GitLab

When communicating via mailboxes, the max latency and thus the retransmission interval are very long, so we need to be careful about any circumstances that could cause messages to be lost.

On the receiver side, if an error (such as an IO error, app shutdown, app crash or device crash) occurs while we're reading messages from a file, we need to ensure that the file can be read again after recovering from the error. Reading the file twice isn't possible with the current protocol stack because the pseudo-random tag at the start of the file is recognised on the first read and can't be recognised again.

To fix this we should divide the process of recognising a tag into two steps. The first step looks up the tag and returns the keys needed for authenticating and decrypting the stream header. The second step marks the tag as recognised and updates the reordering window.

When processing a file downloaded from a mailbox, we should defer the second step until the file has been completely processed.