Push problems in GCP-deployed V23 application

We’re periodically seeing significant numbers of push failures that seem to result in node tree corruption, and I’m looking for suggestions how to make the application better behaved.
The application is a Vaadin (flow) 23.3 SpringBoot app deployed in GKE. We deploy with completely standard embedded tomcat and with TLS termination provided by a GCP load balancer. There’s nothing special about our network environment in google or at the client side.

We see lots of these:
2023-10-19T07:49:28,477 WARN [http-nio-8080-exec-12] (ServerRpcHandler.java:324) - Resynchronizing UI by client's request. A network message was lost before reaching the client and the client is reloading the full UI state. This typically happens because of a bad network connection with packet loss or because of some part of the network infrastructure (load balancer, proxy) terminating a push (websocket or long-polling) connection. If you are using push with a proxy, make sure the push timeout is set to be smaller than the proxy connection timeout

The symptoms look like https://github.com/vaadin/flow/issues/14232 , which was marked as closed earlier this year, certainly for v23.3.25 so I’m not sure if the issue is different or if this suggests a regression.

Terrifyingly, enabling logging.level.com.vaadin.flow.component.internal = trace seems to have resulted in the problem going away. I hate Heisenbugs