Description
Description
The HttpServerTransport will by default wait indefinitely for all client connections to close to avoid spurious errors, and this wait happens fairly early on in the shutdown of the node. In particular, there's no signal to any plugins or other widely-available components to indicate that the node has started to shut down. That means that long-polling actions such as GET /{index}/_fleet/global_checkpoints may delay the shutdown of the node for many minutes (or more). Yet, long-polling actions may safely return early (including Connection: close in the response, see #127984) without this being considered a spurious error.
Could we have a mechanism to let actions like these know that the shutdown has started and therefore to curtail any waiting?