Only set disconnected time when it is not already set.

[agronick]

This is fix 1 of 2 for django/channels#1181.

[andrewgodwin]

So looking at this change in isolation, I'm curious what effect it would have - did you write it anticipating solving a race condition, or something else? I'd like to see a comment with it because otherwise it looks like you're avoiding replacing a number with another similar-size number which would not really affect memory.

[agronick]

I wrote the details in the bug thread. Avoiding writing a bigger number is what we are doing because if it does the time difference will never be large enough to be cleaned up.

• check_timeouts() runs every two seconds
• It calls finish() which calls self.server.protocol_disconnected(self)
• t keeps overwriting the time of the disconnected key
• The disconnected time difference is never high enough to get cleaned up:

            if (
                disconnected
                and time.time() - disconnected > self.application_close_timeout
            ):

• check_timeouts() continues to be called every two seconds.
• It keeps calling send_disconnect() which causes the memory leak with this code:

            self.application_queue.put_nowait({"type": "http.disconnect"})

• This makes a giant queue full of useless data.

To experience this just make a view that has time.sleep(999999) with a low http_timeout.

[agronick]

I added another fix that handles an exception where send is called on an application instance that has been cleaned up. I see that exception a few times a day. I added comments for both changes.

[andrewgodwin]

Great, thanks for adding the comment!

[agronick]

Any idea when this is going out? I had this bug hit last night and this is what it did to my memory usage:

[andrewgodwin]

I'll try to get a release out this week - I forgot to roll one along with Channels last week.

[andrewgodwin]

It's now out in 2.2.4.

[agronick]

Great thanks.

…

On Sat, Dec 15, 2018, 4:36 PM Andrew Godwin ***@***.*** wrote: It's now out in 2.2.4. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#237 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AB8pv1pGPMawQQF3gM6o0cQMeLuPuiohks5u5WtFgaJpZM4Y1_JP> .