- From: Tara Athan <taraathan@gmail.com>
- Date: Fri, 25 Sep 2015 08:39:40 -0400
- To: public-rsp@w3.org
I have a few points for discussion. Since attendance will be limited today, we can perhaps discuss these in a meeting of a semantics/data model task force. * The requirement in the definition of “substream” that it be finite is not consistent with the conventional use of the “sub” prefix in mathematics. Further, it is confusing and wasteful to have two terms for the same thing. I suggest we use “window” for a finite substream, and allow a substream to be infinite. Then, a window is a substream, but not every substream is a window) Example 1: if a stream is filtered to remove all timestamped graphs except those with a certain predicate `p`, then we obtain a sub stream. Example 2: If two streams are merged, then each of the original streams are substreams of the merged stream. * Regarding ordering of timestamps, I propose we consider the partial order over time instants and time intervals defined as follows: 1. time instants are treated as equivalent, for the purposes of this comparison, to a degenerate time interval where the start and end times are the same 2. time intervals are treated as closed for purposes of this comparison 3. If two timestamps have different end times, then the timestamp with the later end time is greater than the timestamp with the earlier end time (irregardless of the start time). 4. If two timestamps have the same end times, then the timestamp with the later start time is greater than the timestamp with the earlier start time. Note: if we define the `closure` of a temporal entity as the smallest closed time interval containing it, then the partial order above can be defined in terms of a total order on closed time intervals. * Clarification is needed in regard to stream, substream and window definitions: there may be timestamped graphs with the same timestamp or incomparable timestamps, and as currently written, the identity of a stream ```does``` depend on the order of occurrence of such timestamped graphs within the stream sequence. This is necessary for window functions, e.g. count-based, to have a deterministic output. However, this has the consequence that the merger of two streams is not unique. If we adopt the ordering proposed in the previous bullet, then we may consider defining the count-based window function not on the basis of number of timestamped graphs, but on the number of timestamp closures. Then we could modify the definition of stream to be independent of the order between timestamped graphs with the same or incomparable timestamps. This would allow us to define a unique merger of streams while maintaining the deterministic nature of count-based window functions. * In regard to the concern about multiple triples needed to express metadata of a graph - we need an example. But in general, RDF metadata, however complex it is, is just another graph, so such metadata could be represented as another time-stamped graph in the stream, e.g. with the same timestamp but different predicate expressing the relation “describes an observation that was observed at". This metadata graph would refer to the other timestamped graph by name (which therefore could not be a blank node). Possible disadvantage - as currently defined, a count-based window function that does not consider the predicate could capture an observation but not its metadata, or vice versa. This would be solved by redefining count-based as in the previous bullet. Regards, Tara
Received on Friday, 25 September 2015 12:40:09 UTC