move stuff from synapse/doc

drafts/model/presence.rst

@@ -0,0 +1,149 @@
+API Efficiency
+==============
+
+A simple implementation of presence messaging has the ability to cause a large
+amount of Internet traffic relating to presence updates. In order to minimise
+the impact of such a feature, the following observations can be made:
+
+ * There is no point in a Home Server polling status for peers in a user's
+   presence list if the user has no clients connected that care about it.
+
+ * It is highly likely that most presence subscriptions will be symmetric - a
+   given user watching another is likely to in turn be watched by that user.
+
+ * It is likely that most subscription pairings will be between users who share
+   at least one Room in common, and so their Home Servers are actively
+   exchanging message PDUs or transactions relating to that Room.
+
+ * Presence update messages do not need realtime guarantees. It is acceptable to
+   delay delivery of updates for some small amount of time (10 seconds to a
+   minute).
+
+The general model of presence information is that of a HS registering its
+interest in receiving presence status updates from other HSes, which then
+promise to send them when required. Rather than actively polling for the
+current state all the time, HSes can rely on their relative stability to only
+push updates when required.
+
+A Home Server should not rely on the longterm validity of this presence
+information, however, as this would not cover such cases as a user's server
+crashing and thus failing to inform their peers that users it used to host are
+no longer available online. Therefore, each promise of future updates should
+carry with a timeout value (whether explicit in the message, or implicit as some
+defined default in the protocol), after which the receiving HS should consider
+the information potentially stale and request it again.
+
+However, because of the likelihood that two home servers are exchanging messages
+relating to chat traffic in a room common to both of them, the ongoing receipt
+of these messages can be taken by each server as an implicit notification that
+the sending server is still up and running, and therefore that no status changes
+have happened; because if they had the server would have sent them. A second,
+larger timeout should be applied to this implicit inference however, to protect
+against implementation bugs or other reasons that the presence state cache may
+become invalid; eventually the HS should re-enquire the current state of users
+and update them with its own.
+
+The following workflows can therefore be used to handle presence updates:
+
+ 1 When a user first appears online their HS sends a message to each other HS
+   containing at least one user to be watched; each message carrying both a
+   notification of the sender's new online status, and a request to obtain and
+   watch the target users' presence information. This message implicitly
+   promises the sending HS will now push updates to the target HSes.
+
+ 2 The target HSes then respond a single message each, containing the current
+   status of the requested user(s). These messages too implicitly promise the
+   target HSes will themselves push updates to the sending HS.
+
+   As these messages arrive at the sending user's HS they can be pushed to the
+   user's client(s), possibly batched again to ensure not too many small
+   messages which add extra protocol overheads.
+
+At this point, all the user's clients now have the current presence status
+information for this moment in time, and have promised to send each other
+updates in future.
+
+ 3 The HS maintains two watchdog timers per peer HS it is exchanging presence
+   information with. The first timer should have a relatively small expiry
+   (perhaps 1 minute), and the second timer should have a much longer time
+   (perhaps 1 hour).
+
+ 4 Any time any kind of message is received from a peer HS, the short-term
+   presence timer associated with it is reset.
+
+ 5 Whenever either of these timers expires, an HS should push a status reminder
+   to the target HS whose timer has now expired, and request again from that
+   server the status of the subscribed users.
+
+ 6 On receipt of one of these presence status reminders, an HS can reset both
+   of its presence watchdog timers.
+
+To avoid bursts of traffic, implementations should attempt to stagger the expiry
+of the longer-term watchdog timers for different peer HSes.
+
+When individual users actively change their status (either by explicit requests
+from clients, or inferred changes due to idle timers or client timeouts), the HS
+should batch up any status changes for some reasonable amount of time (10
+seconds to a minute). This allows for reduced protocol overheads in the case of
+multiple messages needing to be sent to the same peer HS; as is the likely
+scenario in many cases, such as a given human user having multiple user
+accounts.
+
+
+API Requirements
+================
+
+The data model presented here puts the following requirements on the APIs:
+
+Client-Server
+-------------
+
+Requests that a client can make to its Home Server
+
+ * get/set current presence state
+   Basic enumeration + ability to set a custom piece of text
+
+ * report per-device idle time
+   After some (configurable?) idle time the device should send a single message
+   to set the idle duration. The HS can then infer a "start of idle" instant and
+   use that to keep the device idleness up to date. At some later point the
+   device can cancel this idleness.
+
+ * report per-device type
+   Inform the server that this device is a "mobile" device, or perhaps some
+   other to-be-defined category of reduced capability that could be presented to
+   other users.
+
+ * start/stop presence polling for my presence list
+   It is likely that these messages could be implicitly inferred by other
+   messages, though having explicit control is always useful.
+
+ * get my presence list
+   [implicit poll start?]
+   It is possible that the HS doesn't yet have current presence information when
+   the client requests this. There should be a "don't know" type too.
+
+ * add/remove a user to my presence list
+
+Server-Server
+-------------
+
+Requests that Home Servers make to others
+
+ * request permission to add a user to presence list
+
+ * allow/deny a request to add to a presence list
+
+ * perform a combined presence state push and subscription request
+   For each sending user ID, the message contains their new status.
+   For each receiving user ID, the message should contain an indication on
+   whether the sending server is also interested in receiving status from that
+   user; either as an immediate update response now, or as a promise to send
+   future updates.
+
+Server to Client
+----------------
+
+[[TODO(paul): There also needs to be some way for a user's HS to push status
+updates of the presence list to clients, but the general server-client event
+model currently lacks a space to do that.]]