open-consul/agent/consul/state/usage.go
Dan Upton 9fe6c33c0d
xDS Load Balancing (#14397)
Prior to #13244, connect proxies and gateways could only be configured by an
xDS session served by the local client agent.

In an upcoming release, it will be possible to deploy a Consul service mesh
without client agents. In this model, xDS sessions will be handled by the
servers themselves, which necessitates load-balancing to prevent a single
server from receiving a disproportionate amount of load and becoming
overwhelmed.

This introduces a simple form of load-balancing where Consul will attempt to
achieve an even spread of load (xDS sessions) between all healthy servers.
It does so by implementing a concurrent session limiter (limiter.SessionLimiter)
and adjusting the limit according to autopilot state and proxy service
registrations in the catalog.

If a server is already over capacity (i.e. the session limit is lowered),
Consul will begin draining sessions to rebalance the load. This will result
in the client receiving a `RESOURCE_EXHAUSTED` status code. It is the client's
responsibility to observe this response and reconnect to a different server.

Users of the gRPC client connection brokered by the
consul-server-connection-manager library will get this for free.

The rate at which Consul will drain sessions to rebalance load is scaled
dynamically based on the number of proxies in the catalog.
2022-09-09 15:02:01 +01:00

472 lines
14 KiB
Go

package state
import (
"fmt"
"github.com/hashicorp/go-memdb"
"github.com/hashicorp/consul/agent/structs"
)
const (
serviceNamesUsageTable = "service-names"
kvUsageTable = "kv-entries"
connectNativeInstancesTable = "connect-native"
connectPrefix = "connect-mesh"
tableUsage = "usage"
)
var allConnectKind = []string{
string(structs.ServiceKindConnectProxy),
string(structs.ServiceKindIngressGateway),
string(structs.ServiceKindMeshGateway),
string(structs.ServiceKindTerminatingGateway),
connectNativeInstancesTable,
}
// usageTableSchema returns a new table schema used for tracking various indexes
// for the Raft log.
func usageTableSchema() *memdb.TableSchema {
return &memdb.TableSchema{
Name: tableUsage,
Indexes: map[string]*memdb.IndexSchema{
indexID: {
Name: indexID,
AllowMissing: false,
Unique: true,
Indexer: &memdb.StringFieldIndex{
Field: "ID",
Lowercase: true,
},
},
},
}
}
// UsageEntry represents a count of some arbitrary identifier within the
// state store, along with the last seen index.
type UsageEntry struct {
ID string
Index uint64
Count int
}
// ServiceUsage contains all of the usage data related to services
type ServiceUsage struct {
Services int
ServiceInstances int
ConnectServiceInstances map[string]int
EnterpriseServiceUsage
}
// NodeUsage contains all of the usage data related to nodes
type NodeUsage struct {
Nodes int
EnterpriseNodeUsage
}
// PeeringUsage contains all of the usage data related to peerings.
type PeeringUsage struct {
// Number of peerings.
Peerings int
EnterprisePeeringUsage
}
type KVUsage struct {
KVCount int
EnterpriseKVUsage
}
type ConfigEntryUsage struct {
ConfigByKind map[string]int
EnterpriseConfigEntryUsage
}
type uniqueServiceState int
const (
NoChange uniqueServiceState = 0
Deleted uniqueServiceState = 1
Created uniqueServiceState = 2
)
// updateUsage takes a set of memdb changes and computes a delta for specific
// usage metrics that we track.
func updateUsage(tx WriteTxn, changes Changes) error {
usageDeltas := make(map[string]int)
serviceNameChanges := make(map[structs.ServiceName]int)
for _, change := range changes.Changes {
var delta int
if change.Created() {
delta = 1
} else if change.Deleted() {
delta = -1
}
switch change.Table {
case tableNodes:
node := changeObject(change).(*structs.Node)
if node.PeerName != "" {
// TODO(peering) track peered nodes separately. For now not tracking to avoid double billing.
continue
}
usageDeltas[change.Table] += delta
addEnterpriseNodeUsage(usageDeltas, change)
case tablePeering:
usageDeltas[change.Table] += delta
addEnterprisePeeringUsage(usageDeltas, change)
case tableServices:
svc := changeObject(change).(*structs.ServiceNode)
if svc.PeerName != "" {
// TODO(peering) track peered services separately. For now not tracking to avoid double billing.
continue
}
usageDeltas[change.Table] += delta
addEnterpriseServiceInstanceUsage(usageDeltas, change)
connectDeltas(change, usageDeltas, delta)
// Construct a mapping of all of the various service names that were
// changed, in order to compare it with the finished memdb state.
// Make sure to account for the fact that services can change their names.
if serviceNameChanged(change) {
serviceNameChanges[svc.CompoundServiceName()] += 1
before := change.Before.(*structs.ServiceNode)
serviceNameChanges[before.CompoundServiceName()] -= 1
} else {
serviceNameChanges[svc.CompoundServiceName()] += delta
}
case "kvs":
usageDeltas[change.Table] += delta
addEnterpriseKVUsage(usageDeltas, change)
case tableConfigEntries:
entry := changeObject(change).(structs.ConfigEntry)
usageDeltas[configEntryUsageTableName(entry.GetKind())] += delta
addEnterpriseConfigEntryUsage(usageDeltas, change)
}
}
serviceStates, err := updateServiceNameUsage(tx, usageDeltas, serviceNameChanges)
if err != nil {
return err
}
addEnterpriseServiceUsage(usageDeltas, serviceStates)
idx := changes.Index
// This will happen when restoring from a snapshot, just take the max index
// of the tables we are tracking.
if idx == 0 {
// TODO(partitions? namespaces?)
idx = maxIndexTxn(tx, tableNodes, tableServices, "kvs")
}
return writeUsageDeltas(tx, idx, usageDeltas)
}
func updateServiceNameUsage(tx WriteTxn, usageDeltas map[string]int, serviceNameChanges map[structs.ServiceName]int) (map[structs.ServiceName]uniqueServiceState, error) {
serviceStates := make(map[structs.ServiceName]uniqueServiceState, len(serviceNameChanges))
for svc, delta := range serviceNameChanges {
q := Query{
Value: svc.Name,
EnterpriseMeta: svc.EnterpriseMeta,
}
serviceIter, err := tx.Get(tableServices, indexService, q)
if err != nil {
return nil, err
}
// Count the number of service instances associated with the given service
// name at the end of this transaction, and compare that with how many were
// added/removed during the transaction. This allows us to handle a single
// transaction committing multiple changes related to a single service
// name.
var count int
for service := serviceIter.Next(); service != nil; service = serviceIter.Next() {
count += 1
}
var serviceState uniqueServiceState
switch {
case count == 0:
// If no services exist, we know we deleted the last service
// instance.
serviceState = Deleted
usageDeltas[serviceNamesUsageTable] -= 1
case count == delta:
// If the current number of service instances equals the number added,
// than we know we created a new service name.
serviceState = Created
usageDeltas[serviceNamesUsageTable] += 1
default:
serviceState = NoChange
}
serviceStates[svc] = serviceState
}
return serviceStates, nil
}
// serviceNameChanged returns a boolean that indicates whether the
// provided change resulted in an update to the service's service name.
func serviceNameChanged(change memdb.Change) bool {
if change.Updated() {
before := change.Before.(*structs.ServiceNode)
after := change.After.(*structs.ServiceNode)
return before.ServiceName != after.ServiceName
}
return false
}
// connectUsageTableEntry is a convenience function to make prefix addition in 1 place
func connectUsageTableName(kind string) string {
return fmt.Sprintf("%s-%s", connectPrefix, kind)
}
// configEntryUsageTableName is a convenience function to easily get the prefix + config entry kind in 1 place
func configEntryUsageTableName(kind string) string {
return fmt.Sprintf("%s-%s", tableConfigEntries, kind)
}
func connectDeltas(change memdb.Change, usageDeltas map[string]int, delta int) {
// Connect metrics for updated services are more complicated. Check for:
// 1. Did ServiceKind change?
// 2. Is before ServiceKind typical? don't remove from old service kind
// 3. Is After ServiceKind typical? don't add to new service kind
// 4. Add and remove to both ServiceKind's
if change.Updated() {
before := change.Before.(*structs.ServiceNode)
after := change.After.(*structs.ServiceNode)
if before.ServiceKind != structs.ServiceKindTypical {
usageDeltas[connectUsageTableName(string(before.ServiceKind))] -= 1
addEnterpriseConnectServiceInstanceUsage(usageDeltas, before, -1)
}
if after.ServiceKind != structs.ServiceKindTypical {
usageDeltas[connectUsageTableName(string(after.ServiceKind))] += 1
addEnterpriseConnectServiceInstanceUsage(usageDeltas, after, 1)
}
if before.ServiceConnect.Native != after.ServiceConnect.Native {
if before.ServiceConnect.Native {
usageDeltas[connectUsageTableName(string(connectNativeInstancesTable))] -= 1
addEnterpriseConnectServiceInstanceUsage(usageDeltas, before, -1)
} else {
usageDeltas[connectUsageTableName(connectNativeInstancesTable)] += 1
addEnterpriseConnectServiceInstanceUsage(usageDeltas, after, 1)
}
}
} else {
svc := changeObject(change).(*structs.ServiceNode)
if svc.ServiceKind != structs.ServiceKindTypical {
usageDeltas[connectUsageTableName(string(svc.ServiceKind))] += delta
}
if svc.ServiceConnect.Native {
usageDeltas[connectUsageTableName(connectNativeInstancesTable)] += delta
}
addEnterpriseConnectServiceInstanceUsage(usageDeltas, svc, delta)
}
}
// writeUsageDeltas will take in a map of IDs to deltas and update each
// entry accordingly, checking for integer underflow. The index that is
// passed in will be recorded on the entry as well.
func writeUsageDeltas(tx WriteTxn, idx uint64, usageDeltas map[string]int) error {
for id, delta := range usageDeltas {
u, err := tx.First(tableUsage, indexID, id)
if err != nil {
return fmt.Errorf("failed to retrieve existing usage entry: %s", err)
}
if u == nil {
if delta < 0 {
// Don't return an error here, since we don't want to block updates
// from happening to the state store. But, set the delta to 0 so that
// we do not accidentally underflow the uint64 and begin reporting
// large numbers.
delta = 0
}
err := tx.Insert(tableUsage, &UsageEntry{
ID: id,
Count: delta,
Index: idx,
})
if err != nil {
return fmt.Errorf("failed to update usage entry: %s", err)
}
} else if cur, ok := u.(*UsageEntry); ok {
updated := cur.Count + delta
if updated < 0 {
// Don't return an error here, since we don't want to block updates
// from happening to the state store. But, set the delta to 0 so that
// we do not accidentally underflow the uint64 and begin reporting
// large numbers.
updated = 0
}
err := tx.Insert(tableUsage, &UsageEntry{
ID: id,
Count: updated,
Index: idx,
})
if err != nil {
return fmt.Errorf("failed to update usage entry: %s", err)
}
}
}
return nil
}
// NodeUsage returns the latest seen Raft index, a compiled set of node usage
// data, and any errors.
func (s *Store) NodeUsage() (uint64, NodeUsage, error) {
tx := s.db.ReadTxn()
defer tx.Abort()
nodes, err := firstUsageEntry(nil, tx, tableNodes)
if err != nil {
return 0, NodeUsage{}, fmt.Errorf("failed nodes lookup: %s", err)
}
usage := NodeUsage{
Nodes: nodes.Count,
}
results, err := compileEnterpriseNodeUsage(tx, usage)
if err != nil {
return 0, NodeUsage{}, fmt.Errorf("failed nodes lookup: %s", err)
}
return nodes.Index, results, nil
}
// PeeringUsage returns the latest seen Raft index, a compiled set of peering usage
// data, and any errors.
func (s *Store) PeeringUsage() (uint64, PeeringUsage, error) {
tx := s.db.ReadTxn()
defer tx.Abort()
peerings, err := firstUsageEntry(nil, tx, tablePeering)
if err != nil {
return 0, PeeringUsage{}, fmt.Errorf("failed peerings lookup: %s", err)
}
usage := PeeringUsage{
Peerings: peerings.Count,
}
results, err := compileEnterprisePeeringUsage(tx, usage)
if err != nil {
return 0, PeeringUsage{}, fmt.Errorf("failed peerings lookup: %s", err)
}
return peerings.Index, results, nil
}
// ServiceUsage returns the latest seen Raft index, a compiled set of service
// usage data, and any errors.
func (s *Store) ServiceUsage(ws memdb.WatchSet) (uint64, ServiceUsage, error) {
tx := s.db.ReadTxn()
defer tx.Abort()
serviceInstances, err := firstUsageEntry(ws, tx, tableServices)
if err != nil {
return 0, ServiceUsage{}, fmt.Errorf("failed services lookup: %s", err)
}
services, err := firstUsageEntry(ws, tx, serviceNamesUsageTable)
if err != nil {
return 0, ServiceUsage{}, fmt.Errorf("failed services lookup: %s", err)
}
serviceKindInstances := make(map[string]int)
for _, kind := range allConnectKind {
usage, err := firstUsageEntry(ws, tx, connectUsageTableName(kind))
if err != nil {
return 0, ServiceUsage{}, fmt.Errorf("failed services lookup: %s", err)
}
serviceKindInstances[kind] = usage.Count
}
usage := ServiceUsage{
ServiceInstances: serviceInstances.Count,
Services: services.Count,
ConnectServiceInstances: serviceKindInstances,
}
results, err := compileEnterpriseServiceUsage(ws, tx, usage)
if err != nil {
return 0, ServiceUsage{}, fmt.Errorf("failed services lookup: %s", err)
}
return serviceInstances.Index, results, nil
}
func (s *Store) KVUsage() (uint64, KVUsage, error) {
tx := s.db.ReadTxn()
defer tx.Abort()
kvs, err := firstUsageEntry(nil, tx, "kvs")
if err != nil {
return 0, KVUsage{}, fmt.Errorf("failed kvs lookup: %s", err)
}
usage := KVUsage{
KVCount: kvs.Count,
}
results, err := compileEnterpriseKVUsage(tx, usage)
if err != nil {
return 0, KVUsage{}, fmt.Errorf("failed kvs lookup: %s", err)
}
return kvs.Index, results, nil
}
func (s *Store) ConfigEntryUsage() (uint64, ConfigEntryUsage, error) {
tx := s.db.ReadTxn()
defer tx.Abort()
configEntries := make(map[string]int)
var maxIdx uint64
for _, kind := range structs.AllConfigEntryKinds {
configEntry, err := firstUsageEntry(nil, tx, configEntryUsageTableName(kind))
if configEntry.Index > maxIdx {
maxIdx = configEntry.Index
}
if err != nil {
return 0, ConfigEntryUsage{}, fmt.Errorf("failed config entry usage lookup: %s", err)
}
configEntries[kind] = configEntry.Count
}
usage := ConfigEntryUsage{
ConfigByKind: configEntries,
}
results, err := compileEnterpriseConfigEntryUsage(tx, usage)
if err != nil {
return 0, ConfigEntryUsage{}, fmt.Errorf("failed config entry usage lookup: %s", err)
}
return maxIdx, results, nil
}
func firstUsageEntry(ws memdb.WatchSet, tx ReadTxn, id string) (*UsageEntry, error) {
watch, usage, err := tx.FirstWatch(tableUsage, indexID, id)
if err != nil {
return nil, err
}
ws.Add(watch)
// If no elements have been inserted, the usage entry will not exist. We
// return a valid value so that can be certain the return value is not nil
// when no error has occurred.
if usage == nil {
return &UsageEntry{ID: id, Count: 0}, nil
}
realUsage, ok := usage.(*UsageEntry)
if !ok {
return nil, fmt.Errorf("failed usage lookup: type %T is not *UsageEntry", usage)
}
return realUsage, nil
}