Towards an Internet-Scale XML Dissemination Service

1. Towards an Internet-Scale XML Dissemination Service Yanlei Diao Shariq Rizvi Michael J. Franklin EECS, U.C. Berkeley

2. Outline • XML dissemination services • System model • Core techniques • Status and conclusions

3. Applications of XML Dissemination • News feeds via RSS (Really Simple Syndication) • My Yahoo!: updated headlines from BBC, CNet, NPR. • Mobile services • Mobile operators: connect content providers with millions of clients running a multitude of operating systems. • Stock tickers • QuoteMedia: fast access to real-time and historical stock data. • Online auctions: • freebidingtools.com: create your own feed for your favorite eBay search. • Network monitoring: • Ganglia: a distributed monitoring system for clusters and grids.

4. XML streams user queries Data Source Data Source Data Source query results YFilter: An XML Dissemination Service YFilter • User queries: Specification of data interests, written in an XML query language. • Data sources: Continuously publish XML data items. • The service: Delivers to each user the XML data items that match her data interests; the delivered results are presented in a customized format.

5. Broker Broker Broker Broker Broker U1 U2 U4 U3 Data Source U5 Data Source Data Source Broker ONYX: Large-Scale XML Dissemination ONYX • Operator Network using YFilter for XML Dissemination YFilter • An overlay network of information brokers running YFilter. • Underlying infrastructures: • A dedicated network • Peer-to-peer • Collaboration among administrative domains

6. Design Space: Expressiveness Expressiveness: data model + query language a service supports • Subject-based • Messages: a subject label • Queries: a specific label or a wildcard • Predicate-based • Messages: attribute-value pairs • Queries: a set of predicates • XML filtering • Messages: XML • Queries: subset of XPath 1.0 • XML filtering and transformation • Messages: XML • Queries: subset of XQuery

7. Design Space: Why Distributed Processing? • Privacy • Regulations: e.g., CA Senate Bill No. 1386. • Policies: e.g., customers’ data stay behind the firewall. • Locality of data interests • Disseminate regional data directly to local subscribers. • Scalability • Data volume:number of messages per second up to thousands, message size from 1 KB to 20 KB. • Query population: up to millions. • Frequency of query updates: from a daily basis to every few minutes. • Result Volume: can amplify the input data volume by a large factor.

8. Related Systems

9. Content of the Paper • Content-driven routing • Need to handle both structural and value-based constraints. • Leverage YFilter: NFA-based operator networks, distributed construction. • Filtering power of routing (i.e., fraction of messages filtered) • Filtering power can be inherently limited. • Use query partitioning (if possible) to improve it. • Distributed transformation • Currently either at the publishers’ side or at the edge brokers. • Perform cascading message transformation during routing. • Efficient XML transmission • Verbosity of XML, and XML parsing at each routing step. • Investigate different XML formats for XML transmission. • Detailed architectural design • Other optimization techniques…

10. Outline • XML dissemination services • System model • Core techniques • Status and conclusions

11. Broker 2: /nitf/head/pubdata[@edition.area=“NY”] Broker 4: /nitf/head/pubdata[@edition.area=“SF”] Data Source data flow Broker 2: … Broker 1 Broker 5: /nitf//toject.subject[@toject.subject.type=“Stock”] or /nitf//toject.subject[@toject.subject.matter=“fishing”] Broker 6: /nitf//toject.subject[@toject.subject.type =“Sports”] Broker 2 Broker 4 query flow Broker 3 Q4: … Q5: … Broker 6 Broker 5 Q1: /nitf[head/pubdata[@edition.area=“SF”]] [.//toject.subject[@toject.subject.type=“Stock”]] Q2: /nitf[head/pubdata[@edition.area=”SF”]] [.//toject.subject[@toject.subject.matter=“fishing”]] Q3: /nitf[head/pubdata[@edition.area=“SF”]] [.//series[@series.name=“Tide Forecasts”]] Operations on Data/Query flows [a transformation query*]

12. Build routing tables Lookup in routing tables Execute transformation plans Build transformation plans Execute query plans Build query plans System Tasks on Data/Query Planes Processing planes:query planeanddata plane

13. Outline • XML dissemination services • System model • Core techniques • Status and conclusions

14. Routing Table Design • A routing table: mapping from output links to routing queries. • a routing query: the data interests of queries down from an output link. • data interest of a query: XPathexpressions, for and where clauses of FLWOR expressions. • Routing table design • a canonical form of routing queries; • a representation of routing tables; and • an algorithm constructing them from a distributed query population. • Two (conflicting) goals • High filtering power of routing • Fraction of messages filtered in routing. • High routing efficiency • Number of messages routed per second.

15. 1 nitf head  2 * 3 5 pubdata toject. subject 6 4    Q2 Q1 YFilter Basics • An XML filtering and transformation engine that processes multiple queries in a shared fashion. • ANon-Deterministic Finite Automaton(NFA)-basedoperator network. • Benefits for routing: • Fast structure matching. • A small maintenance cost for query updates. • Extensibility for supporting new operators. Q1: /nitf [head/pubdata[@edition.area=“SF”]] [.//tobject.subject[@tobject.subject.type=“Stock”]] Q2: /nitf [head/pubdata[@edition.area=“SF”]] [.//tobject.subject[@tobject.subject.matter=“fishing”]] • Y. Diao and M.J. Franklin. Query Processing for High-Volume XML Message Brokering. VLDB 2003. • Y. Diao, et al. Path Sharing and Predicate Evaluation for High-Performance XML Filtering. TODS, Dec. 2003. • YFilter v1.0 release: Coming later this month!

16. Our Solution • Routing queries are a disjunction of path expressions • Each XPath expression (equivalent of the for and where clauses of FLOWR expressions) is a routing query. • Multiple routing queries can be connected by or. • Routing table representation • Merge routing queries into a singlecombined operator network. • Construction algorithm • Map(): a user query  a routing query in the canonical form. • Collect(): routing queries sent from child brokers  a routing table. • Aggregate(): all the routing queries (at a node) a new routing query.

17. Broker 4 A new routing query (4d) Aggregate( ) Routing Table (mapping from links to routing queries) (4c) Collect( ) from Broker 5 from Broker 6 Broker 5 Broker 6 (5b) A new routing query (6b) A new routing query Aggregate( ) Aggregate( ) Routing queries Routing queries Map( ) Map( ) (5a) (6a) Q1: /nitf[head/pubdata/@edition.area=“SF”] [.//toject.subject/@toject.subject.type=“Stock” ] Q2: /nitf[head/pubdata/@edition.area=“SF”] [.//toject.subject/@toject.subject.matter=“fishing”] Q3: /nitf[head/pubdata/@edition.area=“SF”] [.//series/series.name=“Tide Forecasts” ] An Example Scenario

18. (5b) (6b) 1 1 1 nitf nitf nitf Collect( ) head head head    2 2 2 * * * 3 3 3 5 5 5 pubdata pubdata pubdata toject. subject toject. subject toject. subject 4 7 6 4 6 4 7 7 6 4 4     series  1 1 nitf nitf Broker5 Broker5 Broker6   head head 2 2 * * 3 3 5 5 series series Map( ) & Aggregate( ) Map( ) &Aggregate( ) (4c’) pubdata pubdata (5a) (6a)          Broker5 Broker5 Broker6 Q3 Q1 Q2 Example (continued)

19. (4c) Broker 5:Broker 6: (5b) (6b) 1 1 nitf nitf head head   2 2 * * 3 3 5 5 pubdata pubdata toject. subject toject. subject 7 7 4 4 4 6 6 (5b) (6b) series 1 nitf  head 2 (4c’) * 3 5 series pubdata      Broker5 Broker5 Broker6     Broker5 Broker6 Broker5 Sharing and Short-cut Evaluation • A problem with sharing • Separate routing query representations: short-cut evaluation. • Combined one: sharing may sacrifice the short-cut evaluation strategy. • Solution: dynamic pruning of the operator network at runtime • Each operator/NFA state has a static set of broker ids that it can reach. • System keeps a dynamic set of broker ids that have been reached. • YFilter execution is extended to prune the operator network using these sets.

20. Other Routing Considerations: • Content Generalization • Large routing tables can be a problem. • Introduce content generation as an additional step in Collect( ) or Aggregate( ). • Generalization methods. • Trade off filtering power for routing (space) efficiency. • Filtering Power of Routing • Fraction of messages filtered by routing. • Selectivity of the union of the user queries at the node. • Loss in precision in the routing queries representing this node. • If inherently low, partition the query population to improve it. • An Exclusiveness Pattern: e.g.,“/a/b[@id=?]” • Identify a set of such patterns, and partition queries using them.

21. Status and Conclusions • Queries bring intelligence to the network routing fabric. • We present a detailed architectural design of ONYX. • We address fundamental issues. • YFilter’s NFA-based operator networks are good for routing! • Locality of data interests is key to filtering power! • Status: YFilter release, XML transmission, other implementation underway. • This is an area full of opportunities for optimization. • Improving routing efficiency. • Improving filtering power of routing. • Incremental message transformation. • Sharing among different processing tasks. • Schema-based optimization…

22. Questions ONYX