The problem is twofold:

1. We have arbitrary timeouts when executing requests, and those timeouts start when we submit the request to the client rather than when the request starts executing. Before HSEARCH-2764 In Progress , this timeout is set up by RestClient.performRequest (the synchronous form), after HSEARCH-2764 In Progress we set it up ourselves in DefaultElasticsearchClient.
2. Retry timeouts (managed by RestClient) also start when we submit requests rather than when the request starts executing. See https://github.com/elastic/elasticsearch/issues/25578