open-consul/agent/consul/discoverychain/compile.go

1027 lines
28 KiB
Go
Raw Normal View History

package discoverychain
import (
"fmt"
"strings"
"time"
"github.com/hashicorp/consul/agent/connect"
"github.com/hashicorp/consul/agent/structs"
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/hashstructure"
"github.com/mitchellh/mapstructure"
)
type CompileRequest struct {
ServiceName string
EvaluateInNamespace string
EvaluateInDatacenter string
EvaluateInTrustDomain string
UseInDatacenter string // where the results will be used from
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
// OverrideMeshGateway allows for the setting to be overridden for any
// resolver in the compiled chain.
OverrideMeshGateway structs.MeshGatewayConfig
// OverrideProtocol allows for the final protocol for the chain to be
// altered.
//
// - If the chain ordinarily would be TCP and an L7 protocol is passed here
// the chain will not include Routers or Splitters.
//
// - If the chain ordinarily would be L7 and TCP is passed here the chain
// will not include Routers or Splitters.
OverrideProtocol string
// OverrideConnectTimeout allows for the ConnectTimeout setting to be
// overridden for any resolver in the compiled chain.
OverrideConnectTimeout time.Duration
Entries *structs.DiscoveryChainConfigEntries
}
// Compile assembles a discovery chain in the form of a graph of nodes using
// raw config entries and local context.
//
// "Node" referenced in this file refers to a node in a graph and not to the
// Consul construct called a "Node".
//
// Omitting router and splitter entries for services not using an L7 protocol
// (like HTTP) happens during initial fetching, but for sanity purposes a quick
// reinforcement of that happens here, too.
//
// May return a *structs.ConfigEntryGraphError, but that is only expected when
// being used to validate modifications to the config entry graph. It should
// not be expected when compiling existing entries at runtime that are already
// valid.
func Compile(req CompileRequest) (*structs.CompiledDiscoveryChain, error) {
var (
serviceName = req.ServiceName
evaluateInNamespace = req.EvaluateInNamespace
evaluateInDatacenter = req.EvaluateInDatacenter
evaluateInTrustDomain = req.EvaluateInTrustDomain
useInDatacenter = req.UseInDatacenter
entries = req.Entries
)
if serviceName == "" {
return nil, fmt.Errorf("serviceName is required")
}
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 evaluateInNamespace == "" {
return nil, fmt.Errorf("evaluateInNamespace is required")
}
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 evaluateInDatacenter == "" {
return nil, fmt.Errorf("evaluateInDatacenter is required")
}
if evaluateInTrustDomain == "" {
return nil, fmt.Errorf("evaluateInTrustDomain is required")
}
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 useInDatacenter == "" {
return nil, fmt.Errorf("useInDatacenter is required")
}
if entries == nil {
return nil, fmt.Errorf("entries is required")
}
c := &compiler{
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
serviceName: serviceName,
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
evaluateInNamespace: evaluateInNamespace,
evaluateInDatacenter: evaluateInDatacenter,
evaluateInTrustDomain: evaluateInTrustDomain,
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
useInDatacenter: useInDatacenter,
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
overrideMeshGateway: req.OverrideMeshGateway,
overrideProtocol: req.OverrideProtocol,
overrideConnectTimeout: req.OverrideConnectTimeout,
entries: entries,
resolvers: make(map[structs.ServiceID]*structs.ServiceResolverConfigEntry),
splitterNodes: make(map[string]*structs.DiscoveryGraphNode),
resolveNodes: make(map[string]*structs.DiscoveryGraphNode),
nodes: make(map[string]*structs.DiscoveryGraphNode),
loadedTargets: make(map[string]*structs.DiscoveryTarget),
retainedTargets: make(map[string]struct{}),
}
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
if req.OverrideProtocol != "" {
c.disableAdvancedRoutingFeatures = !enableAdvancedRoutingForProtocol(req.OverrideProtocol)
}
// Clone this resolver map to avoid mutating the input map during compilation.
if len(entries.Resolvers) > 0 {
for k, v := range entries.Resolvers {
c.resolvers[k] = v
}
}
return c.compile()
}
// compiler is a single-use struct for handling intermediate state necessary
// for assembling a discovery chain from raw config entries.
type compiler struct {
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
serviceName string
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
evaluateInNamespace string
evaluateInDatacenter string
evaluateInTrustDomain string
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
useInDatacenter string
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
overrideMeshGateway structs.MeshGatewayConfig
overrideProtocol string
overrideConnectTimeout time.Duration
// config entries that are being compiled (will be mutated during compilation)
//
// This is an INPUT field.
entries *structs.DiscoveryChainConfigEntries
// resolvers is initially seeded by copying the provided entries.Resolvers
// map and default resolvers are added as they are needed.
resolvers map[structs.ServiceID]*structs.ServiceResolverConfigEntry
// cached nodes
splitterNodes map[string]*structs.DiscoveryGraphNode
resolveNodes map[string]*structs.DiscoveryGraphNode
// usesAdvancedRoutingFeatures is set to true if config entries for routing
// or splitting appear in the compiled chain
usesAdvancedRoutingFeatures bool
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
// disableAdvancedRoutingFeatures is set to true if overrideProtocol is set to tcp
disableAdvancedRoutingFeatures bool
// customizedBy indicates which override values customized how the
// compilation behaved.
//
// This is an OUTPUT field.
customizedBy customizationMarkers
// protocol is the common protocol used for all referenced services. These
// cannot be mixed.
//
// This is an OUTPUT field.
protocol string
// startNode is computed inside of assembleChain()
//
// This is an OUTPUT field.
startNode string
// nodes is computed inside of compile()
//
// This is an OUTPUT field.
nodes map[string]*structs.DiscoveryGraphNode
// This is an OUTPUT field.
loadedTargets map[string]*structs.DiscoveryTarget
retainedTargets map[string]struct{}
}
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
type customizationMarkers struct {
MeshGateway bool
Protocol bool
ConnectTimeout bool
}
// serviceIDString deviates from the standard formatting you would get with
// the String() method on the type itself. It is this way to be more
// consistent with other string ids within the discovery chain.
func serviceIDString(sid structs.ServiceID) string {
return fmt.Sprintf("%s.%s", sid.ID, sid.NamespaceOrDefault())
}
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
func (m *customizationMarkers) IsZero() bool {
return !m.MeshGateway && !m.Protocol && !m.ConnectTimeout
}
// recordNode stores the node internally in the compiled chain.
func (c *compiler) recordNode(node *structs.DiscoveryGraphNode) {
// Some types have their own type-specific lookups, so record those, too.
switch node.Type {
case structs.DiscoveryGraphNodeTypeRouter:
// no special storage
case structs.DiscoveryGraphNodeTypeSplitter:
c.splitterNodes[node.Name] = node
case structs.DiscoveryGraphNodeTypeResolver:
c.resolveNodes[node.Resolver.Target] = node
default:
panic("unknown node type '" + node.Type + "'")
}
c.nodes[node.MapKey()] = node
}
func (c *compiler) recordServiceProtocol(sid structs.ServiceID) error {
if serviceDefault := c.entries.GetService(sid); serviceDefault != nil {
return c.recordProtocol(sid, serviceDefault.Protocol)
}
if c.entries.GlobalProxy != nil {
var cfg proxyConfig
// Ignore errors and fallback on defaults if it does happen.
_ = mapstructure.WeakDecode(c.entries.GlobalProxy.Config, &cfg)
if cfg.Protocol != "" {
return c.recordProtocol(sid, cfg.Protocol)
}
}
return c.recordProtocol(sid, "")
}
// proxyConfig is a snippet from agent/xds/config.go:ProxyConfig
type proxyConfig struct {
Protocol string `mapstructure:"protocol"`
}
func (c *compiler) recordProtocol(fromService structs.ServiceID, protocol string) error {
if protocol == "" {
protocol = "tcp"
} else {
protocol = strings.ToLower(protocol)
}
if c.protocol == "" {
c.protocol = protocol
} else if c.protocol != protocol {
return &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"discovery chain %q uses inconsistent protocols; service %q has %q which is not %q",
c.serviceName, fromService.String(), protocol, c.protocol,
),
}
}
return nil
}
func (c *compiler) compile() (*structs.CompiledDiscoveryChain, error) {
if err := c.assembleChain(); err != nil {
return nil, err
}
// We don't need these intermediates anymore.
c.splitterNodes = nil
c.resolveNodes = nil
if c.startNode == "" {
panic("impossible to return no results")
}
if err := c.detectCircularReferences(); err != nil {
return nil, err
}
c.flattenAdjacentSplitterNodes()
if err := c.removeUnusedNodes(); err != nil {
return nil, err
}
for targetID := range c.loadedTargets {
if _, ok := c.retainedTargets[targetID]; !ok {
delete(c.loadedTargets, targetID)
}
}
if !enableAdvancedRoutingForProtocol(c.protocol) && c.usesAdvancedRoutingFeatures {
return nil, &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"discovery chain %q uses a protocol %q that does not permit advanced routing or splitting behavior",
c.serviceName, c.protocol,
),
}
}
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
if c.overrideProtocol != "" {
if c.overrideProtocol != c.protocol {
c.protocol = c.overrideProtocol
c.customizedBy.Protocol = true
}
}
var customizationHash string
if !c.customizedBy.IsZero() {
var customization struct {
OverrideMeshGateway structs.MeshGatewayConfig
OverrideProtocol string
OverrideConnectTimeout time.Duration
}
if c.customizedBy.MeshGateway {
customization.OverrideMeshGateway = c.overrideMeshGateway
}
if c.customizedBy.Protocol {
customization.OverrideProtocol = c.overrideProtocol
}
if c.customizedBy.ConnectTimeout {
customization.OverrideConnectTimeout = c.overrideConnectTimeout
}
v, err := hashstructure.Hash(customization, nil)
if err != nil {
return nil, fmt.Errorf("cannot create customization hash key: %v", err)
}
customizationHash = fmt.Sprintf("%x", v)[0:8]
}
return &structs.CompiledDiscoveryChain{
ServiceName: c.serviceName,
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
Namespace: c.evaluateInNamespace,
Datacenter: c.evaluateInDatacenter,
CustomizationHash: customizationHash,
Protocol: c.protocol,
StartNode: c.startNode,
Nodes: c.nodes,
Targets: c.loadedTargets,
}, nil
}
func (c *compiler) detectCircularReferences() error {
var (
todo stringStack
visited = make(map[string]struct{})
visitChain stringStack
)
todo.Push(c.startNode)
for {
current, ok := todo.Pop()
if !ok {
break
}
if current == "_popvisit" {
if v, ok := visitChain.Pop(); ok {
delete(visited, v)
}
continue
}
visitChain.Push(current)
if _, ok := visited[current]; ok {
return &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"detected circular reference: [%s]",
strings.Join(visitChain.Items(), " -> "),
),
}
}
visited[current] = struct{}{}
todo.Push("_popvisit")
node := c.nodes[current]
switch node.Type {
case structs.DiscoveryGraphNodeTypeRouter:
for _, route := range node.Routes {
todo.Push(route.NextNode)
}
case structs.DiscoveryGraphNodeTypeSplitter:
for _, split := range node.Splits {
todo.Push(split.NextNode)
}
case structs.DiscoveryGraphNodeTypeResolver:
// Circular redirects are detected elsewhere and failover isn't
// recursive so there's nothing more to do here.
default:
return fmt.Errorf("unexpected graph node type: %s", node.Type)
}
}
return nil
}
func (c *compiler) flattenAdjacentSplitterNodes() {
for {
anyChanged := false
for _, node := range c.nodes {
if node.Type != structs.DiscoveryGraphNodeTypeSplitter {
continue
}
fixedSplits := make([]*structs.DiscoverySplit, 0, len(node.Splits))
changed := false
for _, split := range node.Splits {
nextNode := c.nodes[split.NextNode]
if nextNode.Type != structs.DiscoveryGraphNodeTypeSplitter {
fixedSplits = append(fixedSplits, split)
continue
}
changed = true
for _, innerSplit := range nextNode.Splits {
effectiveWeight := split.Weight * innerSplit.Weight / 100
newDiscoverySplit := &structs.DiscoverySplit{
Weight: structs.NormalizeServiceSplitWeight(effectiveWeight),
NextNode: innerSplit.NextNode,
}
fixedSplits = append(fixedSplits, newDiscoverySplit)
}
}
if changed {
node.Splits = fixedSplits
anyChanged = true
}
}
if !anyChanged {
return
}
}
}
// removeUnusedNodes walks the chain from the start and prunes any nodes that
// are no longer referenced. This can happen as a result of operations like
// flattenAdjacentSplitterNodes().
func (c *compiler) removeUnusedNodes() error {
var (
visited = make(map[string]struct{})
todo = make(map[string]struct{})
)
todo[c.startNode] = struct{}{}
getNext := func() string {
if len(todo) == 0 {
return ""
}
for k := range todo {
delete(todo, k)
return k
}
return ""
}
for {
next := getNext()
if next == "" {
break
}
if _, ok := visited[next]; ok {
continue
}
visited[next] = struct{}{}
node := c.nodes[next]
if node == nil {
return fmt.Errorf("compilation references non-retained node %q", next)
}
switch node.Type {
case structs.DiscoveryGraphNodeTypeRouter:
for _, route := range node.Routes {
todo[route.NextNode] = struct{}{}
}
case structs.DiscoveryGraphNodeTypeSplitter:
for _, split := range node.Splits {
todo[split.NextNode] = struct{}{}
}
case structs.DiscoveryGraphNodeTypeResolver:
// nothing special
default:
return fmt.Errorf("unknown node type %q", node.Type)
}
}
if len(visited) == len(c.nodes) {
return nil
}
for name := range c.nodes {
if _, ok := visited[name]; !ok {
delete(c.nodes, name)
}
}
return nil
}
// assembleChain will do the initial assembly of a chain of DiscoveryGraphNode
// entries from the provided config entries.
func (c *compiler) assembleChain() error {
if c.startNode != "" || len(c.nodes) > 0 {
return fmt.Errorf("assembleChain should only be called once")
}
sid := structs.NewServiceID(c.serviceName, c.GetEnterpriseMeta())
// Check for short circuit path.
if len(c.resolvers) == 0 && c.entries.IsChainEmpty() {
// Materialize defaults and cache.
c.resolvers[sid] = newDefaultServiceResolver(sid)
}
// The only router we consult is the one for the service name at the top of
// the chain.
router := c.entries.GetRouter(sid)
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
if router != nil && c.disableAdvancedRoutingFeatures {
router = nil
c.customizedBy.Protocol = true
}
if router == nil {
// If no router is configured, move on down the line to the next hop of
// the chain.
node, err := c.getSplitterOrResolverNode(c.newTarget(c.serviceName, "", "", ""))
if err != nil {
return err
}
c.startNode = node.MapKey()
return nil
}
routerID := structs.NewServiceID(router.Name, router.GetEnterpriseMeta())
routeNode := &structs.DiscoveryGraphNode{
Type: structs.DiscoveryGraphNodeTypeRouter,
Name: serviceIDString(routerID),
Routes: make([]*structs.DiscoveryRoute, 0, len(router.Routes)+1),
}
c.usesAdvancedRoutingFeatures = true
if err := c.recordServiceProtocol(routerID); err != nil {
return err
}
for i := range router.Routes {
// We don't use range variables here because we'll take the address of
// this route and store that in a DiscoveryGraphNode and the range
// variables share memory addresses between iterations which is exactly
// wrong for us here.
route := router.Routes[i]
compiledRoute := &structs.DiscoveryRoute{Definition: &route}
routeNode.Routes = append(routeNode.Routes, compiledRoute)
dest := route.Destination
if dest == nil {
dest = &structs.ServiceRouteDestination{
Service: c.serviceName,
Namespace: router.NamespaceOrDefault(),
}
}
svc := defaultIfEmpty(dest.Service, c.serviceName)
destNamespace := defaultIfEmpty(dest.Namespace, router.NamespaceOrDefault())
// Check to see if the destination is eligible for splitting.
var (
node *structs.DiscoveryGraphNode
err error
)
if dest.ServiceSubset == "" {
node, err = c.getSplitterOrResolverNode(
c.newTarget(svc, "", destNamespace, ""),
)
} else {
node, err = c.getResolverNode(
c.newTarget(svc, dest.ServiceSubset, destNamespace, ""),
false,
)
}
if err != nil {
return err
}
compiledRoute.NextNode = node.MapKey()
}
// If we have a router, we'll add a catch-all route at the end to send
// unmatched traffic to the next hop in the chain.
defaultDestinationNode, err := c.getSplitterOrResolverNode(c.newTarget(router.Name, "", router.NamespaceOrDefault(), ""))
if err != nil {
return err
}
defaultRoute := &structs.DiscoveryRoute{
Definition: newDefaultServiceRoute(router.Name, router.NamespaceOrDefault()),
NextNode: defaultDestinationNode.MapKey(),
}
routeNode.Routes = append(routeNode.Routes, defaultRoute)
c.startNode = routeNode.MapKey()
c.recordNode(routeNode)
return nil
}
func newDefaultServiceRoute(serviceName string, namespace string) *structs.ServiceRoute {
return &structs.ServiceRoute{
Match: &structs.ServiceRouteMatch{
HTTP: &structs.ServiceRouteHTTPMatch{
PathPrefix: "/",
},
},
Destination: &structs.ServiceRouteDestination{
Service: serviceName,
Namespace: namespace,
},
}
}
func (c *compiler) newTarget(service, serviceSubset, namespace, datacenter string) *structs.DiscoveryTarget {
if service == "" {
panic("newTarget called with empty service which makes no sense")
}
t := structs.NewDiscoveryTarget(
service,
serviceSubset,
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
defaultIfEmpty(namespace, c.evaluateInNamespace),
defaultIfEmpty(datacenter, c.evaluateInDatacenter),
)
// Set default connect SNI. This will be overridden later if the service
// has an explicit SNI value configured in service-defaults.
t.SNI = connect.TargetSNI(t, c.evaluateInTrustDomain)
// Use the same representation for the name. This will NOT be overridden
// later.
t.Name = t.SNI
prev, ok := c.loadedTargets[t.ID]
if ok {
return prev
}
c.loadedTargets[t.ID] = t
return t
}
func (c *compiler) rewriteTarget(t *structs.DiscoveryTarget, service, serviceSubset, namespace, datacenter string) *structs.DiscoveryTarget {
var (
service2 = t.Service
serviceSubset2 = t.ServiceSubset
namespace2 = t.Namespace
datacenter2 = t.Datacenter
)
if service != "" && service != service2 {
service2 = service
// Reset the chosen subset if we reference a service other than our own.
serviceSubset2 = ""
}
if serviceSubset != "" {
serviceSubset2 = serviceSubset
}
if namespace != "" {
namespace2 = namespace
}
if datacenter != "" {
datacenter2 = datacenter
}
return c.newTarget(service2, serviceSubset2, namespace2, datacenter2)
}
func (c *compiler) getSplitterOrResolverNode(target *structs.DiscoveryTarget) (*structs.DiscoveryGraphNode, error) {
nextNode, err := c.getSplitterNode(target.ServiceID())
if err != nil {
return nil, err
} else if nextNode != nil {
return nextNode, nil
}
return c.getResolverNode(target, false)
}
func (c *compiler) getSplitterNode(sid structs.ServiceID) (*structs.DiscoveryGraphNode, error) {
name := serviceIDString(sid)
// Do we already have the node?
if prev, ok := c.splitterNodes[name]; ok {
return prev, nil
}
// Fetch the config entry.
splitter := c.entries.GetSplitter(sid)
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
if splitter != nil && c.disableAdvancedRoutingFeatures {
splitter = nil
c.customizedBy.Protocol = true
}
if splitter == nil {
return nil, nil
}
// Build node.
splitNode := &structs.DiscoveryGraphNode{
Type: structs.DiscoveryGraphNodeTypeSplitter,
Name: name,
Splits: make([]*structs.DiscoverySplit, 0, len(splitter.Splits)),
}
// If we record this exists before recursing down it will short-circuit
// sanely if there is some sort of graph loop below.
c.recordNode(splitNode)
var hasLB bool
for _, split := range splitter.Splits {
compiledSplit := &structs.DiscoverySplit{
Weight: split.Weight,
}
splitNode.Splits = append(splitNode.Splits, compiledSplit)
svc := defaultIfEmpty(split.Service, sid.ID)
splitID := structs.ServiceID{
ID: svc,
EnterpriseMeta: *split.GetEnterpriseMeta(&sid.EnterpriseMeta),
}
// Check to see if the split is eligible for additional splitting.
if !splitID.Matches(&sid) && split.ServiceSubset == "" {
nextNode, err := c.getSplitterNode(splitID)
if err != nil {
return nil, err
} else if nextNode != nil {
compiledSplit.NextNode = nextNode.MapKey()
continue
}
// fall through to group-resolver
}
node, err := c.getResolverNode(
c.newTarget(splitID.ID, split.ServiceSubset, splitID.NamespaceOrDefault(), ""),
false,
)
if err != nil {
return nil, err
}
compiledSplit.NextNode = node.MapKey()
// There exists the possibility that a splitter may split between two distinct service names
// with distinct hash-based load balancer configs specified in their service resolvers.
// We cannot apply multiple hash policies to a splitter node's route action.
// Therefore, we attach the first hash-based load balancer config we encounter.
if !hasLB {
2020-09-11 15:21:43 +00:00
if lb := node.LoadBalancer; lb != nil && lb.IsHashBased() {
splitNode.LoadBalancer = node.LoadBalancer
hasLB = true
}
}
}
c.usesAdvancedRoutingFeatures = true
return splitNode, nil
}
// getResolverNode handles most of the code to handle redirection/rewriting
// capabilities from a resolver config entry. It recurses into itself to
// _generate_ targets used for failover out of convenience.
func (c *compiler) getResolverNode(target *structs.DiscoveryTarget, recursedForFailover bool) (*structs.DiscoveryGraphNode, error) {
var (
// State to help detect redirect cycles and print helpful error
// messages.
redirectHistory = make(map[string]struct{})
redirectOrder []string
)
RESOLVE_AGAIN:
// Do we already have the node?
if prev, ok := c.resolveNodes[target.ID]; ok {
return prev, nil
}
targetID := target.ServiceID()
if err := c.recordServiceProtocol(targetID); err != nil {
return nil, err
}
// Fetch the config entry.
resolver, ok := c.resolvers[targetID]
if !ok {
// Materialize defaults and cache.
resolver = newDefaultServiceResolver(targetID)
c.resolvers[targetID] = resolver
}
if _, ok := redirectHistory[target.ID]; ok {
redirectOrder = append(redirectOrder, target.ID)
return nil, &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"detected circular resolver redirect: [%s]",
strings.Join(redirectOrder, " -> "),
),
}
}
redirectHistory[target.ID] = struct{}{}
redirectOrder = append(redirectOrder, target.ID)
// Handle redirects right up front.
//
// TODO(rb): What about a redirected subset reference? (web/v2, but web redirects to alt/"")
if resolver.Redirect != nil {
redirect := resolver.Redirect
redirectedTarget := c.rewriteTarget(
target,
redirect.Service,
redirect.ServiceSubset,
redirect.Namespace,
redirect.Datacenter,
)
if redirectedTarget.ID != target.ID {
target = redirectedTarget
goto RESOLVE_AGAIN
}
}
// Handle default subset.
if target.ServiceSubset == "" && resolver.DefaultSubset != "" {
target = c.rewriteTarget(
target,
"",
resolver.DefaultSubset,
"",
"",
)
goto RESOLVE_AGAIN
}
if target.ServiceSubset != "" && !resolver.SubsetExists(target.ServiceSubset) {
return nil, &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"service %q does not have a subset named %q",
target.Service,
target.ServiceSubset,
),
}
}
connectTimeout := resolver.ConnectTimeout
if connectTimeout < 1 {
connectTimeout = 5 * time.Second
}
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
if c.overrideConnectTimeout > 0 {
if connectTimeout != c.overrideConnectTimeout {
connectTimeout = c.overrideConnectTimeout
c.customizedBy.ConnectTimeout = true
}
}
// Build node.
node := &structs.DiscoveryGraphNode{
Type: structs.DiscoveryGraphNodeTypeResolver,
Name: target.ID,
Resolver: &structs.DiscoveryResolver{
Default: resolver.IsDefault(),
Target: target.ID,
ConnectTimeout: connectTimeout,
},
LoadBalancer: resolver.LoadBalancer,
}
target.Subset = resolver.Subsets[target.ServiceSubset]
if serviceDefault := c.entries.GetService(targetID); serviceDefault != nil && serviceDefault.ExternalSNI != "" {
// Override the default SNI value.
target.SNI = serviceDefault.ExternalSNI
target.External = true
}
// If using external SNI the service is fundamentally external.
if target.External {
if resolver.Redirect != nil {
return nil, &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"service %q has an external SNI set; cannot define redirects for external services",
target.Service,
),
}
}
if len(resolver.Subsets) > 0 {
return nil, &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"service %q has an external SNI set; cannot define subsets for external services",
target.Service,
),
}
}
if len(resolver.Failover) > 0 {
return nil, &structs.ConfigEntryGraphError{
Message: fmt.Sprintf(
"service %q has an external SNI set; cannot define failover for external services",
target.Service,
),
}
}
}
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
// TODO (mesh-gateway)- maybe allow using a gateway within a datacenter at some point
if target.Datacenter == c.useInDatacenter {
target.MeshGateway.Mode = structs.MeshGatewayModeDefault
} else if target.External {
// Bypass mesh gateways if it is an external service.
target.MeshGateway.Mode = structs.MeshGatewayModeDefault
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
} else {
// Default mesh gateway settings
if serviceDefault := c.entries.GetService(targetID); serviceDefault != nil {
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
target.MeshGateway = serviceDefault.MeshGateway
}
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 c.entries.GlobalProxy != nil && target.MeshGateway.Mode == structs.MeshGatewayModeDefault {
target.MeshGateway.Mode = c.entries.GlobalProxy.MeshGateway.Mode
}
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 c.overrideMeshGateway.Mode != structs.MeshGatewayModeDefault {
if target.MeshGateway.Mode != c.overrideMeshGateway.Mode {
target.MeshGateway.Mode = c.overrideMeshGateway.Mode
c.customizedBy.MeshGateway = true
}
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
}
}
// Retain this target in the final results.
c.retainedTargets[target.ID] = struct{}{}
if recursedForFailover {
// If we recursed here from ourselves in a failover context, just emit
// this node without caching it or even processing failover again.
// This is a little weird but it keeps the redirect/default-subset
// logic in one place.
return node, nil
}
// If we record this exists before recursing down it will short-circuit
// sanely if there is some sort of graph loop below.
c.recordNode(node)
if len(resolver.Failover) > 0 {
f := resolver.Failover
// Determine which failover section applies.
failover, ok := f[target.ServiceSubset]
if !ok {
failover, ok = f["*"]
}
if ok {
// Determine which failover definitions apply.
var failoverTargets []*structs.DiscoveryTarget
if len(failover.Datacenters) > 0 {
for _, dc := range failover.Datacenters {
// Rewrite the target as per the failover policy.
failoverTarget := c.rewriteTarget(
target,
failover.Service,
failover.ServiceSubset,
failover.Namespace,
dc,
)
if failoverTarget.ID != target.ID { // don't failover to yourself
failoverTargets = append(failoverTargets, failoverTarget)
}
}
} else {
// Rewrite the target as per the failover policy.
failoverTarget := c.rewriteTarget(
target,
failover.Service,
failover.ServiceSubset,
failover.Namespace,
"",
)
if failoverTarget.ID != target.ID { // don't failover to yourself
failoverTargets = append(failoverTargets, failoverTarget)
}
}
// If we filtered everything out then no point in having a failover.
if len(failoverTargets) > 0 {
df := &structs.DiscoveryFailover{}
node.Resolver.Failover = df
// Take care of doing any redirects or configuration loading
// related to targets by cheating a bit and recursing into
// ourselves.
for _, target := range failoverTargets {
failoverResolveNode, err := c.getResolverNode(target, true)
if err != nil {
return nil, err
}
failoverTarget := failoverResolveNode.Resolver.Target
df.Targets = append(df.Targets, failoverTarget)
}
}
}
}
return node, nil
}
func newDefaultServiceResolver(sid structs.ServiceID) *structs.ServiceResolverConfigEntry {
return &structs.ServiceResolverConfigEntry{
Kind: structs.ServiceResolver,
Name: sid.ID,
EnterpriseMeta: sid.EnterpriseMeta,
}
}
func defaultIfEmpty(val, defaultVal string) string {
if val != "" {
return val
}
return defaultVal
}
func enableAdvancedRoutingForProtocol(protocol string) bool {
return structs.IsProtocolHTTPLike(protocol)
}