+

Under the hood: Query Optimization, Query

Execution plans
+
Relational algebra, SQL, and
the DBMS?
Optimized
Relational
Relational Query Executable
algebra
SQL algebra execution code
expression
expression plan
Code
parser
generator

Query optimizer

DBMS
+
Introduction
 Alternative ways of evaluating a given query
 Equivalent expressions
 Different algorithms for each operation

 Cost difference between a good and a bad way of

evaluating a query can be enormous
 Need to estimate the cost of operations
 Statistical information about relations. Examples:
 number of tuples,
 number of distinct values for an attributes, etc.
 Statistics estimation for intermediate results
 to compute cost of complex expressions
+
Introduction (Cont.)
 Relations generated by two equivalent expressions
have the same set of attributes and contain the same
set of tuples
 although their tuples/attributes may be ordered differently.
+ Introduction (Cont.)
 Generation of query-evaluation plans for an
expression involves several steps:
1. Generating logically equivalent expressions using
equivalence rules. (see Apr 1 handout)
2. Annotating resultant expressions to get alternative
query plans
3. Choosing the cheapest plan based on estimated
cost

 The overall process is called cost based

optimization.
+
Query Representation
 Query tree
 Tree data structure that corresponds to a relational algebra
expression
 Input relations of the query as leaf nodes of the tree
 The relational algebra operations as internal nodes

 An execution of the query tree consists of

 executing an internal node operation whenever its operands
are available
 replacing that internal node by the relation that results from
executing the operation

 There are many trees for the same query

 Trees always have a strict order among their operations
 Query optimization must find “best” order
+
Sample Query
Example: For every project located in‘Stafford’,
retrieve the project number, the controlling
department number and the department
manager’s last name, address and birthdate

SELECT [Link],[Link],[Link], [Link], [Link]

FROM PROJECT AS P,DEPARTMENT AS D, EMPLOYEE AS E
WHERE [Link]=[Link] AND [Link]=[Link] AND
[Link]=‘STAFFORD’;

Relational algebra:
PNUMBER, DNUM, LNAME, ADDRESS, BDATE (((PLOCATION=‘Stafford’(PROJECT))
DNUM=DNUMBER (DEPARTMENT)) MGRSSN=SSN
(EMPLOYEE))
 Internal nodes are
- Executed when
inputs are ready
- Replaced by results

Internal nodes

Input relations
Different representation for the same algebra expression
… assumed to be the initial form

There are many trees for the same query

- strict order among their operations
- Query optimization must find “best” order
+
Transformation of Relational
Expressions
 Two relational algebra expressions are said to be
equivalent if on every legal database instance the two
expressions generate the same set of tuples
 Note: order of tuples is irrelevant

 In SQL, inputs and outputs are multisets of tuples

 Multiset (or bag) is a generalization of the notion of set in which
members are allowed to appear more than once
 Two expressions in the multiset version of the relational algebra
are said to be equivalent if on every legal database instance
the two expressions generate the same multiset of tuples

 An equivalence rule says that expressions of two forms

are equivalent
 Can replace expression of first form by second, or vice versa
+
Equivalence Rules
1. Conjunctive selection operations can be
deconstructed into a sequence of individual
selections.

2. Selection operations are commutative.

3. Only the last in a sequence of projection operations is

needed, the others can be omitted.
 L1 ( L2 ( ( Ln ( E )) ))  L1 ( E )

4. Selections can be combined with Cartesian products

and theta joins.
a. (E1 X E2) = E1  E2
b. 1(E1 2 E2) = E 1 1 2 E2
+
Equivalence Rules (Cont.)
5. Theta-join operations (and natural joins) are
commutative.
E1  E2 = E2  E1

6. (a) Natural join operations are associative:

(E1 E 2) E3 = E1 (E2 E 3)

(b) Theta joins are associative in the following

manner:

(E1 1 E 2) 2 3 E3 = E1 1 3 (E2 2 E 3)

where 2 involves attributes from only E2 and E3.

+ Pictorial Depiction of Equivalence
Rules
+
Equivalence Rules (Cont.)
7. The selection operation distributes over the theta
join operation under the following two conditions:
(a) When all the attributes in 0 involve only the attributes
of one
of the expressions (E1) being joined.

0E1  E2) = (0(E1))  E2

(b) When  1 involves only the attributes of E1 and 2

involves
only the attributes of E2.

1 E1  E2) = (1(E1))  ( (E2))

+ Equivalence Rules (Cont.)
[Link] projections operation distributes over the theta join
operation as follows:

(a) if  involves only attributes from L1  L2:

L1 L2 ( E1  E2 )  ( L1 ( E1 ))  ( L2 ( E2 ))

(b) Consider a join E1  E 2.

 Let L1 and L2 be sets of attributes from E1 and E2, respectively.
 Let L3 be attributes of E1 that are involved in join condition ,
but are not in L1  L2, and
 let L4 be attributes of E2 that are involved in join condition ,
but are notL1in
L2L(1E
1 L2. E2 )
 L1 L2 (( L1 L3 ( E1 ))  ( L2 L4 ( E2 )))
+
Equivalence Rules (Cont.)
9. The set operations union and intersection are commutative
E1  E2 = E2  E1
E1  E2 = E2  E1
9. (set difference is not commutative).

10. Set union and intersection are associative.

(E1  E2)  E3 = E1  (E2  E3)

(E1  E2)  E3 = E1  (E2  E3)

9. The selection operation distributes over ,  and –.

 (E1 – E2) =  (E1) – (E2)
and similarly for  and  in place of –
Also:  (E1 – E2) = (E1) – E2
and similarly for  in place of –, but not for 

[Link] projection operation distributes over union

L(E1  E2) = (L(E1))  (L(E2))

+
Housekeeping (Apr 3)

 Assignment 4 marking
 Project requirements and QA
+
Schema for examples

branch(branch_name, branch_city, assets)

customer (customer_name, customer street, customer
city)
loan (loan_number, branch name, amount)
borrower (customer_name, loan_number)
account (account_number, branch name, balance)
depositor (customer_ name, account_ number)
+
Transformation Example
 Query: Find the names of all customers who have an
account at some branch located in Brooklyn.
customer_name(branch_city = “Brooklyn”
(branch (account depositor)))

 Transformation using rule 7a.

customer_name
((branch_city =“Brooklyn” (branch))
(account depositor))
 Performing the selection as early as possible reduces
the size of the relation to be joined.
+ Example with Multiple
Transformations
 Query: Find the names of all customers with an account
at a Brooklyn branch whose account balance is over
$1000.
customer_name((branch_city = “Brooklyn”  balance > 1000
(branch (account depositor)))

 Transformation using join associatively (Rule 6a):

customer_name((branch_city = “Brooklyn”  balance > 1000
(branch account)) depositor)

 Second form provides an opportunity to apply the

“perform selections early” rule, resulting in the
subexpression

branch_city = “Brooklyn” (branch)  balance > 1000 (account)

 Thus a sequence of transformations can be useful

+
Multiple Transformations (Cont.)
+
Projection Operation Example
customer_name((branch_city = “Brooklyn” (branch) account) depositor)

 When we compute

(branch_city = “Brooklyn” (branch) account )

we obtain a relation whose schema is:

(branch_name, branch_city, assets, account_number, balance)
 Push projections using equivalence rules 8a and 8b; eliminate unneeded
attributes from intermediate results to get:
customer_name ((
account_number ( (branch_city = “Brooklyn” (branch) account ))
depositor )
 Performing the projection as early as possible reduces the size of the
relation to be joined.
+
Join Ordering Example
 For all relations r1, r2, and r3,

(r1 r2) r3 = r1 (r2 r3 )

 If r2 r3 is quite large and r1 r2 is small, we choose

(r1 r2) r3

so that we compute and store a smaller temporary

relation.
+
Join Ordering Example (Cont.)
 Consider the expression

customer_name ((branch_city = “Brooklyn” (branch))

(account depositor))
 Could compute account depositor first, and join
result with
branch_city = “Brooklyn” (branch)
but account depositor is likely to be a large relation.
 Only a small fraction of the bank’s customers are likely
to have accounts in branches located in Brooklyn
 it is better to compute

branch_city = “Brooklyn” (branch) account

first.
+ Enumeration of Equivalent
Expressions
 Query optimizers use equivalence rules to systematically
generate expressions equivalent to the given expression
 Conceptually, generate all equivalent expressions by
repeatedly executing the following step until no more
expressions can be found:
 for each expression found so far, use all applicable equivalence
rules
 add newly generated expressions to the set of expressions
found so far
 The above approach is very expensive in space and time
 Space requirements reduced by sharing common
subexpressions:
 when E1 is generated from E2 by an equivalence rule, usually only
the top level of the two are different, subtrees below are the same
and can be shared
 E.g. when applying join associativity

 Time requirements are reduced by not generating all

expressions
+
Cost Estimates in Query
Optimization
 Cost Components for Query Execution
 Access cost to secondary storage
 Searching, reading, writing, updating, etc …
 Memory usage cost
 Number of memory buffers needed for the query
 Storage cost
 Storing any intermediate files that are generated by an
execution strategy for the query
 Communication cost
 Shipping the results from the database site to the user’s
site
 Computation cost
 Of performing in-memory operations on the data buffers
during the execution plan (searching, sorting, joining,
arithmetic)
+
Cost Estimation

 Cost of each operator computer

 Need statistics of input relations
 E.g. number of tuples, sizes of tuples

 Inputs can be results of sub-expressions

 Need to estimate statistics of expression results
 To do so, we require additional statistics
 E.g. number of distinct values for an attribute
+
Evaluation Plan
 An evaluation plan defines exactly what algorithm is used for each
operation, and how the execution of the operations is coordinated.
+ Query Optimization
 Multi-operator Queries: Pipelined Evaluation
 On-the-fly: The result of one operator is pipelined to
another operator without creating a temporary table to
hold intermediate result, called on-the-fly.
 Materialized : Otherwise, intermediate results must be
materialized.

A B
+
Choice of Evaluation Plans
 Must consider the interaction of evaluation techniques
when choosing evaluation plans: choosing the cheapest
algorithm for each operation independently may not yield
best overall algorithm. E.g.
 merge-join may be costlier than hash-join, but may provide a
sorted output which reduces the cost for an outer level
aggregation.
 nested-loop join may provide opportunity for pipelining

 Practical query optimizers incorporate elements of the

following two broad approaches:
[Link] all the plans and choose the best plan in a
cost-based fashion. (slides at back of deck if interested)
2. Uses heuristics to choose a plan.
+
Heuristic Optimization
 Cost-based optimization is expensive, even with dynamic
programming.
 Systems may use heuristics to reduce the number of
choices that must be made in a cost-based fashion.
 Heuristic optimization transforms the query-tree by using a
set of rules that typically (but not in all cases) improve
execution performance:
 Perform selection early (reduces the number of tuples)
 Perform projection early (reduces the number of attributes)
 Perform most restrictive selection and join operations before
other similar operations.
 Some systems use only heuristics, others combine heuristics
with partial cost-based optimization.
+
Steps in Typical Heuristic
Optimization
1. Deconstruct conjunctive selections into a
sequence of single selection operations (Equiv. rule 1.).

2. Move selection operations down the query tree for

the earliest possible execution (Equiv. rules 2, 7a, 7b, 11).

3. Execute first those selection and join operations

that will produce the smallest relations (Equiv. rule 6).

4. Replace Cartesian product operations that are

followed by a selection condition by join operations
(Equiv. rule 4a).

5. Deconstruct and move as far down the tree as

possible lists of projection attributes, creating new
projections where needed (Equiv. rules 3, 8a, 8b, 12).

6. Identify those subtrees whose operations can be

pipelined, and execute them using pipelining).
+
Additional slides
Structure of Query Optimizers
More details about cost estimation
+
Structure of Query Optimizers

 The System R/Starburst optimizer considers only left-deep

join orders. This reduces optimization complexity and
generates plans amenable to pipelined evaluation.
System R/Starburst also uses heuristics to push selections
and projections down the query tree.
 Heuristic optimization used in some versions of Oracle:
 Repeatedly pick “best” relation to join next
 Starting from each of n starting points. Pick best among
these.
 For scans using secondary indices, some optimizers take
into account the probability that the page containing the
tuple is in the buffer.
 Intricacies of SQL complicate query optimization
 E.g. nested subqueries
+
Structure of Query Optimizers
(Cont.)
 Some query optimizers integrate heuristic selection
and the generation of alternative access plans.
 System R and Starburst use a hierarchical procedure
based on the nested-block concept of SQL: heuristic
rewriting followed by cost-based join-order optimization.

 Even with the use of heuristics, cost-based query

optimization imposes a substantial overhead.
 This expense is usually more than offset by savings at
query-execution time, particularly by reducing the
number of slow disk accesses.
+ Statistical Information for Cost
Estimation
 nr: number of tuples in a relation r.
 br: number of blocks containing tuples of r.
 lr: size of a tuple of r.
 fr: blocking factor of r — i.e., the number of tuples of r that fit into one
block.
 V(A, r): number of distinct values that appear in r for attribute A; same as
the size of A(r).

 If tuples of r are stored together physically in a file, then:

+
Histograms

 Histogram on attribute age of relation person

 Equi-width histograms
 Equi-depth histograms
Selection Size Estimation
 A=v(r)
 nr / V(A,r) : number of records that will satisfy the selection
 Equality condition on a key attribute: size estimate = 1

 AV(r) (case of A  V(r) is symmetric)

 Let c denote the estimated number of tuples satisfying the
condition.
 If min(A,r) and max(A,r) are available in catalog
 c = 0 if v < min(A,r)

 c=

 If histograms available, can refine above estimate

 In absence of statistical information c is assumed to be nr / 2.
+ Size Estimation of Complex
Selections
 The selectivity of a condition i is the probability that a tuple in
the relation r satisfies i .
 If si is the number of satisfying tuples in r, the selectivity of i is
given by si /nr.

 Conjunction: 1 2. . .  n (r). Assuming indepdence, estimate of

tuples in the result is:

 Disjunction:1 2 . . .  n (r). Estimated number of tuples:

 Negation: (r). Estimated number of tuples:

nr – size((r))
+
Join Operation: Running
Example
Running example:
depositor customer

Catalog information for join examples:

 ncustomer = 10,000.
 fcustomer = 25, which implies that
bcustomer =10000/25 = 400.

 ndepositor = 5000.
 fdepositor = 50, which implies that
bdepositor = 5000/50 = 100.
 V(customer_name, depositor) = 2500, which implies that , on
average, each customer has two accounts.
 Also assume that customer_name in depositor is a foreign key on
customer.
 V(customer_name, customer) = 10000 (primary key!)
+
Estimation of the Size of Joins
 The Cartesian product r x s contains nr .ns tuples; each tuple
occupies sr + ss bytes.

 If R  S = , then r s is the same as r x s.

 If R  S is a key for R, then a tuple of s will join with at most

one tuple from r
 therefore, the number of tuples in r s is no greater than the
number of tuples in s.

 If R  S in S is a foreign key in S referencing R, then the

number of tuples in r s is exactly the same as the number of
tuples in s.
 The case for R  S being a foreign key referencing S is symmetric.

 In the example query depositor customer, customer_name in

depositor is a foreign key of customer
 hence, the result has exactly ndepositor tuples, which is 5000
+ Estimation of the Size of Joins
(Cont.)
 If R  S = {A} is not a key for R or S.
If we assume that every tuple t in R produces tuples in R
S, the number of tuples in R S is estimated to be:

If the reverse is true, the estimate obtained will be:

The lower of these two estimates is probably the more

accurate one.
 Can improve on above if histograms are available
 Use formula similar to above, for each cell of histograms on the
two relations
+ Estimation of the Size of Joins
(Cont.)

 Compute the size estimates for depositor customer

without using information about foreign keys:
 V(customer_name, depositor) = 2500, and
V(customer_name, customer) = 10000
 The two estimates are 5000 * 10000/2500 - 20,000 and
5000 * 10000/10000 = 5000
 We choose the lower estimate, which in this case, is the
same as our earlier computation using foreign keys.
+
Size Estimation for Other
Operations
 Projection: estimated size of A(r) = V(A,r)

 Aggregation : estimated size of AgF(r) = V(A,r)

 Set operations
 For unions/intersections of selections on the same relation:
rewrite and use size estimate for selections
 E.g. 1 (r)  2 (r) can be rewritten as 1 2 (r)
 For operations on different relations:
 estimated size of r  s = size of r + size of s.
 estimated size of r  s = minimum size of r and size of s.
 estimated size of r – s = r.
 All the three estimates may be quite inaccurate, but provide
upper bounds on the sizes.
+
Size Estimation (Cont.)

 Outer join:
 Estimated size of r s = size of r s + size of r
 Case of right outer join is symmetric
 Estimated size of r s = size of r s + size of r + size
of s
+ Estimation of Number of Distinct
Values
Selections:  (r)

 If  forces A to take a specified value: V(A, (r)) = 1.

 e.g., A = 3

 If  forces A to take on one of a specified set of values:

V(A, (r)) = number of specified values.
 (e.g., (A = 1 V A = 3 V A = 4 )),

 If the selection condition  is of the form A op r

estimated V(A, (r)) = V(A.r) * s
 where s is the selectivity of the selection.

 In all the other cases: use approximate estimate of

min(V(A,r), n (r) )
 More accurate estimate can be got using probability
theory, but this one works fine generally
+
Estimation of Distinct Values
(Cont.)
Joins: r s
 If all attributes in A are from r
estimated V(A, r s) = min (V(A,r), n r s )

 If A contains attributes A1 from r and A2 from s, then

estimated
V(A,r s) =

min(V(A1,r)*V(A2 – A1,s), V(A1 – A2,r)*V(A2,s),

nr s )
 More accurate estimate can be got using probability
theory, but this one works fine generally
+
Estimation of Distinct Values
(Cont.)
 Estimation of distinct values are straightforward for
projections.
 They are the same in A (r) as in r.

 The same holds for grouping attributes of aggregation.

 For aggregated values
 For min(A) and max(A), the number of distinct values can be
estimated as min(V(A,r), V(G,r)) where G denotes grouping
attributes
 For other aggregates, assume all values are distinct, and use
V(G,r)
+
Optimizing Nested
Subqueries**
SQL conceptually treats nested subqueries in the where clause

as functions that take parameters and return a single value or
set of values
 Parameters are variables from outer level query that are used in the
nested subquery; such variables are called correlation variables

 E.g.
select customer_name
from borrower
where exists (select *
from depositor
where depositor.customer_name =

borrower.customer_name)
 Conceptually, nested subquery is executed once for each tuple
in the cross-product generated by the outer level from clause
 Such evaluation is called correlated evaluation
 Note: other conditions in where clause may be used to compute a
join (instead of a cross-product) before executing the nested
subquery
+ Optimizing Nested Subqueries
(Cont.)
 Correlated evaluation may be quite inefficient since
 a large number of calls may be made to the nested query
 there may be unnecessary random I/O as a result

 SQL optimizers attempt to transform nested subqueries to

joins where possible, enabling use of efficient join techniques
 E.g.: earlier nested query can be rewritten as
select customer_name
from borrower, depositor
where depositor.customer_name = borrower.customer_name
 Note: above query doesn’t correctly deal with duplicates, can be
modified to do so as we will see

 In general, it is not possible/straightforward to move the entire

nested subquery from clause into the outer level query from
clause
 A temporary relation is created instead, and used in body of outer
level query
+ Optimizing Nested Subqueries
(Cont.)
In general, SQL queries of the form below can be rewritten as
shown
 Rewrite: select …
from L1
where P1 and exists (select *
from L2
where P2)
 To: create table t1 as
select distinct V
from L2
where P21
select …
from L1, t1
where P1 and P22
 P21 contains predicates in P2 that do not involve any correlation
variables
 P22 reintroduces predicates involving correlation variables, with
relations renamed appropriately
 V contains all attributes used in predicates with correlation
variables
+
Optimizing Nested Subqueries
(Cont.)
 In our example, the original nested query would be transformed
to
create table t1 as
select distinct customer_name
from depositor
select customer_name
from borrower, t1
where t1.customer_name = borrower.customer_name
 The process of replacing a nested query by a query with a join
(possibly with a temporary relation) is called decorrelation.
 Decorrelation is more complicated when
 the nested subquery uses aggregation, or
 when the result of the nested subquery is used to test for equality, or
 when the condition linking the nested subquery to the other
query is not exists,
 and so on.
+
Cost-Based Optimization

 Consider finding the best join-order for r1 r2 . . . rn.

 There are (2(n – 1))!/(n – 1)! different join orders for

above expression. With n = 7, the number is 665280,
with n = 10, the number is greater than 176 billion!
 No need to generate all the join orders. Using dynamic
programming, the least-cost join order for any subset
of
{r1, r2, . . . rn} is computed only once and stored for
future use.
+Dynamic Programming in
Optimization
 To find best join tree for a set of n relations:
 To find best plan for a set S of n relations, consider all
possible plans of the form: S1 (S – S1) where S1 is any
non-empty subset of S.
 Recursively compute costs for joining subsets of S to find
the cost of each plan. Choose the cheapest of the 2n – 1
alternatives.
 When plan for any subset is computed, store it and reuse
it when it is required again, instead of re-computing it
 Dynamic programming [Link]
Dynamic_programming
+
Join Order Optimization
Algorithm
procedure findbestplan(S)
if (bestplan[S].cost  )
return bestplan[S]
// else bestplan[S] has not been computed earlier, compute it now
if (S contains only 1 relation)
set bestplan[S].plan and bestplan[S].cost based on the best
way
of accessing S

else for each non-empty subset S1 of S such that S1  S

P1= findbestplan(S1)
P2= findbestplan(S - S1)
A = best algorithm for joining results of P1 and P2
cost = [Link] + [Link] + cost of A
if cost < bestplan[S].cost
bestplan[S].cost = cost
bestplan[S].plan = “execute [Link]; execute
[Link];
join results of P1 and P2 using A”
return bestplan[S]
+
Left Deep Join Trees
 In left-deep join trees, the right-hand-side input
for each join is a relation, not the result of an
intermediate join.

See
[Link]
ees-and-bushy-trees-oh-my/
for a discussion of tree types and optimization
+ Cost of Optimization
 With dynamic programming, time complexity of optimization with bushy
trees is O(3n). (see [Link]
 With n = 10, this number is 59000 instead of 176 billion!
 time complexity of an algorithm quantifies the amount of time taken by an
algorithm to run as a function of the length of the string representing the input

 Space complexity is O(2n)

 space complexity = amount of memory an algorithm needs

 To find best left-deep join tree for a set of n relations:

 Consider n alternatives with one relation as right-hand side input and the other
relations as left-hand side input.
 Using (recursively computed and stored) least-cost join order for each
alternative on left-hand-side, choose the cheapest of the n alternatives.

 If only left-deep trees are considered, time complexity of finding best

join order is O(n 2n)
 Space complexity remains at O(2n)

 Cost-based optimization is expensive, but worthwhile for queries on

large datasets (typical queries have small n, generally < 10)
+Interesting Orders in Cost-Based Optimization
 Consider the expression (r1 r2 r3) r4 r5
 An interesting sort order is a particular sort order of
tuples that could be useful for a later operation.
 Generating the result of r1 r2 r3 sorted on the attributes
common with r4 or r5 may be useful, but generating it
sorted on the attributes common only r1 and r2 is not useful.
 Using merge-join to compute r1 r2 r3 may be costlier, but
may provide an output sorted in an interesting order.

 Not sufficient to find the best join order for each subset
of the set of n given relations; must find the best join
order for each subset, for each interesting sort order
 Simple extension of earlier dynamic programming
algorithms
 Usually, number of interesting orders is quite small and
doesn’t affect time/space complexity significantly