From 91ca36509b3112a83680a97a0d672a428b4d7fef Mon Sep 17 00:00:00 2001 From: Kegan Dougal Date: Wed, 7 Oct 2015 11:51:49 +0100 Subject: [PATCH] Flesh out IM module --- specification/0-intro.rst | 2 + specification/modules/instant_messaging.rst | 89 ++++++++++++++++++--- 2 files changed, 80 insertions(+), 11 deletions(-) diff --git a/specification/0-intro.rst b/specification/0-intro.rst index 445b32ef2..7c27e0d78 100644 --- a/specification/0-intro.rst +++ b/specification/0-intro.rst @@ -434,6 +434,8 @@ Some requests have unique error codes: :``M_LOGIN_EMAIL_URL_NOT_YET``: Encountered when polling for an email link which has not been clicked yet. +.. _sect:txn_ids: + The C-S API typically uses ``HTTP POST`` to submit requests. This means these requests are not idempotent. The C-S API also allows ``HTTP PUT`` to make requests idempotent. In order to use a ``PUT``, paths should be suffixed with diff --git a/specification/modules/instant_messaging.rst b/specification/modules/instant_messaging.rst index 7934b80a2..426189b58 100644 --- a/specification/modules/instant_messaging.rst +++ b/specification/modules/instant_messaging.rst @@ -53,7 +53,14 @@ display the fallback plain text ``body`` key instead. Client behaviour ---------------- -Events which have attachments (e.g. ``m.image``, ``m.file``) are advised to be +Clients SHOULD verify the structure of incoming events to ensure that the +expected keys exist and that they are of the right type. Clients can discard +malformed events or display a placeholder message to the user. Redacted +``m.room.message`` events MUST be removed from the client. This can either be +replaced with placeholder text (e.g. "[REDACTED]") or the redacted message can +be removed entirely from the messages view. + +Events which have attachments (e.g. ``m.image``, ``m.file``) SHOULD be uploaded using the `content repository module`_ where available. The resulting ``mxc://`` URI can then be used in the ``url`` key. The attachment SHOULD be uploaded *prior* to sending the event in order to stop a @@ -63,27 +70,87 @@ race condition where the recipient receives a link to a non-existent attachment. Recommendations when sending messages ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- Advise using idempotent PUTs to send messages (and why) -- Retries (exp backoff, giveup eventually allowing manual retry) -- Queueing (bucket per room) -Implementing local echo -~~~~~~~~~~~~~~~~~~~~~~~ -- Local echo (document bug with races) - sending state. Pairing returned event ID. +Clients can send messages using ``POST`` or ``PUT`` requests. Clients SHOULD use +``PUT`` requests with `transaction IDs`_ to make requests idempotent. This +ensures that messages are sent exactly once even under poor network conditions. +Clients SHOULD retry requests using an exponential-backoff algorithm for a +certain amount of time T. It is recommended that T is no longer than 5 minutes. +After this time, the client should stop retrying and mark the message as "unsent". +Users should be able to manually resend unsent messages. + +Users may type several messages at once and send them all in quick succession. +Clients SHOULD preserve the order in which they were sent by the user. This +means that clients should wait for the response to the previous request before +sending the next request. This can lead to head-of-line blocking. In order to +reduce the impact of head-of-line blocking, clients should use a queue per room +rather than a global queue, as ordering is only relevant within a single room +rather than between rooms. + +.. _`transaction IDs`: `sect:txn_ids`_ + +Local echo +~~~~~~~~~~ + +Messages SHOULD appear immediately in the message view when a user presses the +"send" button. This should occur even if the message is still sending. This is +referred to as "local echo". Clients SHOULD implement "local echo" of messages. + +Clients need to be able to pair up the "remote echo" from the server with the +"local echo" to prevent duplicate messages being displayed. Ideally this pairing +would occur transparently to the user: the UI would not flicker as it transitions +from local to remote. Flickering cannot be fully avoided in version 1 of the +client-server API. Two scenarios need to be considered: + +- The client sends a message and the remote echo arrives on the event stream + *after* the request to send the message completes. +- The client sends a message and the remote echo arrives on the event stream + *before* the request to send the message completes. + +In the first scenario, the client will receive an event ID when the request to +send the message completes. This ID can be used to identify the duplicate event +when it arrives on the event stream. However, in the second scenario, the event +arrives before the client has obtained an event ID. This makes it impossible to +identify it as a duplicate event. This results in the client displaying the +message twice for a fraction of a second before the the original request to send +the message completes. Once it completes, the client can take remedial actions +to remove the duplicate event by looking for duplicate event IDs. Version 2 of +the client-server API resolves this by attaching the transaction ID of the +sending request to the event itself. + Displaying membership information with messages ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- Member name linking (incl. pagination aka historical display names) +Clients may wish to show the display name and avatar URL of the room member who +sent a message. This can be achieved by inspecting the ``m.room.member`` event +for that user ID. + +When a user paginates the message history, clients may wish to show the +**historical** display name and avatar URL for a room member. This is possible +because older ``m.room.member`` events are returned when paginating. This can +be implemented efficiently by keeping two sets of room state: old and current. +As new events arrive and/or the user paginates back in time, these two sets of +state diverge from each other. New events update the current state and paginated +events update the old state. When paginated events are processed sequentially, +the old state represents the state of the room *at the time the event was sent*. +This can then be used to set the historical display name and avatar URL. Server behaviour ---------------- -- SHOULD enforce the body/msgtype keys are present (can 400 them) +Homeservers SHOULD enforce that ``m.room.message`` events have textual ``body`` +and ``msgtype`` keys by 400ing the request to send a message. Security considerations ----------------------- -- Not encrypted, link to E2E module. -- XSS: Should sanitise ALL KEYS before injecting as unsafe HTML (name/topic/body/etc) +Messages sent using this module are not encrypted. Encryption can be layered +over the top of this module: where the plaintext format is an ``m.room.message`` +conforming to this module. This can be achieved using the `E2E module`_. + +Clients should sanitise **all keys** for unsafe HTML to prevent Cross-Site +Scripting (XSS) attacks. This includes room names and topics. + +.. _`E2E module`: `module:e2e`_