// queryFailover runs an algorithm to determine which DCs to try and then calls
// them to try to locate alternative services.
func queryFailover(q queryServer, query *structs.PreparedQuery,
limit int, options structs.QueryOptions,
reply *structs.PreparedQueryExecuteResponse) error {
// Pull the list of other DCs. This is sorted by RTT in case the user
// has selected that.
nearest, err := q.GetOtherDatacentersByDistance()
if err != nil {
return err
}
// This will help us filter unknown DCs supplied by the user.
known := make(map[string]struct{})
for _, dc := range nearest {
known[dc] = struct{}{}
}
// Build a candidate list of DCs to try, starting with the nearest N
// from RTTs.
var dcs []string
index := make(map[string]struct{})
if query.Service.Failover.NearestN > 0 {
for i, dc := range nearest {
if !(i < query.Service.Failover.NearestN) {
break
}
dcs = append(dcs, dc)
index[dc] = struct{}{}
}
}
// Then add any DCs explicitly listed that weren't selected above.
for _, dc := range query.Service.Failover.Datacenters {
// This will prevent a log of other log spammage if we do not
// attempt to talk to datacenters we don't know about.
if _, ok := known[dc]; !ok {
q.GetLogger().Printf("[DEBUG] consul.prepared_query: Skipping unknown datacenter '%s' in prepared query", dc)
continue
}
// This will make sure we don't re-try something that fails
// from the NearestN list.
if _, ok := index[dc]; !ok {
dcs = append(dcs, dc)
}
}
// Now try the selected DCs in priority order.
failovers := 0
for _, dc := range dcs {
// This keeps track of how many iterations we actually run.
failovers++
// Be super paranoid and set the nodes slice to nil since it's
// the same slice we used before. We know there's nothing in
// there, but the underlying msgpack library has a policy of
// updating the slice when it's non-nil, and that feels dirty.
// Let's just set it to nil so there's no way to communicate
// through this slice across successive RPC calls.
reply.Nodes = nil
// Note that we pass along the limit since it can be applied
// remotely to save bandwidth. We also pass along the consistency
// mode information we were given, so that applies to the remote
// query as well.
remote := &structs.PreparedQueryExecuteRemoteRequest{
Datacenter: dc,
Query: *query,
Limit: limit,
QueryOptions: options,
}
if err := q.ForwardDC("PreparedQuery.ExecuteRemote", dc, remote, reply); err != nil {
q.GetLogger().Printf("[WARN] consul.prepared_query: Failed querying for service '%s' in datacenter '%s': %s", query.Service.Service, dc, err)
continue
}
// We can stop if we found some nodes.
if len(reply.Nodes) > 0 {
break
}
}
// Set this at the end because the response from the remote doesn't have
// this information.
reply.Failovers = failovers
return nil
}
