Transcript Slides - SEAS - University of Pennsylvania

Introduction
Zachary G. Ives
University of Pennsylvania
CIS 650 – Implementing Data Management Systems
September 4, 2008
Welcome to CIS 650,
Database and Information Systems!
Instructor: Zachary Ives, zives@cis
 576 Levine Hall North
 Office hours: Wednesdays, 2PM
Home page: www.cis.upenn.edu/~zives/cis650/
Discussion group: [email protected]
Texts and readings:
 Hellerstein & Stonebraker: Readings in Database Systems, 4th ed.
 Most papers will be linked via the Web, but it’s often nice to have the book
 Supplementary papers (will be linked via schedule)
2
Course Format and Prerequisites
 Read “classic” and important research papers
 Lectures will be very discussion-oriented; about one topic area per
week or two
 Gain experience building some sort of data management
“engine” and experimentally validating it – this is a systems
course
 At end: you should be equipped to do research in this field,
or apply ideas from data management to your field
 Prerequisites:
 Strong undergraduate DB course, or CIS 550
 SQL, data modeling, basics of query optimization and execution, ACID, …
 Strong Java coding abilities
3
Grading
Summaries/commentary on papers (20%)
“Midterm report” (25%)
 Take one of the topics we’ve discussed and write a summary
and synthesis paper
 Graded for organization, clarity, grammar, etc. as well as
content
Project (50%) – team or individual:
 One focus: a SIGMOD demo to build a “smart research
environment” – instrumented machines, labs, building (more
shortly)
 Implementation
 Experimentation / validation
 Project report (should be in the style of a research paper)
 Brief (~15-minute) presentation for each group / project
Participation, discussion, intangibles (5%)
4
Potential Projects
 “Smart CIS”: integrate data from sensors, machines,
power monitoring, calendars, etc. to build a queriable
building, labs, machines
 Goal: demo at SIGMOD, possibly some research papers!
 Cloud computing: adapt a query processor to run on
Hadoop
 Sensors: build a real app with Crossbow motes
 Data visualizer: help understand and manipulate data
 Transformation reverse engineering: create data
instances to determine what a Perl or other tool is doing
when converting from one format to another
5
So What Is This Course About?
Not how to build an Oracle-driven Web site…
… nor even how to build Oracle…
6
What Is Unique about Data
Management?
 It’s been said that databases and data
management focus on scalability to huge
volumes of data
 What is it that makes this possible – and what
makes the work interesting if NOT at huge
scale?
7
The Key Principle: Data Independence
 Most methods of programming don’t separate
the logical and physical representations of data
 The data structures, access methods, etc. are all
given via interfaces!
 The relational data model was the first model
for data that is independent of its data
structures and implementation
8
What Is Data Independence?
 Codd points out that previous methods had:
 Order dependence
 Index dependence
 Access path dependence
 Still true in today’s Java/C#: what is the
drawback?
 What might you be able to do in removing
those?
9
The Relational Data Model
More than just tables!
 True relations: sets of tuples
 The only data representation a user/programmer “sees”
 Explicit encoding of everything in values
General and universal means of encoding everything!
 Connections are explicitly represented as values
 All semantics are pushed to queries
Additional integrity constraints
 Key constraints, functional dependencies, …
A secondary concept: views
 Define derived relations that are always “live”
 A way of encapsulating, abstracting, protecting, integrating data
10
Constraints and Normalization
 Fundamental idea: we don’t want to build semantics
into the data model, but we want to be able to encode
certain constraints
 Functional dependencies, key constraints, foreign-key
constraints, multivalued dependencies, join dependencies, etc.
 Allows limited data validation, plus opportunities for optimization
 The theory of normalization (see CSE 330, CIS 550)
makes use of known constraints
 Idea: eliminate redundancy, in order to maintain consistency in
the presence of updates
 (Note that there’s no reason for normalization of data in views!)
 Ergo, XML???
11
Relational Completeness
(Plus Extensions): Declarativity
What is special about relational query languages that
makes them amenable to scalability?
 Limited expressiveness – particularly when we consider
conjunctive queries (even with recursion)
 Guaranteed polytime execution in size of data
 Can reason about containment, invert them, etc.
 Equivalence between relational calculus and algebra
 Calculus  fully declarative, basis of query languages
 Algebra  imperative but polytime, basis of runtime systems
 Predictability of operations (in bulk)  cost models
 Ability to supplement data with auxiliary structures for
performance
 Interfaces to other “external” languages
12
Concurrency and Consistency
 Traditionally, DB efforts provide “ACID” properties:
 Atomicity, Consistency, Isolation, Durability
 Transaction : an atomic sequence of database actions
(read/write) on data items (e.g. calendar entry)
 Recoverability via a log: keeping track of all actions carried out
by the database
 But there’s a cost to all of this!
 How do distributed systems, Web services, serviceoriented architectures, and the like affect these
properties?
 We’ll consider one “relaxation” of these properties – the
MapReduce / BigTable style of computing
13
Other Data Models
 Concepts from the relational data model have
been adapted to form object-oriented data
models (with classes and subclasses), XML
models, etc.
 Doesn’t this result in some loss of logical-physical
independence?
14
What Is a Data Management System?
 Of course, there are traditional databases
 The focus of most work in the past 25 years
 “Tight loops” due to locally controlled data
 Indexing, transactions, concurrency, recovery,
optimization
 Also, today there are DB-like components in:
 Your email client and server (transactional storage)
 Enterprise Java Beans (distributed transactions)
 Google Base, BigTable, … (distributed indexing,
storage)
 But…
15
80% of the World’s Data is Not in
Databases!
Examples:
 Scientific data
(large images, complex programs that analyze the data)
 Personal data
 WWW and email
 Network traffic logs
 Sensor data, network router data, stream data, …
 Are there benefits to declarative techniques and data
independence in tackling these issues?
 Need to deal with data we don’t control and can’t guarantee
consistency over
 In recent years: increasing connection between databases, data
integration, information retrieval, information extraction,
sensors…
16
Some Questions We’ll Consider
 What are the “right” architectures for data
sharing? How do they change as consistency
needs (or other requirements) change?
 How much can we abstract away heterogeneity,
physical properties, etc.?
 How do we get good performance from
declarative queries?
17
Some Classes of Systems
We’ll Consider
Databases
How do we optimize and execute queries or ensure ACID?
Data integration
How do we handle heterogeneity in data and meaning?
Data streams and sensor data
How do we process infinite amounts of data?
Cloud computing, Web search
How do we partition computation along 1000s of machines and
achieve reliable execution?
Peer-to-peer architectures
What’s the best way of finding data?
18
Our Agenda this Semester
 Reading the canonical papers in the data
management literature, starting with databases
and later going to other data management
systems
 Some are very systems-y
 Some are very experimental
 Some are highly algorithmic, complexity-oriented
 Gaining an understanding of the principles of
building systems to handle declarative queries
over large volumes of data
19
Recap: Query Answering in a Data
Management System
 Based on declarative query languages
 Based on restricted first-order logic expressions over relations
 Not procedural – defines constraints on the output
 Converted into a query plan that exploits properties; run over the
data by the query optimizer and query execution engine
 Data may be local or remote
 Data may be heterogeneous or homogeneous
 Data sources may have different interfaces, access methods, etc.
 Most common query languages:
 SQL (based on tuple relational calculus)
 Datalog (based on domain relational calculus, plus fixpoint)
 XQuery (functional language; has an XML calculus core)
20
Recap: Layers of a Typical
Data Management System
API/GUI
(Simplification!)
Query
Optimizer
Stats
Physical plan
Exec. Engine
Catalog
Schemas Data/etc
Logging, recovery
Requests
Access Methods
Data/etc
Requests
Buffer Mgr
Pages
Pages
Physical retrieval
Data
Red = logical
Blue = physical
Requests
Source
21
Processing the Query
Web Server /
UI / etc
Hash
STUDENT
Optimizer
Takes
by cid
Execution
Engine
Merge
COURSE
by cid
Storage
Subsystem
SELECT *
FROM STUDENT, Takes, COURSE
WHERE STUDENT.sid = Takes.sID
AND Takes.cID = cid
22
DBMSs in the Real World
 Big, mature relational databases
 IBM, Oracle, Microsoft
 “Middleware” above these
 SAP, PeopleSoft, dozens of special-purpose apps
 “Application servers”
 Integration and warehousing systems
 DB2 Integrator, Oracle Fusion, …
 Current trends:
 Web services; XML everywhere
 Smarter, self-tuning systems (AutoAdmin, …)
 Stream systems (Vertica, Microsoft, IBM)
23
For Next Time
 Skim Codd if you haven’t already
 Read the overview papers of the two first
database systems:
 Astrahan et al., pp. 117 Wong et al. (skip Section 2; focus on pp. 200-)
 Write a summary of your assigned paper and
post to the Google Group: [email protected]
 Key question: how well did this system mesh with
Codd’s relational model? (You may need to skim
through other aspects of your assigned paper to help
answer that question)
24