Queuing failed badly on bad connectivity

[ara4n]

Steps to reproduce

1. Was on the tube; zero connectivity
2. Tried to send a message
3. It immediately hard-failed (red circle), rather than queuing.
4. I automatically tapped the red-circle to try to force a retry (or to get it actually queue, rather than hard-fail)... except there's no retry button; only a remove button :/
5. Later, going back into the room when on good connectivity, the message is still there... except now it's in queued state (hollow-circle). Except despite now being on good connectivity, it doesn't send.
6. Later still, the message has now vanished entirely from the timeline, and looks to have never been sent at all.

Screenshot from step 4, when it started to go wrong:

Outcome

What did you expect?

• On bad connectivity, msgs should be soft-failed and queued, not hard-failed. Hard-fail should only happen if the server sends back an explicit permanent error.
• When connectivity returns (as detected by connectivity checks and/or observing traffic) the retry schedule should be reset and restarted, to try to send the msg asap.
• The user should be able to kick off a retry on permanently failed messages by selecting retry (rather than forcing them to manually copy-and-resend).
• The user should be able to kick off a retry on soft-failed messages (i.e. force a connectivity check) somehow - e.g. by re-foregrounding the app, or re-opening a room (or worst case, sending a msg).
• The app should never ever spontaneously discard unsent messages.

What happened instead?

• Messages got hard-failed, and then got un-hard-failed, and then disappeared.
• There wasn't an obvious way to trigger retries, especially on hard-failed msgs.

Your phone model

iPhone 12 Pro Max

Operating system version

iOS 17.5.1

Application version

631

Homeserver

matrix.org

Will you send logs?

Yes

[bnjbvr]

It immediately hard-failed (red circle), rather than queuing.

Sending a message requires doing a /members request, and we try to avoid running more than one at the same time; so these requests are deduplicated; the first one failed, causing the second (waiting) to fail and be flagged as unrecoverable. There's a real bug there, and we should likely flag this specific kind of error as recoverable too.

That's the one new bug from looking at this rageshake, the rest has been fixed in the SDK, as far as I can tell:

Later, going back into the room when on good connectivity, the message is still there... except now it's in queued state (hollow-circle). Except despite now being on good connectivity, it doesn't send.

Here's, it's the timeline state that's incorrect; it should still have stayed in hard fail mode. This bug has been fixed in the SDK, but the app was using an old version of the SDK at this point that didn't include the patch. (SDK fix)

The app should never ever spontaneously discard unsent messages.

The app is using SDK's commit e89659b6, which didn't have the on-disk persistence yet, so an app restart would lose unsent messages. SDK patch.

[bnjbvr]

matrix-org/matrix-rust-sdk#3619 should fix it.

[pixlwave]

Closing this as it should be included in Nightly by now and we've been testing queuing this morning.