- From: Adam Bergkvist <adam.bergkvist@ericsson.com>
- Date: Wed, 23 Jan 2013 11:44:19 +0100
- To: "public-webrtc@w3.org" <public-webrtc@w3.org>
Hi I'm currently working on putting version 2 of the stats into the spec as decided on the last telco. Overall I think this works, but always getting objects by id makes the code more complicated. There are situations where ids are needed (e.g., to correlate stats objects between getStats() calls), but I think we should get objects directly when possible. Some comments: The current approach to gathering stats is to gather all possible kinds of stats for a chosen selector and then, after you get the report and the stats objects, pick the interesting parts. (I know that Ekr commented this on the last telco.) I propose we do the filtering at getStats()-time instead of after the report is created. The result would be more conservative stats gathering and smaller reports. void getStats (MediaStreamTrack? selector, RTCStatsCategory[]? categories, RTCStatsCallback successCallback, RTCPeerConnectionErrorCallback failureCallback); // names taken from the stats dictionaries in the proposal enum RTCStatsCategory { "rtp-stream", "outgoing-rtp-stream", "incoming-rtp-stream", "incoming-local-rtp-stream", "rtp-session", "candidate", "candidate-pair" // ... }; The same enum would also be reused as argument values the objects() method. > interface RTCStatsReport { > sequence<DOMString> objects(DOMString type?); > RTCStatsObject get(DOMString id); > } > interface RTCStatsObject { > ... > readonly DOMString id; > ... > }; It's not obvious what the objects() method returns; the name suggests stat objcets but the return value is a sequence of DOMStrings. From the example it seems like it returns ids. This seems a bit complicated. Why not let it return a sequence of RTCStatObjects (of a certain type) directly as the name implies instead of requiring detour with ids and get(). var statsObjects = report.objects("rtp-stream"); A corresponding stats object from a later report (for the first object) would then be retrieved with: secondReport.get(statsObjects[0].id); So from what I understand, the id links stats objects that correspond to the same selector sub-component for a specific stats category between reports. For example, sub component=SSRC when selector=MediaStreamTrack and category=rtp-stream. This makes sense since it isn't practical to use indexes if the number of stats objects differs between the first and the second report. Could we have a name that emphasizes the difference between this id and ids on MediaStream and MediaStreamTrack? I.e. that two objects with the same id aren't the same object, instedad they are results of reporting stats from the same underlying component. E.g. sourceId, componentId, providerId or something. > In order to have a well known syntax for RTCStats objects, we use > Dictionary declarations. This does NOT mean that an RTCStatsObject is > or contains a dictionary; due to implementation issues, we still have > only the names() and getValue functions as interfaces to the > RTCStatsObject. I'm not sure it's good idea to take a syntax that is well defined is this context and say that it means something else. I think the described dictionaries looks good and it would be more straight forward to use them as dictionaries (and update implementations to support this). RTCStatsObject could also be a dictionary since the two methods would become redundant (names() is replaced by property enumeration and getValue() by direct property access). This would also flatten each level of the stats structure a bit and make it easier to use. dictionary RTCStatsObject { readonly attribute long timestamp; readonly DOMString type; readonly DOMString id; }; dictionary RTPStream : RTCStatsObject { int ssrc; int? maxBandwidth; // as specified by user, if present ... } ... > dictionary RTPStream { > ... > StatsObjectID? otherEndStats; // Reference to the stats signalled > // from our partner. > ... > } What's the reason for having an id here and not referencing the otherEndStats directly? Can these stats be found via the object() method as well? > interface RTCStatsObject { > readonly attribute long timestamp; > ... We should use DOMTimeStamp here as defined by WebIDL [1]. ============================================================ The changes proposed above would make the API a bit easier to use by working directly with objcets and with less "get by id" operations. (new example) pc.getStats(selector, ["rtp-stream"], ...); // proccess stats var objects = secondReport.objects(); for (i = 0; i < objects.length; i++) { second = objects[i]; // get the corresponding stats object from the first report first = firstReport.get(second.id); if (first && first.remote) { var packetsSent = first.packetsSsent - second.packetsSent; var packetsReceived = first.remote.packetsReceived - second.remote.packetsReceived; // if fractionLost is > 0.3, we have probably found the culprit var fractionLost = (packetsSent - packetsReceived) / packetsSent; } } ------------------------------------------------------------ (existing example) pc.getStats(selector, ...); // proccess stats var ssrcIds = now.objects("RTPStream"); for (i = 0; i < ssrcIds.length; i++) { var ssrcStatsId = ssrcIds[i]; nowState = now.get(ssrcStatsId); prevState = baseline.get(ssrcStatsId); remoteStatsId = nowstate.get("otherEndStats"); remoteNowState = now.get(remoteStatsId); remotePrevState = baseline.get(remoteStatsId); if (prevstate && remotePrevState) { var packetsSent = nowstate.get(packetsSent) - prevstate.get(packetsSent); var packetsReceived = remoteNowState.get(packetsReceived) - remotePrevState.get(packetsReceived); // if fractionLost is > 0.3, we have probably found the culprit var fractionLost = (packetsSent - packetsReceived) / packetsSent; } } /Adam [1] http://dev.w3.org/2006/webapi/WebIDL/#common-DOMTimeStamp
Received on Wednesday, 23 January 2013 10:44:43 UTC