly command. Some refactoring, comment, and new addReply*() family function added in the process.

first attempt to non blocking implementation of slave replication and SYNC bulk data download. Never compiled so far...