The Sibdata Revolution September 2009 Nick Roussopoulos DCS & UMIACS & Univ. of Maryland

Nick Roussopoulos Data Management: Past to Current Structured Data Structured architectures

Nick Roussopoulos Data Management: Huh???

Nick Roussopoulos The Landscape Bell’s Law: Every decade, a new, lower cost, class of computers emerges, defined by platform, interface, and interconnect Mainframes 1960s Minicomputers 1970s Microcomputers/PCs 1980s Web-based computing 1990s Devices (Smart phones, PDAs, wireless sensors, RFID) 2000’s Enabling a new generation of applications that Mandate new data management methods & tools.

Nick Roussopoulos Data Then and Now The “Data Industrial Revolution”: Data used to be “hand-crafted”, now it’s generated by computers!!! The Data Integration quagmire: 40 years of continuous successes (sic) and still a long way to the end. Structure provides crucial understanding for making data usable and leads to discovery/innovation.

Nick Roussopoulos Data Streaming  Data Explosion Clickstream Barcodes PoS System Sensors RFID Telematics Inventory Exponential data growth New challenges: continuous, inter- connected, distributed, physical Shrinking business cycles More complex decisions Phones Transactional Systems

Nick Roussopoulos The Structure Spectrum Structured data (schema-first) regular, known, conforming, … e.g., Relational database Unstructured data (schema-never) freeform, irregular, e.g., plain text, images, audio, … Semi-structured data (schema-later) Provides structural information, but less constrained. e.g., XML, tagged text/media

Nick Roussopoulos Data Integration Integration is the ultimate schema-first problem. Requires complete understanding & disambiguation Structure (semantics) is both a key enabler and a key impediment here.

Nick Roussopoulos Structured Data: How much Conventional Wisdom: ~20% of data is structured currently. Consumer apps, enterprise search, multimedia apps are placing downward pressure on this.

Nick Roussopoulos State of the Art: Integration-in-the-large Team work, huge & expensive effort, excruciating pain Extremely long time lag between data generation and availability Custom-coded implementations that are often unsuccessful Clearing house of already discovered knowledge (the high overhead is for disambiguating the semantics of the heterogeneous data)

Nick Roussopoulos Future: Integration-in-the-small End-user, limited in scope, requires training Continuous as the data sources and equipment evolve End-user tools are needed Small cost, enormous opportunity for discovery and innovation

Nick Roussopoulos Sibling Data Aggregation and naming of disparate data regardless location Includes actual data, references to external data, queries that generate data, & programs to process data May include other sibdata Open vs Closed Open: continuous accumulation Closed: fixed snapshot (archival) Location Independent semantics

Nick Roussopoulos Web search results

Nick Roussopoulos Content vs URL Content  moore.com/

Nick Roussopoulos Deep-Web Queries SELECT y.title FROM Yahoo_Movies m WHERE m.title like Moore;

Nick Roussopoulos Result vs. Query Results are associated with the time the query was run Queries can be captured in sibdata and executed at will; thus the sibdata would be open and captures a different result each time it executes

Nick Roussopoulos Queries to Relational Databases Yahoo_Actors

Nick Roussopoulos Sibdata Deal with all the data from everywhere & in whatever form they come Data co-existence no integrated schema, no single warehouse Expand-as-you-go Integrate little by little as you need ETL Data mapping-integrating as you add more data

Nick Roussopoulos Sibdata Properties Lightweight Metadata captures the encapsulation, name, and provenance data Location-independent Accessible from anywhere Isolated Generated with no interference Durable Persist until dropped Secure Guarantee security defined by the creators and sources Compose multiple levels of security to its components

Nick Roussopoulos Comparison to Transactions Transactions grouping of many actions into an atomic transaction- ACID properties Substrate: database Sibdata Grouping of data into an atomic sibdata – LLADS Substrate: actions/transactions/data generators

Nick Roussopoulos Sibdata Infrastructure

Nick Roussopoulos Sibdata Servers Establish a global sibdata ID and name Creates and maintains metadata with provenance, users, security, etc. Provides searchable catalog Provides storage for non-sib compliant data sources Fault tolerance (replication)

Nick Roussopoulos Sib Protocols Establish Sibdata protocol Concurrency-Consistency issues (?) Sharing of data Name conventions Dispute resolution Distributed Logging Security Using chits Group and multi-valued ownership and visibility

Nick Roussopoulos User Interface Simple OS support Query Languages Graphical Languages ETL tools Extra functionality High dimensional indexing Mining

Nick Roussopoulos Conclusions Need to build Sib Infrastructure Refine the sibdata semantics Refine the security protocols For data aggregates User groups Great opportunities for innovation

Nick Roussopoulos Presentations & Project 3 X 7 students = 21 presentations ~2 per lecture Lecture dates Sep: 15, 22, 29 Oct: 6, 13, 20, 27 Nov: 3, 10, 17, 24 Dec: 1, 8 Project: Proposal due Sep 29 Discussion: Every lecture be prepared to give a 2-3 min progress report, papers found, etc.

Nick Roussopoulos Network Data Independence Hellerstein Berkeley Physical Data Independence Decoupling data from layout (not hard coded applications) Permits reorganization of data w/o affecting the apps Declarative query languages Using the schema Distributed Databases Transparency hides location from the user who acts as if he is accessing a centralized database Limited sites- not capable to expand to the mobility of and constant change of the configuration

Nick Roussopoulos Pilars of Data independence Indexes- offer indirection allowing modification of the underlying structure Schema based and declarative query languages & optimization table R 1 occurrence file

Nick Roussopoulos Sibdata Independence Encapsulation of dissimilar data Data can be moved, rearranged, altered Additional indices on top of Sibdata becomes part of the sibdata Naming and provenance data are fixed Do not change to the outside world Containment information (sibdata encapsulation within other sibdata) is guaranteed

Nick Roussopoulos DHT (Chord) Data centric distribution according to content- total data independence very large number of distributed servers Configuration changes rapidly (although this may not be really that important) Fault-tolerance (extra machines) Limited to single key searches (not range or join queries

Nick Roussopoulos Network Names & Services Internet Indirection Infrastructure (i3) Triggers (id,r) where id = global ID and r is an address to forward packets When a mobile user moves to r’, he modifies his trigger to (id,r’) It also supports 1-to-n mappings (anycast) Content Distribution Networks (Akamai) Replicates heavy data (images, videos) to multiple sites and redirects user accesses to those that are closer (indirection via location independence)

Nick Roussopoulos Relevant DB Technologies Distributed Aggregation Monitor networks (collecting stats) Computing synopses and pass it along Adaptive execution plans Feedback to the execution Commutative tasks to avoid extended delays Range search over DHT Trie hashing Still limited P2P & Mobile Databases

Nick Roussopoulos Pier: A P2P in situ Query Engine Goals Massively distributed processing Scallability Relaxed consistency (best effort) Architecture P2P Built on top of DHT Multicast to all related nodes (lscan) Pipelining the intermediate results

Nick Roussopoulos Pier Joins Stored in DHT Namespace=relation NR, NS resourceID =Primary Key (PK) instanceID =tuple # if not a PK Assume R and S are already DHT hashed using and Symmetric Join building phase lscan NR and NS eliminate unqualified tuples and not needed attributes Rehash all above tuples using namespace NQ resourceID=R.pkey*S.pkey Tuples are tagged with relation name SymmetricJoin Probing phase Probing in parallel with building (with callbacks) locally Satisfying tuples are either sent to the Qsite or DHT-ed for the pipelined op Consumes a lot of bandwidth

Nick Roussopoulos Better Joins Fetch Matches Hash only S lscan R and fetch NS tuples Rewriting Join using 2-way semijoin Project R & R on their PK and joining attribute Do symmetric join on these projections Rewriting Join using Bloom filters Create and DHT the Bloom filters Do lscan and access the Bloom filter to eliminate not joinable tuples

Nick Roussopoulos Conclusions for Pier P2P bring massive parallelism Repetitive data comparison over DHT brings along massive waste of bandwidth Smarter in situ distillation (2-way semijoins, Bloom filters) work better