Traffic Shifting/Splitting

Envoy’s router can split traffic to a route in a virtual host across two or more upstream clusters. There are two common use cases.

1. Version upgrades: traffic to a route is shifted gradually from one cluster to another. The traffic shifting section describes this scenario in more detail.

2. A/B testing or multivariate testing: two or more versions of the same service are tested simultaneously. The traffic to the route has to be split between clusters running different versions of the same service. The traffic splitting section describes this scenario in more detail.

Traffic shifting between two upstreams

The runtime object in the route configuration determines the probability of selecting a particular route (and hence its cluster). By using the runtime_fraction configuration, traffic to a particular route in a virtual host can be gradually shifted from one cluster to another. Consider the following example configuration, where two versions helloworld_v1 and helloworld_v2 of a service named helloworld are declared in the envoy configuration file.

  1. virtual_hosts:
  2.    - name: www2
  3.      domains:
  4.      - '*'
  5.      routes:
  6.        - match:
  7.            prefix: /
  8.            runtime_fraction:
  9.              default_value:
  10.                numerator: 50
  11.                denominator: HUNDRED
  12.              runtime_key: routing.traffic_shift.helloworld
  13.          route:
  14.            cluster: helloworld_v1
  15.        - match:
  16.            prefix: /
  17.          route:
  18.            cluster: helloworld_v2

Envoy matches routes with a first match policy. If the route has a runtime_fraction object, the request will be additionally matched based on the runtime_fraction value (or the default, if no value is specified). Thus, by placing routes back-to-back in the above example and specifying a runtime_fraction object in the first route, traffic shifting can be accomplished by changing the runtime_fraction value. The following are the approximate sequence of actions required to accomplish the task.

  1. In the beginning, set routing.traffic_shift.helloworld to 100, so that all requests to the helloworld virtual host would match with the v1 route and be served by the helloworld_v1 cluster.

  2. To start shifting traffic to helloworld_v2 cluster, set routing.traffic_shift.helloworld to values 0 < x < 100. For instance at 90, 1 out of every 10 requests to the helloworld virtual host will not match the v1 route and will fall through to the v2 route.

  3. Gradually decrease the value set in routing.traffic_shift.helloworld so that a larger percentage of requests match the v2 route.

  4. When routing.traffic_shift.helloworld is set to 0, no requests to the helloworld virtual host will match to the v1 route. All traffic would now fall through to the v2 route and be served by the helloworld_v2 cluster.

Traffic splitting across multiple upstreams

Consider the helloworld example again, now with three versions (v1, v2 and v3) instead of two. To split traffic evenly across the three versions (i.e., 33%, 33%, 34%), the weighted_clusters option can be used to specify the weight for each upstream cluster.

Unlike the previous example, a single route entry is sufficient. The weighted_clusters configuration block in a route can be used to specify multiple upstream clusters along with weights that indicate the percentage of traffic to be sent to each upstream cluster.

  1. virtual_hosts:
  2.    - name: www2
  3.      domains:
  4.      - '*'
  5.      routes:
  6.        - match: { prefix: / }
  7.          route:
  8.            weighted_clusters:
  9.              runtime_key_prefix: routing.traffic_split.helloworld
  10.              clusters:
  11.                - name: helloworld_v1
  12.                  weight: 33
  13.                - name: helloworld_v2
  14.                  weight: 33
  15.                - name: helloworld_v3
  16.                  weight: 34

By default, the weights must sum to exactly 100. In the V2 API, the total weight defaults to 100, but can be modified to allow finer granularity.

The weights assigned to each cluster can be dynamically adjusted using the following runtime variables: routing.traffic_split.helloworld.helloworld_v1, routing.traffic_split.helloworld.helloworld_v2 and routing.traffic_split.helloworld.helloworld_v3.