Under highly conncurrent write load, the coordinating node would
create a connection to any other node that is part of the replica
group. Since each connection can be expensive, OOM sitations could
occur because there was no bounds on the number of new connections
that would be created. If writes on a remote node were slow, connections
could pile up an exacerbate the problem.
This switches the pool to be bounded and has a checkout that is blocking
with a timeout. If a connection is available, it's returned immediately.
If the pool still has room for more connections, it will create one if needed.
Otherwise, the call will block until a connection becomes available or
the timeout expires. In the case of a timeout, it is propogated back up
to the PointsWriter that determine what do return to the client.
With this change remote mapping no longer uses HTTP, as the HTTP ports
exposed by nodes on the cluster are not known cluster wide. The TCP
ports exposed by the cluster service are, so this change uses that
functionality. Each RemoteMapper has its own dedicated connection pool
for each node, and remote mapping TCP connections are in no way coupled
with query TCP connections.
Jason Wilder
Philip O'Toole
Philip O'Toole
Ben Johnson