consul/agent/proxycfg/state.go

854 lines
28 KiB
Go
Raw Normal View History

package proxycfg
import (
"context"
"errors"
"fmt"
"log"
"reflect"
"strings"
"time"
"github.com/hashicorp/consul/agent/cache"
cachetype "github.com/hashicorp/consul/agent/cache-types"
"github.com/hashicorp/consul/agent/structs"
"github.com/mitchellh/copystructure"
connect: reconcile how upstream configuration works with discovery chains (#6225) * connect: reconcile how upstream configuration works with discovery chains The following upstream config fields for connect sidecars sanely integrate into discovery chain resolution: - Destination Namespace/Datacenter: Compilation occurs locally but using different default values for namespaces and datacenters. The xDS clusters that are created are named as they normally would be. - Mesh Gateway Mode (single upstream): If set this value overrides any value computed for any resolver for the entire discovery chain. The xDS clusters that are created may be named differently (see below). - Mesh Gateway Mode (whole sidecar): If set this value overrides any value computed for any resolver for the entire discovery chain. If this is specifically overridden for a single upstream this value is ignored in that case. The xDS clusters that are created may be named differently (see below). - Protocol (in opaque config): If set this value overrides the value computed when evaluating the entire discovery chain. If the normal chain would be TCP or if this override is set to TCP then the result is that we explicitly disable L7 Routing and Splitting. The xDS clusters that are created may be named differently (see below). - Connect Timeout (in opaque config): If set this value overrides the value for any resolver in the entire discovery chain. The xDS clusters that are created may be named differently (see below). If any of the above overrides affect the actual result of compiling the discovery chain (i.e. "tcp" becomes "grpc" instead of being a no-op override to "tcp") then the relevant parameters are hashed and provided to the xDS layer as a prefix for use in naming the Clusters. This is to ensure that if one Upstream discovery chain has no overrides and tangentially needs a cluster named "api.default.XXX", and another Upstream does have overrides for "api.default.XXX" that they won't cross-pollinate against the operator's wishes. Fixes #6159
2019-08-02 03:03:34 +00:00
"github.com/mitchellh/mapstructure"
)
type CacheNotifier interface {
Notify(ctx context.Context, t string, r cache.Request,
correlationID string, ch chan<- cache.UpdateEvent) error
}
const (
coalesceTimeout = 200 * time.Millisecond
rootsWatchID = "roots"
leafWatchID = "leaf"
intentionsWatchID = "intentions"
serviceListWatchID = "service-list"
datacentersWatchID = "datacenters"
serviceResolversWatchID = "service-resolvers"
svcChecksWatchIDPrefix = cachetype.ServiceHTTPChecksName + ":"
serviceIDPrefix = string(structs.UpstreamDestTypeService) + ":"
preparedQueryIDPrefix = string(structs.UpstreamDestTypePreparedQuery) + ":"
defaultPreparedQueryPollInterval = 30 * time.Second
)
// state holds all the state needed to maintain the config for a registered
// connect-proxy service. When a proxy registration is changed, the entire state
// is discarded and a new one created.
type state struct {
// logger, source and cache are required to be set before calling Watch.
logger *log.Logger
source *structs.QuerySource
cache CacheNotifier
// ctx and cancel store the context created during initWatches call
ctx context.Context
cancel func()
kind structs.ServiceKind
service string
proxyID string
address string
port int
taggedAddresses map[string]structs.ServiceAddress
proxyCfg structs.ConnectProxyConfig
token string
ch chan cache.UpdateEvent
snapCh chan ConfigSnapshot
reqCh chan chan *ConfigSnapshot
}
// newState populates the state struct by copying relevant fields from the
// NodeService and Token. We copy so that we can use them in a separate
// goroutine later without reasoning about races with the NodeService passed
// (especially for embedded fields like maps and slices).
//
// The returned state needs its required dependencies to be set before Watch
// can be called.
func newState(ns *structs.NodeService, token string) (*state, error) {
if ns.Kind != structs.ServiceKindConnectProxy && ns.Kind != structs.ServiceKindMeshGateway {
return nil, errors.New("not a connect-proxy or mesh-gateway")
}
// Copy the config map
proxyCfgRaw, err := copystructure.Copy(ns.Proxy)
if err != nil {
return nil, err
}
proxyCfg, ok := proxyCfgRaw.(structs.ConnectProxyConfig)
if !ok {
return nil, errors.New("failed to copy proxy config")
}
taggedAddresses := make(map[string]structs.ServiceAddress)
for k, v := range ns.TaggedAddresses {
taggedAddresses[k] = v
}
return &state{
kind: ns.Kind,
service: ns.Service,
proxyID: ns.ID,
address: ns.Address,
port: ns.Port,
taggedAddresses: taggedAddresses,
proxyCfg: proxyCfg,
token: token,
// 10 is fairly arbitrary here but allow for the 3 mandatory and a
// reasonable number of upstream watches to all deliver their initial
// messages in parallel without blocking the cache.Notify loops. It's not a
// huge deal if we do for a short period so we don't need to be more
// conservative to handle larger numbers of upstreams correctly but gives
// some head room for normal operation to be non-blocking in most typical
// cases.
ch: make(chan cache.UpdateEvent, 10),
snapCh: make(chan ConfigSnapshot, 1),
reqCh: make(chan chan *ConfigSnapshot, 1),
}, nil
}
// Watch initialized watches on all necessary cache data for the current proxy
// registration state and returns a chan to observe updates to the
// ConfigSnapshot that contains all necessary config state. The chan is closed
// when the state is Closed.
func (s *state) Watch() (<-chan ConfigSnapshot, error) {
s.ctx, s.cancel = context.WithCancel(context.Background())
err := s.initWatches()
if err != nil {
s.cancel()
return nil, err
}
go s.run()
return s.snapCh, nil
}
// Close discards the state and stops any long-running watches.
func (s *state) Close() error {
if s.cancel != nil {
s.cancel()
}
return nil
}
// initWatches sets up the watches needed for the particular service
func (s *state) initWatches() error {
switch s.kind {
case structs.ServiceKindConnectProxy:
return s.initWatchesConnectProxy()
case structs.ServiceKindMeshGateway:
return s.initWatchesMeshGateway()
default:
return fmt.Errorf("Unsupported service kind")
}
}
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
func (s *state) watchMeshGateway(ctx context.Context, dc string, upstreamID string) error {
return s.cache.Notify(ctx, cachetype.InternalServiceDumpName, &structs.ServiceDumpRequest{
Datacenter: dc,
QueryOptions: structs.QueryOptions{Token: s.token},
ServiceKind: structs.ServiceKindMeshGateway,
UseServiceKind: true,
Source: *s.source,
}, "mesh-gateway:"+dc+":"+upstreamID, s.ch)
}
func (s *state) watchConnectProxyService(ctx context.Context, correlationId string, service string, dc string, filter string) error {
return s.cache.Notify(ctx, cachetype.HealthServicesName, &structs.ServiceSpecificRequest{
Datacenter: dc,
QueryOptions: structs.QueryOptions{
Token: s.token,
Filter: filter,
},
ServiceName: service,
Connect: true,
// Note that Identifier doesn't type-prefix for service any more as it's
// the default and makes metrics and other things much cleaner. It's
// simpler for us if we have the type to make things unambiguous.
Source: *s.source,
}, correlationId, s.ch)
}
// initWatchesConnectProxy sets up the watches needed based on current proxy registration
// state.
func (s *state) initWatchesConnectProxy() error {
// Watch for root changes
err := s.cache.Notify(s.ctx, cachetype.ConnectCARootName, &structs.DCSpecificRequest{
Datacenter: s.source.Datacenter,
QueryOptions: structs.QueryOptions{Token: s.token},
Source: *s.source,
}, rootsWatchID, s.ch)
if err != nil {
return err
}
// Watch the leaf cert
err = s.cache.Notify(s.ctx, cachetype.ConnectCALeafName, &cachetype.ConnectCALeafRequest{
Datacenter: s.source.Datacenter,
Token: s.token,
Service: s.proxyCfg.DestinationServiceName,
}, leafWatchID, s.ch)
if err != nil {
return err
}
// Watch for intention updates
err = s.cache.Notify(s.ctx, cachetype.IntentionMatchName, &structs.IntentionQueryRequest{
Datacenter: s.source.Datacenter,
QueryOptions: structs.QueryOptions{Token: s.token},
Match: &structs.IntentionQueryMatch{
Type: structs.IntentionMatchDestination,
Entries: []structs.IntentionMatchEntry{
{
Namespace: structs.IntentionDefaultNamespace,
Name: s.proxyCfg.DestinationServiceName,
},
},
},
}, intentionsWatchID, s.ch)
if err != nil {
return err
}
// Watch for service check updates
err = s.cache.Notify(s.ctx, cachetype.ServiceHTTPChecksName, &cachetype.ServiceHTTPChecksRequest{
ServiceID: s.proxyCfg.DestinationServiceID,
}, svcChecksWatchIDPrefix+s.proxyCfg.DestinationServiceID, s.ch)
if err != nil {
return err
}
connect: reconcile how upstream configuration works with discovery chains (#6225) * connect: reconcile how upstream configuration works with discovery chains The following upstream config fields for connect sidecars sanely integrate into discovery chain resolution: - Destination Namespace/Datacenter: Compilation occurs locally but using different default values for namespaces and datacenters. The xDS clusters that are created are named as they normally would be. - Mesh Gateway Mode (single upstream): If set this value overrides any value computed for any resolver for the entire discovery chain. The xDS clusters that are created may be named differently (see below). - Mesh Gateway Mode (whole sidecar): If set this value overrides any value computed for any resolver for the entire discovery chain. If this is specifically overridden for a single upstream this value is ignored in that case. The xDS clusters that are created may be named differently (see below). - Protocol (in opaque config): If set this value overrides the value computed when evaluating the entire discovery chain. If the normal chain would be TCP or if this override is set to TCP then the result is that we explicitly disable L7 Routing and Splitting. The xDS clusters that are created may be named differently (see below). - Connect Timeout (in opaque config): If set this value overrides the value for any resolver in the entire discovery chain. The xDS clusters that are created may be named differently (see below). If any of the above overrides affect the actual result of compiling the discovery chain (i.e. "tcp" becomes "grpc" instead of being a no-op override to "tcp") then the relevant parameters are hashed and provided to the xDS layer as a prefix for use in naming the Clusters. This is to ensure that if one Upstream discovery chain has no overrides and tangentially needs a cluster named "api.default.XXX", and another Upstream does have overrides for "api.default.XXX" that they won't cross-pollinate against the operator's wishes. Fixes #6159
2019-08-02 03:03:34 +00:00
// TODO(namespaces): pull this from something like s.source.Namespace?
currentNamespace := "default"
// Watch for updates to service endpoints for all upstreams
for _, u := range s.proxyCfg.Upstreams {
dc := s.source.Datacenter
if u.Datacenter != "" {
// TODO(rb): if we ASK for a specific datacenter, do we still use the chain?
dc = u.Datacenter
}
connect: reconcile how upstream configuration works with discovery chains (#6225) * connect: reconcile how upstream configuration works with discovery chains The following upstream config fields for connect sidecars sanely integrate into discovery chain resolution: - Destination Namespace/Datacenter: Compilation occurs locally but using different default values for namespaces and datacenters. The xDS clusters that are created are named as they normally would be. - Mesh Gateway Mode (single upstream): If set this value overrides any value computed for any resolver for the entire discovery chain. The xDS clusters that are created may be named differently (see below). - Mesh Gateway Mode (whole sidecar): If set this value overrides any value computed for any resolver for the entire discovery chain. If this is specifically overridden for a single upstream this value is ignored in that case. The xDS clusters that are created may be named differently (see below). - Protocol (in opaque config): If set this value overrides the value computed when evaluating the entire discovery chain. If the normal chain would be TCP or if this override is set to TCP then the result is that we explicitly disable L7 Routing and Splitting. The xDS clusters that are created may be named differently (see below). - Connect Timeout (in opaque config): If set this value overrides the value for any resolver in the entire discovery chain. The xDS clusters that are created may be named differently (see below). If any of the above overrides affect the actual result of compiling the discovery chain (i.e. "tcp" becomes "grpc" instead of being a no-op override to "tcp") then the relevant parameters are hashed and provided to the xDS layer as a prefix for use in naming the Clusters. This is to ensure that if one Upstream discovery chain has no overrides and tangentially needs a cluster named "api.default.XXX", and another Upstream does have overrides for "api.default.XXX" that they won't cross-pollinate against the operator's wishes. Fixes #6159
2019-08-02 03:03:34 +00:00
ns := currentNamespace
if u.DestinationNamespace != "" {
ns = u.DestinationNamespace
}
cfg, err := parseReducedUpstreamConfig(u.Config)
if err != nil {
// Don't hard fail on a config typo, just warn. We'll fall back on
// the plain discovery chain if there is an error so it's safe to
// continue.
s.logger.Printf("[WARN] envoy: failed to parse Upstream[%s].Config: %s",
u.Identifier(), err)
}
switch u.DestinationType {
case structs.UpstreamDestTypePreparedQuery:
err = s.cache.Notify(s.ctx, cachetype.PreparedQueryName, &structs.PreparedQueryExecuteRequest{
Datacenter: dc,
QueryOptions: structs.QueryOptions{Token: s.token, MaxAge: defaultPreparedQueryPollInterval},
QueryIDOrName: u.DestinationName,
Connect: true,
Source: *s.source,
}, "upstream:"+u.Identifier(), s.ch)
connect: reconcile how upstream configuration works with discovery chains (#6225) * connect: reconcile how upstream configuration works with discovery chains The following upstream config fields for connect sidecars sanely integrate into discovery chain resolution: - Destination Namespace/Datacenter: Compilation occurs locally but using different default values for namespaces and datacenters. The xDS clusters that are created are named as they normally would be. - Mesh Gateway Mode (single upstream): If set this value overrides any value computed for any resolver for the entire discovery chain. The xDS clusters that are created may be named differently (see below). - Mesh Gateway Mode (whole sidecar): If set this value overrides any value computed for any resolver for the entire discovery chain. If this is specifically overridden for a single upstream this value is ignored in that case. The xDS clusters that are created may be named differently (see below). - Protocol (in opaque config): If set this value overrides the value computed when evaluating the entire discovery chain. If the normal chain would be TCP or if this override is set to TCP then the result is that we explicitly disable L7 Routing and Splitting. The xDS clusters that are created may be named differently (see below). - Connect Timeout (in opaque config): If set this value overrides the value for any resolver in the entire discovery chain. The xDS clusters that are created may be named differently (see below). If any of the above overrides affect the actual result of compiling the discovery chain (i.e. "tcp" becomes "grpc" instead of being a no-op override to "tcp") then the relevant parameters are hashed and provided to the xDS layer as a prefix for use in naming the Clusters. This is to ensure that if one Upstream discovery chain has no overrides and tangentially needs a cluster named "api.default.XXX", and another Upstream does have overrides for "api.default.XXX" that they won't cross-pollinate against the operator's wishes. Fixes #6159
2019-08-02 03:03:34 +00:00
case structs.UpstreamDestTypeService:
fallthrough
connect: reconcile how upstream configuration works with discovery chains (#6225) * connect: reconcile how upstream configuration works with discovery chains The following upstream config fields for connect sidecars sanely integrate into discovery chain resolution: - Destination Namespace/Datacenter: Compilation occurs locally but using different default values for namespaces and datacenters. The xDS clusters that are created are named as they normally would be. - Mesh Gateway Mode (single upstream): If set this value overrides any value computed for any resolver for the entire discovery chain. The xDS clusters that are created may be named differently (see below). - Mesh Gateway Mode (whole sidecar): If set this value overrides any value computed for any resolver for the entire discovery chain. If this is specifically overridden for a single upstream this value is ignored in that case. The xDS clusters that are created may be named differently (see below). - Protocol (in opaque config): If set this value overrides the value computed when evaluating the entire discovery chain. If the normal chain would be TCP or if this override is set to TCP then the result is that we explicitly disable L7 Routing and Splitting. The xDS clusters that are created may be named differently (see below). - Connect Timeout (in opaque config): If set this value overrides the value for any resolver in the entire discovery chain. The xDS clusters that are created may be named differently (see below). If any of the above overrides affect the actual result of compiling the discovery chain (i.e. "tcp" becomes "grpc" instead of being a no-op override to "tcp") then the relevant parameters are hashed and provided to the xDS layer as a prefix for use in naming the Clusters. This is to ensure that if one Upstream discovery chain has no overrides and tangentially needs a cluster named "api.default.XXX", and another Upstream does have overrides for "api.default.XXX" that they won't cross-pollinate against the operator's wishes. Fixes #6159
2019-08-02 03:03:34 +00:00
case "": // Treat unset as the default Service type
err = s.cache.Notify(s.ctx, cachetype.CompiledDiscoveryChainName, &structs.DiscoveryChainRequest{
Datacenter: s.source.Datacenter,
QueryOptions: structs.QueryOptions{Token: s.token},
Name: u.DestinationName,
EvaluateInDatacenter: dc,
EvaluateInNamespace: ns,
OverrideMeshGateway: s.proxyCfg.MeshGateway.OverlayWith(u.MeshGateway),
OverrideProtocol: cfg.Protocol,
OverrideConnectTimeout: cfg.ConnectTimeout(),
}, "discovery-chain:"+u.Identifier(), s.ch)
if err != nil {
return err
}
default:
return fmt.Errorf("unknown upstream type: %q", u.DestinationType)
}
}
return nil
}
connect: reconcile how upstream configuration works with discovery chains (#6225) * connect: reconcile how upstream configuration works with discovery chains The following upstream config fields for connect sidecars sanely integrate into discovery chain resolution: - Destination Namespace/Datacenter: Compilation occurs locally but using different default values for namespaces and datacenters. The xDS clusters that are created are named as they normally would be. - Mesh Gateway Mode (single upstream): If set this value overrides any value computed for any resolver for the entire discovery chain. The xDS clusters that are created may be named differently (see below). - Mesh Gateway Mode (whole sidecar): If set this value overrides any value computed for any resolver for the entire discovery chain. If this is specifically overridden for a single upstream this value is ignored in that case. The xDS clusters that are created may be named differently (see below). - Protocol (in opaque config): If set this value overrides the value computed when evaluating the entire discovery chain. If the normal chain would be TCP or if this override is set to TCP then the result is that we explicitly disable L7 Routing and Splitting. The xDS clusters that are created may be named differently (see below). - Connect Timeout (in opaque config): If set this value overrides the value for any resolver in the entire discovery chain. The xDS clusters that are created may be named differently (see below). If any of the above overrides affect the actual result of compiling the discovery chain (i.e. "tcp" becomes "grpc" instead of being a no-op override to "tcp") then the relevant parameters are hashed and provided to the xDS layer as a prefix for use in naming the Clusters. This is to ensure that if one Upstream discovery chain has no overrides and tangentially needs a cluster named "api.default.XXX", and another Upstream does have overrides for "api.default.XXX" that they won't cross-pollinate against the operator's wishes. Fixes #6159
2019-08-02 03:03:34 +00:00
// reducedProxyConfig represents the basic opaque config values that are now
// managed with the discovery chain but for backwards compatibility reasons
// should still affect how the proxy is configured.
//
// The full-blown config is agent/xds.UpstreamConfig
type reducedUpstreamConfig struct {
Protocol string `mapstructure:"protocol"`
ConnectTimeoutMs int `mapstructure:"connect_timeout_ms"`
}
func (c *reducedUpstreamConfig) ConnectTimeout() time.Duration {
return time.Duration(c.ConnectTimeoutMs) * time.Millisecond
}
func parseReducedUpstreamConfig(m map[string]interface{}) (reducedUpstreamConfig, error) {
var cfg reducedUpstreamConfig
err := mapstructure.WeakDecode(m, &cfg)
return cfg, err
}
// initWatchesMeshGateway sets up the watches needed based on the current mesh gateway registration
func (s *state) initWatchesMeshGateway() error {
// Watch for root changes
err := s.cache.Notify(s.ctx, cachetype.ConnectCARootName, &structs.DCSpecificRequest{
Datacenter: s.source.Datacenter,
QueryOptions: structs.QueryOptions{Token: s.token},
Source: *s.source,
}, rootsWatchID, s.ch)
if err != nil {
return err
}
// Watch for all services
err = s.cache.Notify(s.ctx, cachetype.CatalogListServicesName, &structs.DCSpecificRequest{
Datacenter: s.source.Datacenter,
QueryOptions: structs.QueryOptions{Token: s.token},
Source: *s.source,
}, serviceListWatchID, s.ch)
if err != nil {
return err
}
// Eventually we will have to watch connect enable instances for each service as well as the
// destination services themselves but those notifications will be setup later. However we
// cannot setup those watches until we know what the services are. from the service list
// watch above
err = s.cache.Notify(s.ctx, cachetype.CatalogDatacentersName, &structs.DatacentersRequest{
QueryOptions: structs.QueryOptions{Token: s.token, MaxAge: 30 * time.Second},
}, datacentersWatchID, s.ch)
// Once we start getting notified about the datacenters we will setup watches on the
// gateways within those other datacenters. We cannot do that here because we don't
// know what they are yet.
// Watch service-resolvers so we can setup service subset clusters
err = s.cache.Notify(s.ctx, cachetype.ConfigEntriesName, &structs.ConfigEntryQuery{
Datacenter: s.source.Datacenter,
QueryOptions: structs.QueryOptions{Token: s.token},
Kind: structs.ServiceResolver,
}, serviceResolversWatchID, s.ch)
if err != nil {
s.logger.Printf("[ERR] mesh-gateway: failed to register watch for service-resolver config entries")
return err
}
return err
}
func (s *state) initialConfigSnapshot() ConfigSnapshot {
snap := ConfigSnapshot{
Kind: s.kind,
Service: s.service,
ProxyID: s.proxyID,
Address: s.address,
Port: s.port,
TaggedAddresses: s.taggedAddresses,
Proxy: s.proxyCfg,
Datacenter: s.source.Datacenter,
}
switch s.kind {
case structs.ServiceKindConnectProxy:
snap.ConnectProxy.DiscoveryChain = make(map[string]*structs.CompiledDiscoveryChain)
snap.ConnectProxy.WatchedUpstreams = make(map[string]map[string]context.CancelFunc)
snap.ConnectProxy.WatchedUpstreamEndpoints = make(map[string]map[string]structs.CheckServiceNodes)
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
snap.ConnectProxy.WatchedGateways = make(map[string]map[string]context.CancelFunc)
snap.ConnectProxy.WatchedGatewayEndpoints = make(map[string]map[string]structs.CheckServiceNodes)
snap.ConnectProxy.WatchedServiceChecks = make(map[string][]structs.CheckType)
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
snap.ConnectProxy.UpstreamEndpoints = make(map[string]structs.CheckServiceNodes) // TODO(rb): deprecated
case structs.ServiceKindMeshGateway:
snap.MeshGateway.WatchedServices = make(map[string]context.CancelFunc)
snap.MeshGateway.WatchedDatacenters = make(map[string]context.CancelFunc)
snap.MeshGateway.ServiceGroups = make(map[string]structs.CheckServiceNodes)
snap.MeshGateway.GatewayGroups = make(map[string]structs.CheckServiceNodes)
// there is no need to initialize the map of service resolvers as we
// fully rebuild it every time we get updates
}
return snap
}
func (s *state) run() {
// Close the channel we return from Watch when we stop so consumers can stop
// watching and clean up their goroutines. It's important we do this here and
// not in Close since this routine sends on this chan and so might panic if it
// gets closed from another goroutine.
defer close(s.snapCh)
snap := s.initialConfigSnapshot()
// This turns out to be really fiddly/painful by just using time.Timer.C
// directly in the code below since you can't detect when a timer is stopped
// vs waiting in order to know to reset it. So just use a chan to send
// ourselves messages.
sendCh := make(chan struct{})
var coalesceTimer *time.Timer
for {
select {
case <-s.ctx.Done():
return
case u := <-s.ch:
if err := s.handleUpdate(u, &snap); err != nil {
s.logger.Printf("[ERR] %s watch error: %s", u.CorrelationID, err)
continue
}
case <-sendCh:
// Make a deep copy of snap so we don't mutate any of the embedded structs
// etc on future updates.
snapCopy, err := snap.Clone()
if err != nil {
s.logger.Printf("[ERR] Failed to copy config snapshot for proxy %s",
s.proxyID)
continue
}
s.snapCh <- *snapCopy
// Allow the next change to trigger a send
coalesceTimer = nil
// Skip rest of loop - there is nothing to send since nothing changed on
// this iteration
continue
case replyCh := <-s.reqCh:
if !snap.Valid() {
// Not valid yet just respond with nil and move on to next task.
replyCh <- nil
continue
}
// Make a deep copy of snap so we don't mutate any of the embedded structs
// etc on future updates.
snapCopy, err := snap.Clone()
if err != nil {
s.logger.Printf("[ERR] Failed to copy config snapshot for proxy %s",
s.proxyID)
continue
}
replyCh <- snapCopy
// Skip rest of loop - there is nothing to send since nothing changed on
// this iteration
continue
}
// Check if snap is complete enough to be a valid config to deliver to a
// proxy yet.
if snap.Valid() {
// Don't send it right away, set a short timer that will wait for updates
// from any of the other cache values and deliver them all together.
if coalesceTimer == nil {
coalesceTimer = time.AfterFunc(coalesceTimeout, func() {
// This runs in another goroutine so we can't just do the send
// directly here as access to snap is racy. Instead, signal the main
// loop above.
sendCh <- struct{}{}
})
}
}
}
}
func (s *state) handleUpdate(u cache.UpdateEvent, snap *ConfigSnapshot) error {
switch s.kind {
case structs.ServiceKindConnectProxy:
return s.handleUpdateConnectProxy(u, snap)
case structs.ServiceKindMeshGateway:
return s.handleUpdateMeshGateway(u, snap)
default:
return fmt.Errorf("Unsupported service kind")
}
}
func (s *state) handleUpdateConnectProxy(u cache.UpdateEvent, snap *ConfigSnapshot) error {
if u.Err != nil {
return fmt.Errorf("error filling agent cache: %v", u.Err)
}
switch {
case u.CorrelationID == rootsWatchID:
roots, ok := u.Result.(*structs.IndexedCARoots)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
snap.Roots = roots
case u.CorrelationID == leafWatchID:
leaf, ok := u.Result.(*structs.IssuedCert)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
snap.ConnectProxy.Leaf = leaf
case u.CorrelationID == intentionsWatchID:
// Not in snapshot currently, no op
case strings.HasPrefix(u.CorrelationID, "discovery-chain:"):
resp, ok := u.Result.(*structs.DiscoveryChainResponse)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
svc := strings.TrimPrefix(u.CorrelationID, "discovery-chain:")
snap.ConnectProxy.DiscoveryChain[svc] = resp.Chain
if err := s.resetWatchesFromChain(svc, resp.Chain, snap); err != nil {
return err
}
case strings.HasPrefix(u.CorrelationID, "upstream-target:"):
resp, ok := u.Result.(*structs.IndexedCheckServiceNodes)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
correlationID := strings.TrimPrefix(u.CorrelationID, "upstream-target:")
targetID, svc, ok := removeColonPrefix(correlationID)
if !ok {
return fmt.Errorf("invalid correlation id %q", u.CorrelationID)
}
m, ok := snap.ConnectProxy.WatchedUpstreamEndpoints[svc]
if !ok {
m = make(map[string]structs.CheckServiceNodes)
snap.ConnectProxy.WatchedUpstreamEndpoints[svc] = m
}
snap.ConnectProxy.WatchedUpstreamEndpoints[svc][targetID] = resp.Nodes
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
case strings.HasPrefix(u.CorrelationID, "mesh-gateway:"):
resp, ok := u.Result.(*structs.IndexedCheckServiceNodes)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
}
correlationID := strings.TrimPrefix(u.CorrelationID, "mesh-gateway:")
dc, svc, ok := removeColonPrefix(correlationID)
if !ok {
return fmt.Errorf("invalid correlation id %q", u.CorrelationID)
}
m, ok := snap.ConnectProxy.WatchedGatewayEndpoints[svc]
if !ok {
m = make(map[string]structs.CheckServiceNodes)
snap.ConnectProxy.WatchedGatewayEndpoints[svc] = m
}
snap.ConnectProxy.WatchedGatewayEndpoints[svc][dc] = resp.Nodes
case strings.HasPrefix(u.CorrelationID, "upstream:"+serviceIDPrefix):
resp, ok := u.Result.(*structs.IndexedCheckServiceNodes)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
svc := strings.TrimPrefix(u.CorrelationID, "upstream:"+serviceIDPrefix)
snap.ConnectProxy.UpstreamEndpoints[svc] = resp.Nodes
case strings.HasPrefix(u.CorrelationID, "upstream:"+preparedQueryIDPrefix):
resp, ok := u.Result.(*structs.PreparedQueryExecuteResponse)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
pq := strings.TrimPrefix(u.CorrelationID, "upstream:")
snap.ConnectProxy.UpstreamEndpoints[pq] = resp.Nodes
case strings.HasPrefix(u.CorrelationID, svcChecksWatchIDPrefix):
resp, ok := u.Result.([]structs.CheckType)
if !ok {
return fmt.Errorf("invalid type for service checks response: %T, want: []structs.CheckType", u.Result)
}
svcID := strings.TrimPrefix(u.CorrelationID, svcChecksWatchIDPrefix)
snap.ConnectProxy.WatchedServiceChecks[svcID] = resp
default:
return fmt.Errorf("unknown correlation ID: %s", u.CorrelationID)
}
return nil
}
func removeColonPrefix(s string) (string, string, bool) {
idx := strings.Index(s, ":")
if idx == -1 {
return "", "", false
}
return s[0:idx], s[idx+1:], true
}
func (s *state) resetWatchesFromChain(
id string,
chain *structs.CompiledDiscoveryChain,
snap *ConfigSnapshot,
) error {
if chain == nil {
return fmt.Errorf("not possible to arrive here with no discovery chain")
}
// Initialize relevant sub maps.
if _, ok := snap.ConnectProxy.WatchedUpstreams[id]; !ok {
snap.ConnectProxy.WatchedUpstreams[id] = make(map[string]context.CancelFunc)
}
if _, ok := snap.ConnectProxy.WatchedUpstreamEndpoints[id]; !ok {
snap.ConnectProxy.WatchedUpstreamEndpoints[id] = make(map[string]structs.CheckServiceNodes)
}
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
if _, ok := snap.ConnectProxy.WatchedGateways[id]; !ok {
snap.ConnectProxy.WatchedGateways[id] = make(map[string]context.CancelFunc)
}
if _, ok := snap.ConnectProxy.WatchedGatewayEndpoints[id]; !ok {
snap.ConnectProxy.WatchedGatewayEndpoints[id] = make(map[string]structs.CheckServiceNodes)
}
// We could invalidate this selectively based on a hash of the relevant
// resolver information, but for now just reset anything about this
// upstream when the chain changes in any way.
//
// TODO(rb): content hash based add/remove
for targetID, cancelFn := range snap.ConnectProxy.WatchedUpstreams[id] {
s.logger.Printf("[TRACE] proxycfg: upstream=%q:chain=%q: stopping watch of target %s", id, chain.ServiceName, targetID)
delete(snap.ConnectProxy.WatchedUpstreams[id], targetID)
delete(snap.ConnectProxy.WatchedUpstreamEndpoints[id], targetID)
cancelFn()
}
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
needGateways := make(map[string]struct{})
for _, target := range chain.Targets {
s.logger.Printf("[TRACE] proxycfg: upstream=%q:chain=%q: initializing watch of target %s", id, chain.ServiceName, target.ID)
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
// We'll get endpoints from the gateway query, but the health still has
// to come from the backing service query.
switch target.MeshGateway.Mode {
case structs.MeshGatewayModeRemote:
needGateways[target.Datacenter] = struct{}{}
case structs.MeshGatewayModeLocal:
needGateways[s.source.Datacenter] = struct{}{}
}
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
ctx, cancel := context.WithCancel(s.ctx)
err := s.watchConnectProxyService(
ctx,
"upstream-target:"+target.ID+":"+id,
target.Service,
target.Datacenter,
target.Subset.Filter,
)
if err != nil {
cancel()
return err
}
snap.ConnectProxy.WatchedUpstreams[id][target.ID] = cancel
}
connect: fix failover through a mesh gateway to a remote datacenter (#6259) Failover is pushed entirely down to the data plane by creating envoy clusters and putting each successive destination in a different load assignment priority band. For example this shows that normally requests go to 1.2.3.4:8080 but when that fails they go to 6.7.8.9:8080: - name: foo load_assignment: cluster_name: foo policy: overprovisioning_factor: 100000 endpoints: - priority: 0 lb_endpoints: - endpoint: address: socket_address: address: 1.2.3.4 port_value: 8080 - priority: 1 lb_endpoints: - endpoint: address: socket_address: address: 6.7.8.9 port_value: 8080 Mesh gateways route requests based solely on the SNI header tacked onto the TLS layer. Envoy currently only lets you configure the outbound SNI header at the cluster layer. If you try to failover through a mesh gateway you ideally would configure the SNI value per endpoint, but that's not possible in envoy today. This PR introduces a simpler way around the problem for now: 1. We identify any target of failover that will use mesh gateway mode local or remote and then further isolate any resolver node in the compiled discovery chain that has a failover destination set to one of those targets. 2. For each of these resolvers we will perform a small measurement of comparative healths of the endpoints that come back from the health API for the set of primary target and serial failover targets. We walk the list of targets in order and if any endpoint is healthy we return that target, otherwise we move on to the next target. 3. The CDS and EDS endpoints both perform the measurements in (2) for the affected resolver nodes. 4. For CDS this measurement selects which TLS SNI field to use for the cluster (note the cluster is always going to be named for the primary target) 5. For EDS this measurement selects which set of endpoints will populate the cluster. Priority tiered failover is ignored. One of the big downsides to this approach to failover is that the failover detection and correction is going to be controlled by consul rather than deferring that entirely to the data plane as with the prior version. This also means that we are bound to only failover using official health signals and cannot make use of data plane signals like outlier detection to affect failover. In this specific scenario the lack of data plane signals is ok because the effectiveness is already muted by the fact that the ultimate destination endpoints will have their data plane signals scrambled when they pass through the mesh gateway wrapper anyway so we're not losing much. Another related fix is that we now use the endpoint health from the underlying service, not the health of the gateway (regardless of failover mode).
2019-08-05 18:30:35 +00:00
for dc, _ := range needGateways {
if _, ok := snap.ConnectProxy.WatchedGateways[id][dc]; ok {
continue
}
s.logger.Printf("[TRACE] proxycfg: upstream=%q:chain=%q: initializing watch of mesh gateway in dc %s", id, chain.ServiceName, dc)
ctx, cancel := context.WithCancel(s.ctx)
err := s.watchMeshGateway(ctx, dc, id)
if err != nil {
cancel()
return err
}
snap.ConnectProxy.WatchedGateways[id][dc] = cancel
}
for dc, cancelFn := range snap.ConnectProxy.WatchedGateways[id] {
if _, ok := needGateways[dc]; ok {
continue
}
s.logger.Printf("[TRACE] proxycfg: upstream=%q:chain=%q: stopping watch of mesh gateway in dc %s", id, chain.ServiceName, dc)
delete(snap.ConnectProxy.WatchedGateways[id], dc)
delete(snap.ConnectProxy.WatchedGatewayEndpoints[id], dc)
cancelFn()
}
return nil
}
func (s *state) handleUpdateMeshGateway(u cache.UpdateEvent, snap *ConfigSnapshot) error {
if u.Err != nil {
return fmt.Errorf("error filling agent cache: %v", u.Err)
}
switch u.CorrelationID {
case rootsWatchID:
roots, ok := u.Result.(*structs.IndexedCARoots)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
snap.Roots = roots
case serviceListWatchID:
services, ok := u.Result.(*structs.IndexedServices)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
for svcName := range services.Services {
if _, ok := snap.MeshGateway.WatchedServices[svcName]; !ok {
ctx, cancel := context.WithCancel(s.ctx)
err := s.cache.Notify(ctx, cachetype.HealthServicesName, &structs.ServiceSpecificRequest{
Datacenter: s.source.Datacenter,
QueryOptions: structs.QueryOptions{Token: s.token},
ServiceName: svcName,
Connect: true,
}, fmt.Sprintf("connect-service:%s", svcName), s.ch)
if err != nil {
s.logger.Printf("[ERR] mesh-gateway: failed to register watch for connect-service:%s", svcName)
cancel()
return err
}
snap.MeshGateway.WatchedServices[svcName] = cancel
}
}
for svcName, cancelFn := range snap.MeshGateway.WatchedServices {
if _, ok := services.Services[svcName]; !ok {
delete(snap.MeshGateway.WatchedServices, svcName)
cancelFn()
}
}
snap.MeshGateway.WatchedServicesSet = true
case datacentersWatchID:
datacentersRaw, ok := u.Result.(*[]string)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
if datacentersRaw == nil {
return fmt.Errorf("invalid response with a nil datacenter list")
}
datacenters := *datacentersRaw
for _, dc := range datacenters {
if dc == s.source.Datacenter {
continue
}
if _, ok := snap.MeshGateway.WatchedDatacenters[dc]; !ok {
ctx, cancel := context.WithCancel(s.ctx)
err := s.cache.Notify(ctx, cachetype.InternalServiceDumpName, &structs.ServiceDumpRequest{
Datacenter: dc,
QueryOptions: structs.QueryOptions{Token: s.token},
ServiceKind: structs.ServiceKindMeshGateway,
UseServiceKind: true,
Source: *s.source,
}, fmt.Sprintf("mesh-gateway:%s", dc), s.ch)
if err != nil {
s.logger.Printf("[ERR] mesh-gateway: failed to register watch for mesh-gateway:%s", dc)
cancel()
return err
}
snap.MeshGateway.WatchedDatacenters[dc] = cancel
}
}
for dc, cancelFn := range snap.MeshGateway.WatchedDatacenters {
found := false
for _, dcCurrent := range datacenters {
if dcCurrent == dc {
found = true
break
}
}
if !found {
delete(snap.MeshGateway.WatchedDatacenters, dc)
cancelFn()
}
}
case serviceResolversWatchID:
configEntries, ok := u.Result.(*structs.IndexedConfigEntries)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
resolvers := make(map[string]*structs.ServiceResolverConfigEntry)
for _, entry := range configEntries.Entries {
if resolver, ok := entry.(*structs.ServiceResolverConfigEntry); ok {
resolvers[resolver.Name] = resolver
}
}
snap.MeshGateway.ServiceResolvers = resolvers
default:
switch {
case strings.HasPrefix(u.CorrelationID, "connect-service:"):
resp, ok := u.Result.(*structs.IndexedCheckServiceNodes)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
svc := strings.TrimPrefix(u.CorrelationID, "connect-service:")
if len(resp.Nodes) > 0 {
snap.MeshGateway.ServiceGroups[svc] = resp.Nodes
} else if _, ok := snap.MeshGateway.ServiceGroups[svc]; ok {
delete(snap.MeshGateway.ServiceGroups, svc)
}
case strings.HasPrefix(u.CorrelationID, "mesh-gateway:"):
resp, ok := u.Result.(*structs.IndexedCheckServiceNodes)
if !ok {
return fmt.Errorf("invalid type for response: %T", u.Result)
}
dc := strings.TrimPrefix(u.CorrelationID, "mesh-gateway:")
if len(resp.Nodes) > 0 {
snap.MeshGateway.GatewayGroups[dc] = resp.Nodes
} else if _, ok := snap.MeshGateway.GatewayGroups[dc]; ok {
delete(snap.MeshGateway.GatewayGroups, dc)
}
default:
// do nothing for now
}
}
return nil
}
// CurrentSnapshot synchronously returns the current ConfigSnapshot if there is
// one ready. If we don't have one yet because not all necessary parts have been
// returned (i.e. both roots and leaf cert), nil is returned.
func (s *state) CurrentSnapshot() *ConfigSnapshot {
// Make a chan for the response to be sent on
ch := make(chan *ConfigSnapshot, 1)
s.reqCh <- ch
// Wait for the response
return <-ch
}
// Changed returns whether or not the passed NodeService has had any of the
// fields we care about for config state watching changed or a different token.
func (s *state) Changed(ns *structs.NodeService, token string) bool {
if ns == nil {
return true
}
return ns.Kind != s.kind ||
s.proxyID != ns.ID ||
s.address != ns.Address ||
s.port != ns.Port ||
!reflect.DeepEqual(s.proxyCfg, ns.Proxy) ||
s.token != token
}