Post on 23-Mar-2016
description
CS345: Advanced Databases
Chris Ré
What this course isDatabase fundamentals:– Theory– Old Crusty, Good SQL stuff– No/New/Not-Yet SQL
New stuff: Knowledge bases & Inference
Databases is a strange and beautiful area: Theory, Algorithms, Systems, & Applications
It’s a bit scattered, and I love it.
A Brief, BiasedDatabase History
Three Turing Award Winners
Charles Bachmann
Edgar Codd
JimGray
Seminal contributions made in Industry
The Birth of the Relational Model(1971)
database: a handful of relations (tables) with fixed schema.
WorksIn(Employee,Dept)
Query with small # of operations:Selection (filter),
Projection, Join, Union.
Basically, an operational finite model theory.
Data and Query ModelR(A,B) = { (a1,b2),…,(an,bn) }S(B,C,D) = { (b’1,c1,d1),…,(b’m,cm,dm) }
PA(R) ={ a : exists b. (a,b) in R } Projection
SelectionsF(R) ={ (a,b) : F( (a,b) ) for t in R }
F : D(R) -> {True, False}
Join(R,S) = { (a,b,c,d) : (a,b) in R & (b,c,d) in S} Join
Data
Key idea of the Relational Model
Declarative User says what they want---
not how to get it.
Key question: Can one implement the Relational
Model efficiently?
System R
In,1974 System R shows possible to get good performance.
1st Implementation of SQL.
IBM didn’t Push it,worried about IMS cannibalization, but…
Pat Selinger
Others Come on to the Scene…
Larry Ellison hears about IBM’s Research prototype and founds a company….
Fast Forward to TodayRelational model is dominate model of
data.
Takeaways about Database Research
Started with mathematical elegance and with close ties to industry.
Improve runtime performance as a proxy to increase programmer
productivity.
The Big Ideas
Independence
Declarative languages can improve productivity– Different team members work
independently• Backend, Storage, UI, BI, Etc.
– Transactional model.– Challenge: Support efficient concurrent
access?
Performance
Parallel programming is hard; SQL is most popular parallel programming language.– How do you deal with asymmetry of
memory hierarchy (Disk/MM/Cache)? – How do you structure parallel
optimization?– Concurrency?
Manageability
Systems live over time, and the system should automate many routine tasks.–Maintain derived data products (views)– Self-monitoring systems (autonomic)
Course Topics
A user says what they want—not how to get it.
Topic 1: QP FundamentalsQuery Processing Fundamentals
1. Empirical Join evaluation from 70s!2. System R: The Archetype (Cardinalityw)3. Formal Query Languages4. Acyclic Query Evaluation (Structure)5. Worst-case Optimal Join Algorithms (S
+ C)This will be the most
formal part of the course.
Analyzing your data before it was big (when it was just very large…)
Topic 2: OLAP-Style Analytics
Building new and old data systems:1. Theory of Materialized View2. Gamma (Parallel DBs) 3. MapReduce & the Rise of NoSQL
(2000s)4. NewSQL & Optimizing Joins on MR
(theory)5. Fagin’s Algorithm (theory)6. Statistical Analytic Systems
My biased view of the future…
Topic 3: Next-Generation Systems
1. Information Extraction2. Probabilistic Query Evaluation
(Theory)3. Scalable Inference4. Knowledge Bases
Transactions.
Topic 4: OLTP StyleTransactional Systems1. The rise of Key-Value Stores2. The case for determinism3. CALM & CAPs 4. The Return of Main Memory DBs.5. Spanner, F1, and Data Centers
Course Logistics
Grading• Course Project (More next)– Do something interesting with data.– Teams OK– Form teams soon and email me by Jan
12.
• Midterm Exam
Projects in each topic1. Knowledgebase Construction– Pick a domain and build a KBC system for it with
DeepDive
2. Join Algorithms– Certificate versions (see me)– MapReduce? GraphLab? Spark?
3. Analytics Systems
4. Transactional Systems.
You are free to choose other
projects
Datasets• Snapshot of the web marked up with NLP tools
and structured data (KBP and KBA challenges)
• 500k+ docs used by PaleoBiologists and structured data.
• We can mark up even more stuff.
• Benchmark ML, graphs if you want to work on analytics or join evaluation.
Wednesday
• Wednesday we begin the ancient art of join evaluation. All who pass this way must pass through this ancient topic!
• Read: Shapiro.– not too carefully, we’ll go through
details