A question about Incremental xDS in high concurrency scenarios

[bearslyricattack]

Hello, I would like to know how Envoy handles concurrent scenarios when using the incremental XDS of Envoy. When the control plane sends a request and Envoy has not yet completed processing and responded with an ACK, the control plane issues a new configuration update? Is it maintaining an internal queue for sequential execution? How much does its performance improve compared to full updates when the control surface is frequently updated?

[zuercher]

The control plane messages (which are a stream of responses to Envoy's stream of requests) are processed sequentially. There's no explicit queue within Envoy -- the response stream will accumulate in the grpc client, which will exert back pressure to the control plane in the event that multiple incremental updates are issued quickly enough.

I can't speak to the performance of the update within Envoy. I know that configurations with a large number of listeners, clusters and/or clusters with high cardinality can take significant time just to transmit from the control plane to Envoy. In my experience the network delay dominated the latency between the control plane initiating an update and completing it within Envoy.