Document new behaviour of session.on_disconnect

[Gerold103]

Now session.on_disconnect triggers are called right after disconnect. Not after a last request is finished.

[pgulutzan]

In what version did this occur?

If it's a significant behaviour change, we have to say "starting in version x this incompatible change was made ...". But I wonder whether the version-10 documentation is already correct by accident, because it says
"Define a trigger for execution after a client has disconnected. If the trigger function causes an error, the error is logged but otherwise is ignored. The trigger is invoked while the session associated with the client still exists and can access session properties, such as box.session.id()."
So, what in the current documentation is now incorrect?

[Gerold103]

The new behaviour appears since 1.10. As I know, beforehand on_disconnect triggers were called only after a last request has finished. Now they are called right after disconnect. If the documentation does not specify the time of triggers invocation, then maybe it should start specifying.

[Gerold103]

Also, it is worth documenting that not all session properties are valid after disconnect. box.session.fd() will return -1, box.session.sync() will return 0, and box.session.push will return nil + error after disconnect.

[pgulutzan]

The manual says that the on_disconnect trigger function is called after disconnect. As far as I can tell, it does not wait for the last request to finish, but that is true in both version 1.9 and version 1.10. It does wait for other requests to yield,but that seems obvious -- otherwise how can the trigger function be called?

Here is an example based on how I understand what you are saying.
On shell #1, as admin:
tarantool
box.cfg{listen=3302}
function on_disconnect_impl() print('disconnect ' .. box.session.id()); os.exit(); end
box.session.on_disconnect(on_disconnect_impl)
box.schema.user.grant('guest','read,write,execute','universe')
fiber = require('fiber')
fiber.sleep(1000)
On shell #2, as guest:
tarantoolctl connect 3302
^C i.e. control-C
Result: shell #1 displays and stops, regardless of Tarantool version.

[Gerold103]

box.cfg{listen = 3313}
netbox = require('net.box')
fiber = require('fiber')
box.schema.user.grant('guest', 'read,write,execute', 'universe')
finish_long = false
disconnected = false
function long()
	box.session.on_disconnect(function() disconnected = true end)
	while not finish_long do fiber.sleep(0.01) end
end
c = netbox.connect(box.cfg.listen)
c:call('long', {}, {timeout = 0.01})
c:close()
disconnected -- (1), false in 1.9, true in 1.10
finish_long = true
while not disconnected do fiber.sleep(0.01) end
disconnected -- (2), true in both

In 1.9 label (1) will print false, (2) - true. In 1.10 they both print true.
Under finishing a last request I mean not a last request of Tarantool, but of a connection.

[pgulutzan]

Okay, I can add, following what I quoted previously: "Since version 1.10, the trigger function is invoked immediately after the disconnect, even if requests that were made during the session have not finished." It doesn't seem significant enough to deserve a warning about changed behaviour, since in the version 1.9 manual we don't say something different. And in section https://tarantool.io/en/doc/1.10/book/box/box_session/#box-session-id I will change "The result can be 0 meaning there is no session." because the result might be -1; plus in the overview I can add that some items become invalid after disconnect.

[Gerold103]

Box.session.id() does not change in case of disconnect. Only sync and fd. Also push() returns an error.

[pgulutzan]

Okay, I made the additions that I described, except that instead of putting something in the overview I added notes in id() and push() (we don't document fd). The mention of -1 is not about what one sees from the trigger function after disconnect, it merely is stated that -1 is a possible value.