Modern Database Systems Lecture 1 Aristides Gionis Michael Mathioudakis T.A.: Orestis Kostakis Spring 2016
Modern Database Systems Lecture 1
Aristides Gionis Michael Mathioudakis
T.A.: Orestis Kostakis
Spring 2016
logistics
assignment will be up by Monday (you will receive email)
due Feb 12th
if you’re not registered... I will post material (slides and assignments) also at
http://michalis.co/moderndb/
2
in this lecture...
review past material relational model and sql
storage and indexing access cost analysis
hash index b+ tree
3
relational model and SQL
relational model and sql
what is the relational model? tabular representation of data
why do we study it?
supports simple and intuitive querying good for educational purposes
most widely used
5
definitions relational database
a set of relations
relation
example!
schema name of relation + name and
type of each field fields as columns
instance a table with rows and columns
6
example relation: students
cardinality (number of rows) = 3, degree (number of fields/columns) = 5 > can we have the same value twice in the same column? schema students(sid: integer, name: string, username: string, age: integer, gpa: real)
sid name username age gpa
53666 Sam Jones jones 22 3.4
53688 Alice Smith smith 22 3.8
53650 Jon Edwards jon 23 2.4
7
querying major strength of relational model
simple, intuitive, precise querying of data
the DBMS is responsible for efficient evaluation
Standard Query Language (SQL) the standard language for relational queries
developed by IBM in the 1970s was standardized in 1986
latest standard in 2011
example!
8
example SQL query
to find student records of age 23 SELECT * FROM students WHERE age=23
to find just names and usernames SELECT name, username FROM students WHERE age=23
sid name username age gpa
53666 Kate Jones jones 22 3.4
53688 Alice Smith smith 22 3.8
53650 Jon Edward jon 23 2.4
sid name username age gpa
53650 Jon Edward jon 23 2.4
name username
Jon Edward jon
9
creating, altering, and destroying, relations in SQL
CREATE TABLE students (sid CHAR(20), name CHAR(20), username CHAR(10), age INTEGER, gpa REAL);
the type of each column is enforced by the DBMS
DROP TABLE students;
ALTER TABLE students ADD COLUMN firstYear integer;
every tuple in the current instance is extended with a null value in the new column
CREATE TABLE course (sid CHAR(20), points integer, grade CHAR(2));
destroy relation students (schema and instance) 10
adding and deleting tuples > what do the following statements do?
INSERT INTO students(sid, name, username, age, gpa) VALUES (12345, “Kate Doe”, “kate”, 23, 4.0);
DELETE FROM students WHERE name = ‘Jane Smith’;
11
candidate keys a set of fields is a candidate key (aka ‘key’) for a relation if... 1) distinct tuples cannot have same values in all key fields, and
2) this is not true for any subset of the key
if only part (1) from above is true... we have a superkey
possibly many candidate keys for a relation DBMS admin chooses one (1) of them as primary key
an integrity constraint
condition must be true for any instance of the database other integrity constraints?
12
candidate keys
in SQL, use PRIMARY KEY to specify primary key UNIQUE to specify candidate keys
example
relation enrolled holds information about student enrollment to courses compare the following ‘create table’ statements
use ICs carefully - they might forbid database instances that could arise in practice
CREATE TABLE Enrolled (sid CHAR(20), cid CHAR(20), grade CHAR(2), PRIMARY KEY (sid,cid))
CREATE TABLE Enrolled (sid CHAR(20) cid CHAR(20), grade CHAR(2), PRIMARY KEY (sid), UNIQUE (cid, grade)) 13
storage and indexing
14
storage
setting the DBMS uses disks as external storage to store relations into files of records
disks retrieve random page at fixed cost cheaper to retrieve several consecutive pages than each by random access
why?
file organization method of arranging a file of records on external storage
record: one row of a relation record is internally assigned a record id (rid)
rid is sufficient to physically locate record (address)
15
alternative file organizations heap files random order
suitable when typical access is a file scan to retrieve all records
sorted files records are sorted - typically by column value(s)
suitable if records must be retrieved by same order
indexes data structures that allow organized access to records…
... via search keys - typically column value(s) updates are faster than in sorted files -- why?
16
data structures that allow us to find rids of records with specified column values any subset of the columns of a relation can be the search key for an index search key is not same as primary / candidate key
indexes an index contains a collection of data entries supports efficient retrieval of data entries k*
with a given key value k
index entries
data entries
data records
index file
data file 17
types of data entries
three alternatives 1. data record with key value k
2. (k, rid of data record with search key k) 3. (k, list of rids of data records with search key k)
type of data entries is orthogonal to index structure
example of index structure B+ trees or hash tables
18
data entries of type 1 index structure is a file
organization for data records we just have an ‘index file’
index entries
data records
index file
> how many indexes of a relation can be of type 1?
19
types of data entries - types 2 & 3 data entries typically much smaller than data records
> why?
index entries
data entries
data records
index file
data file
type 3 is more compact than type 2 > why?
20
index classes primary vs secondary
primary: if search key contains a primary key unique index: search key contains a candidate key
clustered vs unclustered
if order of data records is same as that of data entries makes big difference for some queries!
> can alternative 1 indexes be unclustered?
unclustered clustered
21
hash-based indexes
retrieve records with exactly specified search-key values suitable for equality queries index is collection of buckets bucket = 1 or more disk pages
hashing function h
h(r) = bucket where record r belongs, based on its column values
data entries are … ... type 1: the buckets contain data records
... type 2 or 3: the buckets contain (key, rid) or (key, rids) pairs
22
hash-based indexes
Smith, 44, 3000
Jones, 40, 6003
Tracy, 44, 5004
Ashby, 25, 3000
Basu, 33, 4003
Kate, 29, 2007
Cass, 50, 5004
Basu, 33, 6003
age h1
relation employes(name CHAR(100), age INTEGER, salary INTEGER)
3000
3000
5004
5004
4003
2007
6003
6003
salary h2
clustered (type 1) hash index on age unclustered (type 2) hash index on salary 23
leaf pages contain data entries, and are chained (prev & next) non-leaf pages have index entries; only used to direct searches
P 0 K 1 P 1 K 2 P 2 K m P m
index entry
b+ tree indexes
non-leaf pages
leaf pages
(sorted by search key)
24
example b+ tree
find 28*? 29*? all > 15* and < 30*?
insert/delete find data entry in leaf, then update it
need to adjust parent sometimes change sometimes bubbles up the tree
2* 3*
root
17
30
14* 16* 33* 34* 38* 39*
13 5
7* 5* 8* 22* 24*
27
27* 29*
entries < 17 entries >= 17
note that data entries in leaf level are sorted
access-cost analysis
26
access-cost model ● relation students
○ B: number of data pages, R: number of records per page ● execute typical select-from-where query
○ D: (average) time to read or write one disk page
SELECT * FROM students WHERE <...>
● estimate running time of query
○ ignore cpu costs ○ number of disk accesses (read/writes) is the bottleneck
27
file organizations heap file (random order; inserts at eof)
sorted file, sorted on <age, gpa>
clustered B+ tree file (type 1 data entries) on
search key <age, gpa>
heap file with unclustered B+ tree index on search key <age, gpa>
heaf file with unclustered hash index on
search key <age, gpa>
28
queries to compare
insert record
SELECT * FROM students
SELECT * FROM students WHERE age = 22 and gpa = 4.0
SELECT * FROM student WHERE age >= 20
INSERT INTO STUDENTS (sid, name, username, age, gpa) VALUES (12345, “Michael”, “mike”, 32, 2.6)
scan - fetch all records
equality search
range search
29
cost analysis
what is the estimated time for each query to run?
under simplified model how many disk pages are accessed?
time = #disk-accesses x D
30
cost analysis
scan equality range insert
heap
sorted
clustered
unclustered b+ tree
unclustered hash
31
heap file
operation cost and explanation
scan B; simply retrieve all pages
equality search
B in worst case; if we know that exactly one such record exists, the cost is 0.5B in expectation
range search B; must retrieve all records
insert 2; fetch and store back the last page of the file
32
sorted file
operation cost and explanation
scan B; simply retrieve all pages
equality search log2B + #qualifying-pages; since the condition matches the index, we can find the page of the record with binary search that retrieves log2B pages; if more than one records qualify, retrieve sequentially #qualifying-pages after the first
range search log2B + #qualifying-pages; as above, log2B pages are retrieved to find the first matching record, followed possibly by a number (#qualifying-pages) of pages with qualifying records
insert log2B + B; find the position of the record in the file (log2B); then, read the second half of the file, insert the record, write the second half back (0.5B + 0.5B in expectation)
33
clustered b+ tree
operation cost and explanation
scan 1.5B; simply retrieve all record pages
equality search logF1.5B + #qualifying-pages; find the first qualifying record and retrieve consecutive qualifying ones
range search logF1.5B + #qualifying-pages; find the first qualifying record and retrieve consecutive qualifying ones
insert logF1.5B + 1; search for record page (logF1.5B) and add record to it (1)
assumptions: 2/3 = 67% occupancy of record pages, i.e. 1.5B record pages; fanout F
34
unclustered b+ tree
operation cost and explanation
scan B(R+0.15); scan the leaf level of the index (0.15B); for each data entry, fetch the page with the corresponding data record (6.7R x 0.15B = BR)
equality search logF0.15B + #qualifying-records; locate the first data entry (logF0.15B) and do one disk access for every qualifying record (#qualifying-records)
range search logF0.15B + #qualifying-records; locate the first data entry (logF0.15B) and do one disk access for every qualifying record (#qualifying-records)
insert 3 + logF0.15B;insert at end of heap file (2), find page for data entry (logF0.15B) and update it (1)
assumptions: the size of one data entry is 10% the size of one record; also, index pages have 2/3=67% occupancy; therefore, number of index leaf pages is 0.1*1.5B = 0.15B and number of data entries in one page are 10*0.67R = 6.7R
35
unclustered hash index
operation cost and explanation
scan B(R+0.125); retrieve pages that contain data entries (0.125B); for each data entry, fetch the page with the corresponding data record
equality search 2; retrieve page with data entry (1) and page with data record (1)
range search 0.125B + #qualifying-records; the hash index offers no help - scan index (0.125B) and retrieve pages of matching records; typically it’s better to scan entire heapfile (B)
insert 4; insert record into heap file (1 read+1 write); insert record into hash index (1 read + 1 write)
assumptions: the size of one data entry is 10% the size of one record; static hashing, no overflow pages (one bucket is one page); 4/5 = 80% occupancy; therefore , 0.1*1.25B = 0.125B pages for data entries and the number of data entries in a page is 10*0.8R = 8R
36
cost analysis scan equality range insert
heap B B B 2
sorted B log2B + #qualifying-pages
log2B + #qualifying-pages
log2B + B
clustered 1.5B logF1.5B + #qualifying-pages
logF1.5B + #qualifying-pages
logF1.5B + 1
unclustered b+ tree
B(R+0.15) logF0.15B + #qualifying-records
logF0.15B + #qualifying-records
3 + logF0.15B
unclustered hash
B(R+0.125) 2 0.125B + #qualifying-records
4
note we made several assumptions to obtain these numbers 37
the morale
different queries have different cost for different file organizations
> how would you use this analysis as a db admin?
discuss
38
the morale
know your workload what queries? how often?
on what relations? what file organizations? what indexes would speed-up response times for your workload?
hint: see WHERE clause for index key candidates
why?
what trade-offs will you face? hint: queries are faster but updates take time, index takes space
we’ll see more complex cases in ‘query optimization’
39
indexes with composite search keys
composite search keys search on a combination of fields
equality query
every field value is equal to a constant e.g., age=20 and sal =75, wrt <sal,age> index
range query
some field value is not a constant e.g., age =20; or age=20 and sal > 10, wrt <sal,age> index
data entries in index sorted by
search key to support range queries (e.g., b+ trees) <sal, age>
<age>
<sal>
data records sorted by name
data entries sorted by <sal,age>
data entries sorted by <sal>
examples of composite key indexes
11,80
12,10
12,20
13,75
10,12
20,12
75,13
80,11
11
12
12
13
10
20
75
80
name age sal
bob 12 10
cal 11 80
joe 12 20
sue 13 75
<age,sal>
remember also composite indexes are larger,
updated more often 40
composite search keys
if condition is: 3000<sal<5000: <age,sal> index does not help! why?
because the index does not match the selection condition
index matches selection (condition ∧ ... ∧ ... ∧ condition) when:for hash index: only equality conditions for all fields
for tree index: includes equality or range condition for a prefix of the search key
41
to retrieve employee records with age=30 AND sal=4000, an index on <age,sal> or <sal, age> would be better than
an index on <age> or an index on <sal>
if condition is: age=30 AND 3000<sal<5000: <age,sal> index much better than <sal,age> index! why?
hint: allows us to allocate answer with contiguous data entries order can make a difference depending on the selectivity of each condition
if condition is: 20<age<30 AND 3000<sal<5000: tree index on <age,sal> or <sal,age> make no difference
if selectivity of each condition is the same
composite search keys
42
index-only plans
some queries can be answered without retrieving any data records
if a suitable index is available
example employees
(name CHAR(100), depnum INTEGER, age INTEGER, salary INTEGER)
SELECT depnum, COUNT(*)FROM employeesGROUP BY depnum
SELECT AVG(salary)FROM employeesWHERE age=25 ANDsalary BETWEEN 3000 AND 5000
index on <depnum>
b+ tree index on <age,salary>
43
index-only plans are possible with both <dno,age> or <age,dno>
tree index <age, dno> is better
why?
SELECT E.dno, COUNT (*)
FROM Emp E
WHERE E.age=30
GROUP BY E.dno
index-only plans
44
summary
45
summary ● relational model and SQL
○ tabular representation ■ one record per row ■ schema determines names and types of columns
○ simple, intuitive querying language ■ statements to select records that satisfy a condition ■ specify columns to project ■ statements to insert and delete tuples
46
● storage ○ a DBMS might use different file organizations to store relations ○ heap file, sorted file, index ○ different queries have different access costs
for different file organizations ○ having the right index can make a big difference in execution time
● commonly used indexes ○ B+ tree and hash-based index
next
b+ trees and hash-based index
external sorting
joins
query optimization
47
references ● “cowbook”, database management systems, by ramakrishnan and gehrke ● “elmasri”, fundamentals of database systems, elmasri and navathe ● other database textbooks
● disk access analysis ○ cowbook, chapter 8
● b+ tree and hashing algorithms ○ elmasri
■ section 18.2: hash indexes ■ section 18.3.2: b+ trees
○ cowbook ■ chapters 10 and 11
48
credits
slides based on material from database management systems, by ramakrishnan and gehrke
49
joins sid name username age gpa
53666 Sam Jones jones 22 3.4
53688 Alice Smith smith 22 3.8
53650 Jon Edwards jon 23 2.4
students
sid points grade
53666 92 A
53688 35 D
53650 65 C
course
what does this compute?
SELECT S.name, C.grade FROM Students S,Course C WHERE S.sid = C.sid AND
C.points > 60
S.name C.grade
Sam Jones A
Jon Edwards C 50
index-only plans
SELECT E.dno, COUNT (*)
FROM Emp E
WHERE E.age>30
GROUP BY E.dno
what if we consider the second query? we’ll come back to this after external sorting