Get rid of cluster full mesh #3984
Labels
blocked
Not ready to be implemented
feature
A new functionality
replication
vshard
Sharding related issues
Milestone
Get rid of whole cluster full mesh to avoid too big number of sockets. When each instance sees each other even in different replicasets, it will create too many threads and sockets according to current implementation. Now each connection with a master/slave means +1 relay/applier thread, +1 socket. So for a cluster with 100 instances there are 9900 threads and sockets, 99 per each instance. A possible solution is to make the links “weaker”. For example, interconnect instances of different replicasets either via UDP (hi, SWIM), or via temporary TCP connections, which are closed after too long idle time. Relay/applier threads should be removed and possibly merged into WAL thread, or iproto. This is a preparation for sharding integration into the core.
Follows #3983.
The text was updated successfully, but these errors were encountered: