ignore gcc warning about write() return code not checked. It is esplicitily this way since the "max number of clients reached" is a best-effort error

replication slave timeout when receiving the initial bulk data set to 3600 seconds, now that replication is non-blocking the server must save the db before to start the async replication and this can take a lot of time with huge datasets

critical bug about glueoutputbuffers=yes fixed. Under load and with pipelining and clients disconnecting on the middle of the chat with the server, Redis could block. Now it's ok

Partial qsort implemented in SORT command, only when both BY and LIMIT is used. minor fix for a warning compiling under Linux.

Warns if /proc/sys/vm/overcommit_memory is set to 0 on Linux. Also make sure to don't resize the hash tables while the child process is saving in order to avoid copy-on-write of memory pages

FLUSHALL/FLUSHDB no longer sync on disk. Just increment the dirty counter by the number of elements removed, that will probably trigger a background saving operation

Rearrange redisObject struct to reduce memory usage in 64bit environments (as recommended http://groups.google.com/group/redis-db/msg/68f5a743f8f4e287)