1	End of Query Optimization Data Integration May 24, 2004
2	Agenda Questions? Finish last bits of query optimization Data integration: the last frontier
3	Query Execution
4	Query Execution Plans
5	We’ve Seen So Far Transformation rules The cost module: Given a candidate plan: what is its expected cost and size of the result? Now: putting it all together.
6	Plans for Single-Relation Queries (Prep for Join ordering) Task: create a query execution plan for a single Select-project-group-by block. Key idea: consider each possible access path to the relevant tuples of the relation. Choose the cheapest one. The different operations are essentially carried out together (e.g., if an index is used for a selection, projection is done for each retrieved tuple, and the resulting tuples are pipelined into the aggregate computation).
7	Example If we have an Index on rating: (1/NKeys(I)) * NTuples(R) = (1/10) * 40000 tuples retrieved. Clustered index: (1/NKeys(I)) * (NPages(I)+NPages(R)) = (1/10) * (50+500) pages are retrieved (= 55). Unclustered index: (1/NKeys(I)) * (NPages(I)+NTuples(R)) = (1/10) * (50+40000) pages are retrieved. If we have an index on sid: Would have to retrieve all tuples/pages. With a clustered index, the cost is 50+500. Doing a file scan: we retrieve all file pages (500).
8	Determining Join Ordering R1 R2 …. Rn Join tree: A join tree represents a plan. An optimizer needs to inspect many (all ?) join trees
9	Types of Join Trees Left deep:
10	Types of Join Trees Bushy:
11	Types of Join Trees Right deep:
12	Problem Given: a query R1 R2 … Rn Assume we have a function cost() that gives us the cost of every join tree Find the best join tree for the query
13	Join Ordering by Dynamic Programming Idea: for each subset of {R1, …, Rn}, compute the best plan for that subset In increasing order of set cardinality: Step 1: for {R1}, {R2}, …, {Rn} Step 2: for {R1,R2}, {R1,R3}, …, {Rn-1, Rn} … Step n: for {R1, …, Rn} A subset of {R1, …, Rn} is also called a subquery
14	Dynamic Programming: step 1 Step 1: For each {Ri} do: Size({Ri}) = B(Ri) Plan({Ri}) = Ri Cost({Ri}) = (cost of scanning Ri)
15	Dynamic Programming: step i: Step i: For each Q in {R1, …, Rn} of cardinality i do: Compute Size(Q) For every pair of subqueries Q’, Q’’ s.t. Q = Q’ U Q’’ compute cost(Plan(Q’) Plan(Q’’)) Cost(Q) = the smallest such cost Plan(Q) = the corresponding plan
16	A few practical considerations Heuristics for reducing the search space Restrict to left linear trees Restrict to trees “without cartesian product” Need more than just one plan for each subquery: “interesting orders”: save a single plan for every possible ordering of the result. Why?
17	Query Optimization Summary Create initial (naïve) query execution plan. Apply transformation rules: Try to un-nest blocks Move predicates and grouping operators. Consider each block at a time: Determine join order Push selections, projections if possible.
18	Data Integration
19	What is Data Integration Providing Uniform (same query interface to all sources) Access to (queries; eventually updates too) Multiple (we want many, but 2 is hard too) Autonomous (DBA doesn’t report to you) Heterogeneous (data models are different) Structured (or at least semi-structured) Data Sources (not only databases).
20	The Problem: Data Integration Uniform query capability across autonomous, heterogeneous data sources on LAN, WAN, or Internet
21	Motivation(s) Enterprise data integration; web-site construction. WWW: Comparison shopping Portals integrating data from multiple sources B2B, electronic marketplaces Science and culture: Medical genetics: integrating genomic data Astrophysics: monitoring events in the sky. Environment: Puget Sound Regional Synthesis Model Culture: uniform access to all cultural databases produced by countries in Europe.
22	Discussion Why is it hard? How will we solve it?
23	Current Solutions Mostly ad-hoc programming: create a special solution for every case; pay consultants a lot of money. Data warehousing: load all the data periodically into a warehouse. 6-18 months lead time Separates operational DBMS from decision support DBMS. (not only a solution to data integration). Performance is good; data may not be fresh. Need to clean, scrub you data.
24	Data Warehouse Architecture
25	The Virtual Integration Architecture Leave the data in the sources. When a query comes in: Determine the relevant sources to the query Break down the query into sub-queries for the sources. Get the answers from the sources, and combine them appropriately. Data is fresh. Challenge: performance.
26	Virtual Integration Architecture
27	Data Integration: Higher-level Abstraction
28
29	Research Projects Garlic (IBM), Information Manifold (AT&T) Tsimmis, InfoMaster (Stanford) The Internet Softbot/Razor/Tukwila (UW) Hermes (Maryland) DISCO, Agora (INRIA, France) SIMS/Ariadne (USC/ISI) Many, many more!
30	Semantic Mappings Differences in: Names in schema Attribute grouping Coverage of databases Granularity and format of attributes
31	Issues for Semantic Mappings
32	Beyond Data Integration Mediated schema is a bottleneck for large-scale data sharing It’s hard to create, maintain, and agree upon.
33	Peer Data Management Systems
34	PDMS-Related Projects Hyperion (Toronto) PeerDB (Singapore) Local relational models (Trento) Edutella (Hannover, Germany) Semantic Gossiping (EPFL Zurich) Raccoon (UC Irvine) Orchestra (Ives, U. Penn)
35	A Few Comments about Commerce Until 5 years ago: Data integration = Data warehousing. Since then: A wave of startups: Nimble, MetaMatrix, Calixa, Composite, Enosys Big guys made announcements (IBM, BEA). [Delay] Big guys released products. Success: analysts have new buzzword – EII New addition to acronym soup (with EAI). Lessons: Performance was fine. Need management tools.
36	Data Integration: Before
37	Data Integration: After
38	Sound Business Models Explosion of intranet and extranet information 80% of corporate information is unmanaged By 2004 30X more enterprise data than 1999 The average company: maintains 49 distinct enterprise applications spends 35% of total IT budget on integration-related efforts
39	Sound Business Models Explosion of intranet and extranet information 80% of corporate information is unmanaged By 2004 30X more enterprise data than 1999 The average company: maintains 49 distinct enterprise applications spends 35% of total IT budget on integration-related efforts
40	Dimensions to Consider How many sources are we accessing? How autonomous are they? Meta-data about sources? Is the data structured? Queries or also updates? Requirements: accuracy, completeness, performance, handling inconsistencies. Closed world assumption vs. open world?
41	Outline Wrappers Semantic integration and source descriptions: Modeling source completeness Modeling source capabilities Query optimization Query execution Peer-data management systems Creating schema mappings
42	Wrapper Programs Task: to communicate with the data sources and do format translations. They are built w.r.t. a specific source. They can sit either at the source or at the mediator. Often hard to build (very little science). Can be “intelligent”: perform source-specific optimizations.
43	Example
44	Data Source Catalog Contains all meta-information about the sources: Logical source contents (books, new cars). Source capabilities (can answer SQL queries) Source completeness (has all books). Physical properties of source and network. Statistics about the data (like in an RDBMS) Source reliability Mirror sources Update frequency.
45	Content Descriptions User queries refer to the mediated schema. Data is stored in the sources in a local schema. Content descriptions provide the semantic mappings between the different schemas. Data integration system uses the descriptions to translate user queries into queries on the sources.
46	Desiderata from Source Descriptions Expressive power: distinguish between sources with closely related data. Hence, be able to prune access to irrelevant sources. Easy addition: make it easy to add new data sources. Reformulation: be able to reformulate a user query into a query on the sources efficiently and effectively.
47	Reformulation Problem Given: A query Q posed over the mediated schema Descriptions of the data sources Find: A query Q’ over the data source relations, such that: Q’ provides only correct answers to Q, and Q’ provides all possible answers from to Q given the sources.
48	Languages for Schema Mapping
49	Global-as-View Mediated schema: Movie(title, dir, year, genre), Schedule(cinema, title, time). Create View Movie AS select * from S1 [S1(title,dir,year,genre)] union select * from S2 [S2(title, dir,year,genre)] union [S3(title,dir), S4(title,year,genre)] select S3.title, S3.dir, S4.year, S4.genre from S3, S4 where S3.title=S4.title
50	Global-as-View: Example 2 Mediated schema: Movie(title, dir, year, genre), Schedule(cinema, title, time). Create View Movie AS [S1(title,dir,year)] select title, dir, year, NULL from S1 union [S2(title, dir,genre)] select title, dir, NULL, genre from S2
51	Global-as-View: Example 3 Mediated schema: Movie(title, dir, year, genre), Schedule(cinema, title, time). Source S4: S4(cinema, genre) Create View Movie AS select NULL, NULL, NULL, genre from S4 Create View Schedule AS select cinema, NULL, NULL from S4. But what if we want to find which cinemas are playing comedies?
52	Global-as-View Summary Query reformulation boils down to view unfolding. Very easy conceptually. Can build hierarchies of mediated schemas. You sometimes loose information. Not always natural. Adding sources is hard. Need to consider all other sources that are available.
53	Local-as-View (LAV)
54	Query Reformulation
55	Query Reformulation
56	Local-as-View: example 1 Mediated schema: Movie(title, dir, year, genre), Schedule(cinema, title, time). Create Source S1 AS select * from Movie Create Source S3 AS [S3(title, dir)] select title, dir from Movie Create Source S5 AS select title, dir, year from Movie where year > 1960 AND genre=“Comedy”
57	Local-as-View: Example 2 Mediated schema: Movie(title, dir, year, genre), Schedule(cinema, title, time). Source S4: S4(cinema, genre) Create Source S4 select cinema, genre from Movie m, Schedule s where m.title=s.title . Now if we want to find which cinemas are playing comedies, there is hope!
58	Local-as-View Summary Very flexible. You have the power of the entire query language to define the contents of the source. Hence, can easily distinguish between contents of closely related sources. Adding sources is easy: they’re independent of each other. Query reformulation: answering queries using views!
59	The General Problem Given a set of views V1,…,Vn, and a query Q, can we answer Q using only the answers to V1,…,Vn? Many, many papers on this problem. The best performing algorithm: The MiniCon Algorithm, (Pottinger & Levy, 2000). Great survey on the topic: (Halevy, 2001).
60	Local Completeness Information If sources are incomplete, we need to look at each one of them. Often, sources are locally complete. Movie(title, director, year) complete for years after 1960, or for American directors. Question: given a set of local completeness statements, is a query Q’ a complete answer to Q?
61	Example Movie(title, director, year) (complete after 1960). Show(title, theater, city, hour) Query: find movies (and directors) playing in Seattle: Select m.title, m.director From Movie m, Show s Where m.title=s.title AND city=“Seattle” Complete or not?
62	Example #2 Movie(title, director, year), Oscar(title, year) Query: find directors whose movies won Oscars after 1965: select m.director from Movie m, Oscar o where m.title=o.title AND m.year=o.year AND o.year > 1965. Complete or not?
63	Query Optimization Very related to query reformulation! Goal of the optimizer: find a physical plan with minimal cost. Key components in optimization: Search space of plans Search strategy Cost model
64	Optimization in Distributed DBMS A distributed database (2-minute tutorial): Data is distributed over multiple nodes, but is uniform. Query execution can be distributed to sites. Communication costs are significant. Consequences for optimization: Optimizer needs to decide locality Need to exploit independent parallelism. Need operators that reduce communication costs (semi-joins).
65	DDBMS vs. Data Integration In a DDBMS, data is distributed over a set of uniform sites with precise rules. In a data integration context: Data sources may provide only limited access patterns to the data. Data sources may have additional query capabilities. Cost of answering queries at sources unknown. Statistics about data unknown. Transfer rates unpredictable.
66	Modeling Source Capabilities Negative capabilities: A web site may require certain inputs (in an HTML form). Need to consider only valid query execution plans. Positive capabilities: A source may be an ODBC compliant system. Need to decide placement of operations according to capabilities. Problem: how to describe and exploit source capabilities.
67	Example #1: Access Patterns Mediated schema relation: Cites(paper1, paper2) Create Source S1 as select * from Cites given paper1 Create Source S2 as select paper1 from Cites Query: select paper1 from Cites where paper2=“Hal00”
68	Example #1: Continued Create Source S1 as select * from Cites given paper1 Create Source S2 as select paper1 from Cites Select p1 From S1, S2 Where S2.paper1=S1.paper1 AND S1.paper2=“Hal00”
69	Example #2: Access Patterns Create Source S1 as select * from Cites given paper1 Create Source S2 as select paperID from UW-Papers Create Source S3 as select paperID from AwardPapers given paperID Query: select * from AwardPapers
70	Example #2: Solutions Can’t go directly to S3 because it requires a binding. Can go to S1, get UW papers, and check if they’re in S3. Can go to S1, get UW papers, feed them into S2, and feed the results into S3. Can go to S1, feed results into S2, feed results into S2 again, and then feed results into S3. Strictly speaking, we can’t a priori decide when to stop. Need recursive query processing.
71	Handling Positive Capabilities Characterizing positive capabilities: Schema independent (e.g., can always perform joins, selections). Schema dependent: can join R and S, but not T. Given a query, tells you whether it can be handled. Key issue: how do you search for plans? Garlic approach (IBM): Given a query, STAR rules determine which subqueries are executable by the sources. Then proceed bottom-up as in System-R.
72	Matching Objects Across Sources How do I know that A. Halevy in source 1 is the same as Alon Halevy in source 2? If there are uniform keys across sources, no problem. If not: Domain specific solutions (e.g., maybe look at the address, ssn). Use Information retrieval techniques (Cohen, 98). Judge similarity as you would between documents. Use concordance tables. These are time-consuming to build, but you can then sell them for lots of money.
73	Optimization and Execution Problem: Few and unreliable statistics about the data. Unexpected (possibly bursty) network transfer rates. Generally, unpredictable environment. General solution: (research area) Adaptive query processing. Interleave optimization and execution. As you get to know more about your data, you can improve your plan.
74	Tukwila Data Integration System Novel components: Event handler Optimization-execution loop
75	Double Pipelined Join (Tukwila) Hash Join Partially pipelined: no output until inner read Asymmetric (inner vs. outer) — optimization requires source behavior knowledge
76	Semantic Mappings Need mappings in every data sharing architecture “Standards are great, but there are too many.”
77	Why is it so Hard? Schemas never fully capture their intended meaning: We need to leverage any additional information we may have. A human will always be in the loop. Goal is to improve designer’s productivity. Solution must be extensible. Two cases for schema matching: Find a map to a common mediated schema. Find a direct mapping between two schemas.
78	Typical Matching Heuristics We build a model for every element from multiple sources of evidences in the schemas Schema element names BooksAndCDs/Categories ~ BookCategories/Category Descriptions and documentation ItemID: unique identifier for a book or a CD ISBN: unique identifier for any book Data types, data instances DateTime ¹ Integer, addresses have similar formats Schema structure All books have similar attributes
79	Using Past Experience Matching tasks are often repetitive Humans improve over time at matching. A matching system should improve too! LSD: Learns to recognize elements of mediated schema. [Doan, Domingos, H., SIGMOD-01, MLJ-03] Doan: 2003 ACM Distinguished Dissertation Award.
80	Example: Matching Real-Estate Sources
81	Learning Source Descriptions We learn a classifier for each element of the mediated schema. Training examples are provided by the given mappings. Multi-strategy learning: Base learners: name, instance, description Combine using stacking. Accuracy of 70-90% in experiments. Learning about the mediated schema.
82	Multi-Strategy Learning Use a set of base learners: Name learner, Naïve Bayes, Whirl, XML learner And a set of recognizers: County name, zip code, phone numbers. Each base learner produces a prediction weighted by confidence score. Combine base learners with a meta-learner, using stacking.
83	The Semantic Web A web of structured data: The 5-year old vision of Tim Berners-Lee How does it relate to data integration? How are we going to do it? Why should we do it? Do we need a killer app or is the semantic web a killer app?
84	The End