[grpc-web] Looking for Envoy example of load balancing

[Rob Cecil]

Looking for an example of an Envoy configuration that implements session affinity (stickiness) to load balance a cluster of backend servers.  Thanks!

I'm open to using the source IP or something in the header, but probably not cookie.

Thanks

[Rob Cecil]

Just some follow up - in case anyone else is looking for solutions.

There are no obvious usage scenarios documented, but I found a solution using a lb_policy of "ring_hash" where the client provides a header that is unique the client. Just generate a unique, stable identifier throughout the lifetime of the client (browser) and inject that as a header (here using "x-session-hash", but it could be anything) into every request.

I'm using nanoid to generate unique strings in Javascript. I simply generate one and store in local storage.

Here's the relevant .yaml for an example configuration that defines a two-host upstream cluster:

admin:

  access_log_path: /tmp/admin_access.log

  address:

    socket_address: { address: 0.0.0.0, port_value: 9901 }

static_resources:

  listeners:

  - name: listener_0

    address:

      socket_address: { address: 0.0.0.0, port_value: 8080 }

    filter_chains:

    - filters:

      - name: envoy.filters.network.http_connection_manager

        typed_config:

          "@type": type.googleapis.com/envoy.extensions.filters.network.http_connection_manager.v3.HttpConnectionManager

          codec_type: auto

          stat_prefix: ingress_http

          route_config:

            name: local_route

            virtual_hosts:

            - name: local_service

              domains: ["*"]

              routes:

              - match: { prefix: "/" }

                route:

                  cluster: controlweb_backendservice

                  hash_policy:

                    - header:

                        header_name: x-session-hash

                  max_stream_duration:

                    grpc_timeout_header_max: 0s

              cors:

                allow_origin_string_match:

                - prefix: "*"

                allow_methods: GET, PUT, DELETE, POST, OPTIONS

                allow_headers: keep-alive,user-agent,cache-control,content-type,content-transfer-encoding,access-token,x-accept-content-transfer-encoding,x-accept-response-streaming,x-user-agent,x-grpc-web,grpc-timeout,x-session-hash

                max_age: "1728000"

                expose_headers: access-token,grpc-status,grpc-message

          http_filters:

          - name: envoy.filters.http.grpc_web

          - name: envoy.filters.http.cors

          - name: envoy.filters.http.router

  clusters:

  - name: controlweb_backendservice

    connect_timeout: 0.25s

    type: strict_dns

    http2_protocol_options: {}

    lb_policy: ring_hash

    

    load_assignment:

      cluster_name: cluster_0

      endpoints:

        - lb_endpoints:

            - endpoint:

                address:

                  socket_address:

                    address: 172.16.0.219

                    port_value: 50251

              load_balancing_weight: 10

            - endpoint:

                address:

                  socket_address:

                    address: 172.16.0.132

                    port_value: 50251

              load_balancing_weight: 1

[Rob Cecil]

I am aware of maglev as a potential alternative lb_policy, and also there are problems if the cluster configuration is changed while there are outstand connections through envoy.

[Rob Cecil]

This was , in large part how I found the approach I am using:

https://github.com/envoyproxy/envoy/issues/4897