data-plane-api/XDS_PROTOCOL.md

xDS REST and gRPC protocol

Envoy discovers its various dynamic resources via the filesystem or by querying one or more management servers. Collectively, these discovery services and their corresponding APIs are referred to as xDS. Resources are requested via subscriptions, by specifying a filesystem path to watch, initiating gRPC streams or polling a REST-JSON URL. The latter two methods involve sending requests with a DiscoveryRequest proto payload. Resources are delivered in a DiscoveryResponse proto payload in all methods. We discuss each type of subscription below.

Filesystem subscriptions

The simplest approach to delivering dynamic configuration is to place it at a well known path specified in the ConfigSource. Envoy will use inotify (kqueue on Mac OS X) to monitor the file for changes and parse the DiscoveryResponse proto in the file on update. Binary protobufs, JSON, YAML and proto text are supported formats for the DiscoveryResponse.

There is no mechanism available for filesystem subscriptions to ACK/NACK updates beyond stats counters and logs. The last valid configuration for an xDS API will continue to apply if an configuration update rejection occurs.

Streaming gRPC subscriptions

Singleton resource type discovery

A gRPC ApiConfigSource can be specified independently for each xDS API, pointing at an upstream cluster corresponding to a management server. This will initiate an independent bidirectional gRPC stream for each xDS resource type, potentially to distinct management servers. API delivery is eventually consistent. See ADS below for situations in which explicit control of sequencing is required.

Type URLs

Each xDS API is concerned with resources of a given type. There is a 1:1 correspondence between an xDS API and a resource type. That is:

• LDS: envoy.api.v2.Listener
• RDS: envoy.api.v2.RouteConfiguration
• CDS: envoy.api.v2.Cluster
• EDS: envoy.api.v2.ClusterLoadAssignment

The concept of type URLs appears below, and takes the form type.googleapis.com/<resource type>, e.g. type.googleapis.com/envoy.api.v2.Cluster for CDS. In various requests from Envoy and responses by the management server, the resource type URL is stated.

ACK/NACK and versioning

Each stream begins with a DiscoveryRequest from Envoy, specifying the list of resources to subscribe to, the type URL corresponding to the subscribed resources, the node identifier and an empty version_info. An example EDS request might be:

version_info:
node: { id: envoy }
resource_names:
- foo
- bar
type_url: type.googleapis.com/envoy.api.v2.ClusterLoadAssignment
response_nonce:

The management server may reply either immediately or when the requested resources are available with a DiscoveryResponse, e.g.:

version_info: X
resources:
- foo ClusterLoadAssignment proto encoding
- bar ClusterLoadAssignment proto encoding
type_url: type.googleapis.com/envoy.api.v2.ClusterLoadAssignment
nonce: A

After processing the DiscoveryResponse, Envoy will send a new request on the stream, specifying the last version successfully applied and the nonce provided by the management server. If the update was successfully applied, the version_info will be X, as indicated in the sequence diagram:

In this sequence diagram, and below, the following format is used to abbreviate messages:

• DiscoveryRequest: (V=version_info,R=resource_names,N=response_nonce,T=type_url)
• DiscoveryResponse: (V=version_info,R=resources,N=nonce,T=type_url)

The version provides Envoy and the management server a shared notion of the currently applied configuration, as well as a mechanism to ACK/NACK configuration updates. If Envoy had instead rejected configuration update X, it would reply with its previous version, which in this case was the empty initial version:

Later, an API update may succeed at a new version Y:

Each stream has its own notion of versioning, there is no shared versioning across resource types. When ADS is not used, even each resource of a given resource type may have a distinct version, since the Envoy API allows distinct EDS/RDS resources to point at different ConfigSources.

When to send an update

The management server should only send updates to the Envoy client when the resources in the DiscoveryResponse have changed. Envoy replies to any DiscoveryResponse with a DiscoveryRequest containing the ACK/NACK immediately after it has been either accepted or rejected. If the management server provides the same set of resources rather than waiting for a change to occur, it will cause Envoy and the management server to spin and have a severe performance impact.

Within a stream, new DiscoveryRequests supersede any prior DiscoveryRequests having the same resource type. This means that the management server only needs to respond to the latest DiscoveryRequest on each stream for any given resource type.

Resource hints

The resource_names specified in the DiscoveryRequest are a hint. Some resource types, e.g. Clusters and Listeners will specify an empty resource_names list, since Envoy is interested in learning about all the Clusters (CDS) and Listeners (LDS) that the management server(s) know about corresponding to its node identification. Other resource types, e.g. RouteConfigurations (RDS) and ClusterLoadAssignments (EDS), follow from earlier CDS/LDS updates and Envoy is able to explicitly enumerate these resources.

LDS/CDS resource hints will always be empty and it is expected that the management server will provide the complete state of the LDS/CDS resources in each response. An absent Listener or Cluster will be deleted.

For EDS/RDS, the management server does not need to supply every requested resource and may also supply additional, unrequested resources, resource_names is only a hint. Envoy will silently ignore any superfluous resources. When a requested resource is missing in a RDS or EDS update, Envoy will retain the last known value for this resource. The management server may be able to infer all the required EDS/RDS resources from the node identification in the DiscoveryRequest, in which case this hint may be discarded. An empty EDS/RDS DiscoveryResponse is effectively a nop from the perspective of the respective resources in the Envoy.

When a Listener or Cluster is deleted, its corresponding EDS and RDS resources are also deleted inside the Envoy instance. In order for EDS resources to be known or tracked by Envoy, there must exist an applied Cluster definition (e.g. sourced via CDS). A similar relationship exists between RDS and Listeners (e.g. sourced via LDS).

For EDS/RDS, Envoy may either generate a distinct stream for each resource of a given type (e.g. if each ConfigSource has its own distinct upstream cluster for a management server), or may combine together multiple resource requests for a given resource type when they are destined for the same management server. This is left to implementation specifics, management servers should be capable of handling one or more resource_names for a given resource type in each request. Both sequence diagrams below are valid for fetching two EDS resources {foo, bar}:

Resource updates

As discussed above, Envoy may update the list of resource_names it presents to the management server in each DiscoveryRequest that ACK/NACKs a specific DiscoveryResponse. In addition, Envoy may later issue additional DiscoveryRequests at a given version_info to update the management server with new resource hints. For example, if Envoy is at EDS version X and knows only about cluster foo, but then receives a CDS update and learns about bar in addition, it may issue an additional DiscoveryRequest for X with {foo,bar} as resource_names.

There is a race condition that may arise here; if after a resource hint update is issued by Envoy at X, but before the management server processes the update it replies with a new version Y, the resource hint update may be interpreted as a rejection of Y by presenting an X version_info. To avoid this, the management server provides a nonce that Envoy uses to indicate the specific DiscoveryResponse each DiscoveryRequest corresponds to:

The management server should not send a DiscoveryResponse for any DiscoveryRequest that has a stale nonce. A nonce becomes stale following a newer nonce being presented to Envoy in a DiscoveryResponse. A management server does not need to send an update until it determines a new version is available. Earlier requests at a version then also become stale. It may process multiple DiscoveryRequests at a version until a new version is ready.

An implication of the above resource update sequencing is that Envoy does not expect a DiscoveryResponse for every DiscoveryRequest it issues.

Eventual consistency considerations

Since Envoy's xDS APIs are eventually consistent, traffic may drop briefly during updates. For example, if only cluster X is known via CDS/EDS, a RouteConfiguration references cluster X and is then adjusted to cluster Y just before the CDS/EDS update providing Y, traffic will be blackholed until Y is known about by the Envoy instance.

For some applications, a temporary drop of traffic is acceptable, retries at the client or by other Envoy sidecars will hide this drop. For other scenarios where drop can't be tolerated, traffic drop could have been avoided by providing a CDS/EDS update with both X and Y, then the RDS update repointing from X to Y and then a CDS/EDS update dropping X.

In general, to avoid traffic drop, sequencing of updates should follow a make before break model, wherein

• CDS updates (if any) must always be pushed first.
• EDS updates (if any) must arrive after CDS updates for the respective clusters.
• LDS updates must arrive after corresponding CDS/EDS updates.
• RDS updates related to the newly added listeners must arrive in the end.
• Stale CDS clusters and related EDS endpoints (ones no longer being referenced) can then be removed.

xDS updates can be pushed independently if no new clusters/routes/listeners are added or if it's acceptable to temporarily drop traffic during updates. Note that in case of LDS updates, the listeners will be warmed before they receive traffic, i.e. the dependent routes are fetched through RDS if configured. On the other hand, clusters are not warmed when adding/removing/updating clusters. Similarly, routes are not warmed -- i.e., the management plane must ensure that clusters referenced by a route are in place, before pushing the updates for a rotue.

Aggregated Discovery Services (ADS)

It's challenging to provide the above guarantees on sequencing to avoid traffic drop when management servers are distributed. ADS allow a single management server, via a single gRPC stream, to deliver all API updates. This provides the ability to carefully sequence updates to avoid traffic drop. With ADS, a single stream is used with multiple independent DiscoveryRequest/DiscoveryResponse sequences multiplexed via the type URL. For any given type URL, the above sequencing of DiscoveryRequest and DiscoveryResponse messages applies. An example update sequence might look like:

A single ADS stream is available per Envoy instance.

An example minimal bootstrap.yaml fragment for ADS configuration is:

node:
  id: <node identifier>
dynamic_resources:
  cds_config: {ads: {}}
  lds_config: {ads: {}}
  ads_config:
    api_type: GRPC
    cluster_name: [ads_cluster]
static_resources:
  clusters:
  - name: ads_cluster
    connect_timeout: { seconds: 5 }
    type: STATIC
    hosts:
    - socket_address:
        address: <ADS management server IP address>
        port_value: <ADS management server port>
    lb_policy: ROUND_ROBIN
    http2_protocol_options: {}
admin:
  ...

REST-JSON polling subscriptions

Synchronous (long) polling via REST endpoints is also available for the xDS singleton APIs. The above sequencing of messages is similar, except no persistent stream is maintained to the management server. It is expected that there is only a single outstanding request at any point in time, and as a result the response nonce is optional in REST-JSON. The JSON canonical transform of proto3 is used to encode DiscoveryRequest and DiscoveryResponse messages. ADS is not available for REST-JSON polling.

When the poll period is set to a small value, with the intention of long polling, then there is also a requirement to avoid sending a DiscoveryResponse unless a change to the underlying resources has occurred.