func loadLeaderLease(eng engine.Engine, raftID proto.RaftID) (*proto.Lease, error) {
lease := &proto.Lease{}
if _, err := engine.MVCCGetProto(eng, keys.RaftLeaderLeaseKey(raftID), proto.ZeroTimestamp, true, nil, lease); err != nil {
return nil, err
}
return lease, nil
}
// InternalLeaderLease sets the leader lease for this range. The command fails
// only if the desired start timestamp collides with a previous lease.
// Otherwise, the start timestamp is wound back to right after the expiration
// of the previous lease (or zero). After a lease has been set, calls to
// HasLeaderLease() will return true if this replica is the lease holder and
// the lease has not yet expired. If this range replica is already the lease
// holder, the expiration will be extended or shortened as indicated. For a new
// lease, all duties required of the range leader are commenced, including
// clearing the command queue and timestamp cache.
func (r *Range) InternalLeaderLease(batch engine.Engine, ms *engine.MVCCStats, args *proto.InternalLeaderLeaseRequest, reply *proto.InternalLeaderLeaseResponse) {
r.Lock()
defer r.Unlock()
prevLease := r.getLease()
isExtension := prevLease.RaftNodeID == args.Lease.RaftNodeID
effectiveStart := args.Lease.Start
// We return this error in "normal" lease-overlap related failures.
rErr := &proto.LeaseRejectedError{
Existing: *prevLease,
Requested: args.Lease,
}
// Verify details of new lease request. The start of this lease must
// obviously precede its expiration.
if !args.Lease.Start.Less(args.Lease.Expiration) {
reply.SetGoError(rErr)
return
}
// Wind the start timestamp back as far to the previous lease's expiration
// as we can. That'll make sure that when multiple leases are requested out
// of order at the same replica (after all, they use the request timestamp,
// which isn't straight out of our local clock), they all succeed unless
// they have a "real" issue with a previous lease. Example: Assuming no
// previous lease, one request for [5, 15) followed by one for [0, 15)
// would fail without this optimization. With it, the first request
// effectively gets the lease for [0, 15), which the second one can commit
// again (even extending your own lease is possible; see below).
//
// If no old lease exists or this is our lease, we don't need to add an
// extra tick. This allows multiple requests from the same replica to
// merge without ticking away from the minimal common start timestamp.
if prevLease.RaftNodeID == 0 || isExtension {
// TODO(tschottdorf) Think about whether it'd be better to go all the
// way back to prevLease.Start(), so that whenever the last lease is
// the own one, the original start is preserved.
effectiveStart.Backward(prevLease.Expiration)
} else {
effectiveStart.Backward(prevLease.Expiration.Next())
}
if isExtension {
if effectiveStart.Less(prevLease.Start) {
reply.SetGoError(rErr)
return
}
// Note that the lease expiration can be shortened by the holder.
// This could be used to effect a faster lease handoff.
} else if effectiveStart.Less(prevLease.Expiration) {
reply.SetGoError(rErr)
return
}
args.Lease.Start = effectiveStart
// Store the lease to disk & in-memory.
if err := engine.MVCCPutProto(batch, ms, keys.RaftLeaderLeaseKey(r.Desc().RaftID), proto.ZeroTimestamp, nil, &args.Lease); err != nil {
reply.SetGoError(err)
return
}
atomic.StorePointer(&r.lease, unsafe.Pointer(&args.Lease))
// If this replica is a new holder of the lease, update the
// low water mark in the timestamp cache. We add the maximum
// clock offset to account for any difference in clocks
// between the expiration (set by a remote node) and this
// node.
if r.getLease().RaftNodeID == r.rm.RaftNodeID() && prevLease.RaftNodeID != r.getLease().RaftNodeID {
r.tsCache.SetLowWater(prevLease.Expiration.Add(int64(r.rm.Clock().MaxOffset()), 0))
log.Infof("range %d: new leader lease %s", r.Desc().RaftID, args.Lease)
}
// Gossip configs in the event this range contains config info.
r.maybeGossipConfigsLocked(func(configPrefix proto.Key) bool {
return r.ContainsKey(configPrefix)
})
}
