Non-raft nodes need to be notifified when the metastore changes. For
example, a database could be dropped on node 1 (non-raft) and node 2
would not know. Since queries for that database would not be a cache
miss, node 2 would not get updated.
To propogate changes to non-raft nodes, each non-raft node maintains
a blocking connection to a raft node that blocks until a metadata
change occurs. When the change is triggered, the updated metadata
is returned to the client and the client idempotently updates its local
cache. It then reconnects and waits for another change. This is
similar watches in zookeeper or etcd. Since the blocking request is
always recreated, it also serves as a polling mechanism that will retry
another raft member if the current connection is lost.
Some errors would not be returned to the client because something
failed before we could create the appropriate respone. For these
cases, a general error response is returned.
Nodes that are not part of the raft cluster will not reliably know who the
current raft cluster leader is. To make communication simpler, proxy all
rpc and raft calls to the current raft leader if a non-leader receives one.
This adds some basic ability to join a node to an existing cluster. It
uses a rpc layer to initiate a join request to an existing memeber. The
response indicates whether the joining node should take part in the raft
cluster and who it's peers should be. If raft should not be started, the
peers are the addresses of the current raft members that it should delegate
consensus operations.
To keep the meta store implementation agnostic of whether it's running
a local raft or not, a consensusStrategy type was also added.
This adds some basic plumbing to make remote procedure calls to other cluster
members. This first implementation allows a node to contact the raft leader
and fetch a copy of the meta data. This will be used by non-raft members to
pull down the latest metadata.
This change moves tracking of next timestamp and values to simple
slices, as performance measurement showed that Peek() on TagSet cursors
was a huge performance drain. There is much more that can be done here,
but with this in place query performance has been restored to 0.9.1
levels.
This change also uses -1 to indicate that no value is available for a
given timestamp.
Jason Wilder
dgnorton
Josh Horwitz
Philip O'Toole
Ben Johnson
Philip O'Toole
Nathaniel Cook
Nathaniel Cook
Gunnar
gunnaraasen