CAS CS 460/660

Introduction to Database
Systems

Query Evaluation II

1.1
 Cost-based Query Sub-System
Select *
Queries From Blah B
Where B.blah = blah

Query Parser

Query Optimizer

Plan Plan Cost Catalog Manager

Generator Estimator

Schema Statistics
Query Plan Evaluator
1.2
 Review - Relational Operations

 We will consider how to implement:

 Selection (  ) Selects a subset of rows from relation.
 Projection (  ) Deletes unwanted columns from relation.


Join ( ⋈ ) Allows us to combine two relations.
 Set-difference ( - ) Tuples in reln. 1, but not in reln. 2.
 Union (  ) Tuples in reln. 1 and in reln. 2.
 Also: Aggregation (SUM, MIN, etc.) and GROUP BY
 Since each op returns a relation, ops can be composed ! After we cover
the operations, we will discuss how to optimize queries formed by
composing them.

1.3
Selection (filter)
Operators

1.4
 Schema for Examples

Sailors (sid: integer, sname: string, rating: integer, age: real)

Reserves (sid: integer, bid: integer, day: date, rname: string)

 Similar to old schema; rname added for variation.

(assume pages of 4000 bytes each)

 Reserves:
 Each tuple is 40 bytes long, 100 tuples per page, 1000 pages. (100K reseravtions)
 Sailors:
 Each tuple is 50 bytes long, 80 tuples per page, 500 pages. (40K sailors)

1.5
 SELECT *
Simple Selections FROM Reserves R
WHERE R.date > ‘1/1/2015’

 Of the form
 Question: how best to perform? Depends on:
 what indexes/access paths are available
 what is the expected size of the result (in terms of number of tuples and/or
number of pages)
 Size of result approximated as

size of R * reduction factor

 “reduction factor” is usually called selectivity.
 estimate of reduction factors is based on statistics – we will discuss shortly.

1.6
 Alternatives for Simple Selections
 With no index, unsorted:
 Must essentially scan the whole relation
 cost is M (#pages in R). For “Reserves” = 1000 I/Os.
 With no index, sorted on day:
 cost of binary search + number of pages containing results.
 For reserves = 10 I/Os + selectivity*1000
 With an index on selection attribute:
 Use index to find qualifying data entries,
 then retrieve corresponding data records.
 (Hash index useful only for equality selections.)

1.7
 Using an Index for Selections

 Cost depends on #qualifying tuples, and clustering.

 Cost:

 finding qualifying data entries (typically small)

 plus cost of retrieving records (could be large w/o
clustering).
 In example “Reserves” relation, if 10% of tuples qualify (result size estimate:
100 pages, 10000 tuples).

 With a clustered index, cost is little more than 100 I/Os;

 if unclustered, could be more than 10000 I/Os! unless…

1.8
 Selections using Index (cont)
 Important refinement for unclustered indexes:

1. Find qualifying data entries.

2. Sort the rid’s of the data records to be retrieved.
3. Fetch rids in order. This ensures that each data page is
looked at just once (though # of such pages likely to be
higher than with clustering).

Index entries UNCLUSTERED

CLUSTERED direct search for
data entries

Data entries Data entries

(Index File)
(Data file)

Data Records Data Records

1.9
 General Selection Conditions
 (day<8/9/94 AND rname=‘Paul’) OR bid=5 OR sid=3

 Such selection conditions are first converted to conjunctive normal form

(CNF):
 (day<8/9/94 OR bid=5 OR sid=3 ) AND
(rname=‘Paul’ OR bid=5 OR sid=3)
 We only discuss the case with no ORs (a conjunction of terms of the form
attr op value).
 A B-tree index matches (a conjunction of) terms that involve only
attributes in a prefix of the search key.
 Index on <a, b, c> matches a=5 AND b= 3, but not b=3.
 For Hash index, must have all attributes in search key

1.10
 Two Approaches to General Selections

 First approach: Find the most selective access path,

retrieve tuples using it, and apply any remaining
terms that don’t match the index:
 Most selective access path: An index or file scan that we
estimate will require the fewest page I/Os.
 Terms that match this index reduce the number of tuples
retrieved; other terms are used to discard some retrieved
tuples, but do not affect number of tuples/pages fetched.

1.11
 Most Selective Index - Example
 Consider day < 8/9/94 AND bid=5 AND sid=3.
 A B+ tree index on day can be used;
 then, bid=5 and sid=3 must be checked
for each retrieved tuple.
 Similarly, a hash index on <bid, sid> could be used;
 Then, day<8/9/94 must be checked.
 How about a B+tree on <rname,day>?
 How about a B+tree on <day, rname>?
 How about a Hash index on <day, rname>?

1.12
 Intersection of Rids
 Second approach: if we have 2 or more matching indexes
(w/Alternatives (2) or (3) for data entries):
 Get sets of rids of data records using each matching index.
 Then intersect these sets of rids.
 Retrieve the records and apply any remaining terms.
 Consider day<8/9/94 AND bid=5 AND sid=3. With a B+ tree
index on day and an index on sid, we can retrieve rids of
records satisfying day<8/9/94 using the first, rids of recs
satisfying sid=3 using the second, intersect, retrieve records
and check bid=5.
 Note: commercial systems use various tricks to do this:
 bit maps, bloom filters, index joins

1.13
Join Operators

1.14
 Join Operators

 Joins are a very common query operation.

 Joins can be very expensive:
Consider an inner join of R and S each with 1M records. Q: How many
tuples in the answer?
(cross product in worst case, 0 in the best(?))

 Many join algorithms have been developed

 Can have very different join costs.

1.15
 Equality Joins With One Join Column
SELECT *
FROM Reserves R1, Sailors S1
WHERE R1.sid=S1.sid

 In algebra: R ⋈ S. Common! Must be carefully optimized. R  S

is large; so, R  S followed by a selection is inefficient.
 Assume:
 M = 1000 pages in R, pR =100 tuples per page.
 N = 500 pages in S, pS = 80 tuples per page.
 In our examples, R is Reserves and S is Sailors.
 Cost metric : # of I/Os. We will ignore output costs.
 We will consider more complex join conditions later.

1.16
 Simple Nested Loops Join
foreach tuple r in R do
foreach tuple s in S do
if ri == sj then add <r, s> to result

 For each tuple in the outer relation R, we scan the entire inner relation S.
 How much does this Cost?
 (pR * M) * N + M = 100,000*500 + 1000 I/Os. ( about 50M I/Os!!)
 At 10ms/IO, Total: ???
 What if smaller relation (S) was outer?
 (ps * N) *M + N = 40,000*1000 + 500 I/Os. (better….  40M I/Os)
 Prohibitively expensive…

Q: What is cost if one relation can fit entirely in memory?

M+N = 1500 I/Os!!!!!

1.17
 Page-Oriented Nested Loops Join
foreach page bR in R do
foreach page bS in S do
foreach tuple r in bR do
foreach tuple s in bSdo
if ri == sj then add <r, s> to result

 For each page of R, get each page of S, and write out matching pairs of
tuples <r, s>, where r is in R-page and S is in S-page.

 What is the cost of this approach?

 M*N + M= 1000*500 + 1000 = 501000

 If smaller relation (S) is outer, cost = 500*1000 + 500 = 500500

1.18
 Block Nested Loops Join
 Page-oriented NL doesn’t exploit extra buffers.
 Alternative approach: Use one page as an input buffer for scanning
the inner S, one page as the output buffer, and use all remaining
pages to hold ``block’’ of outer R.
 For each matching tuple r in R-block, s in S-page, add <r, s> to
result. Then read next R-block, scan S, etc.

R&S Join Result

block of R tuples
(k <= B-2 pages)
...
... ...
Input buffer for S Output buffer

1.19
 Examples of Block Nested Loops
 Cost:

Scan of outer + #outer blocks * scan of inner

 #outer blocks = ceiling(#pages of outer/blocksize)
 With Reserves (R) as outer, and 100 pages/Block (B=102):
 Cost of scanning R is 1000 I/Os; a total of 10 blocks (B=102).
 Per block of R, we scan Sailors (S); 10*500 I/Os.
 Total cost: 10*500+1000 = 6000 I/Os
 If space for just 90 pages of R, we would scan S 12 times.
 With 100-page block of Sailors as outer:
 Cost of scanning S is 500 I/Os; a total of 5 blocks.
 Per block of S, we scan Reserves; 5*1000 I/Os.
 Total cost: 5 * 1000 + 500 = 5500 I/Os. (Much better )
 We may be able to do even better for different B!

1.20
 Index Nested Loops Join
foreach tuple r in R do
foreach tuple s in S where ri == sj do
add <r, s> to result

 If there is an index on the join column of one relation (say S), can make it
the inner and exploit the index.
 Cost: M + ( (M*pR) * cost of finding matching S tuples)
 For each R tuple, cost of probing S index is about 1.2 for hash index, 2-4
for B+ tree.
 Cost of then finding S tuples (assuming Alt. (2) or (3) for data entries)
depends on clustering.
 Clustered index: 1 I/O per page of matching S tuples.
 Unclustered: up to 1 I/O per matching S tuple.

1.21
 Examples of Index Nested Loops

 Hash-index (Alt. 2) on sid of Sailors (as inner):

 Scan Reserves: 1000 page I/Os, 100*1000 tuples.
 For each Reserves tuple: 1.2 I/Os to get data entry in index, plus 1 I/O to
get (the exactly one) matching Sailors tuple. Total: 1000+100*1000*2.2
 Hash-index (Alt. 2) on sid of Reserves (as inner):
 Scan Sailors: 500 page I/Os, 80*500 tuples.
 For each Sailors tuple: 1.2 I/Os to find index page with data entries, plus
cost of retrieving matching Reserves tuples. Assuming uniform distribution,
2.5 reservations per sailor (100,000 / 40,000). Cost of retrieving them is 1
or 2.5 I/Os depending on whether the index is clustered. Assume clustered.
 Totals: 500 + 80*500*2.2 = 88.5K I/Os!!! (not so good here )
 Other scenarios may be better though.

1.22
 Sort-Merge Join (R S)

 Sort R and S on the join column, then scan them to do a ``merge’’ (on join
col.), and output result tuples.
 Particularly useful if
 one or both inputs are already sorted on join attribute(s)
 output is required to be sorted on join attributes(s)
 “Merge” phase can require some back tracking if duplicate values appear in
join column
 R is scanned once; each S group is scanned once per matching R tuple.
(Multiple scans of an S group will probably find needed pages in buffer.)

1.23
 Example of Sort-Merge Join

 Cost: Sort S +Sort R + (M+N)

 The cost of merging: usually M+N,
 worst case is M*N (but very unlikely!)
 With 35, 100 or 300 buffer pages, both Reserves and Sailors can be sorted
in 2 passes; total join cost: 7500.
(BNL cost: 2500 to 16500 I/Os)
1.24
 Cost of Sort-Merge
 For B = 35
 Sort-Merge:
 sort R: in two passes=> 4M = 4000
 sort S: in two passes => 4N = 2000
 merge: M+N (hopefully…) => 1500
 Total: 7500
 Block Nested Loop:
 celing(N/B-2)*M+N = 16*1000+500 = 16500
 Sort-Merge Better for B=35!!!!
 For B = 300
 Sort-Merge: the same: 7500
 BNLJ:
 celing(N/B-2)*M+N = 2*1000+500 = 2500
 Here BNLJ is better!!!!

1.25
 Refinement of Sort-Merge Join

 We can combine the merging phases in the sorting of R and S with the
merging required for the join.
 Pass 0 as before, but apply to both R then S before merge.
 If B > , where L is the size of the larger relation, using the sorting
refinement that produces runs of length 2B in Pass 0, #runs of each relation is
< B/2.

 In “Merge” phase: Allocate 1 page per run of each relation, and `merge’ while
checking the join condition
 Cost: read+write each relation in Pass 0 + read each relation in (only) merging
pass (+ writing of result tuples).
 In example, cost goes down from 7500 to 4500 I/Os for B=300.

 In practice, the I/O cost of sort-merge join, like the cost of external sorting,
is linear.
1.26
 Impact of Buffering

 If several operations are executing concurrently, estimating the number

of available buffer pages is guesswork.
 Repeated access patterns interact with buffer replacement policy.
 e.g., Inner relation is scanned repeatedly in Simple Nested Loop Join. With
enough buffer pages to hold inner, replacement policy does not matter.
Otherwise, MRU is best, LRU is worst (sequential flooding).
 Does replacement policy matter for Block Nested Loops?
 What about Index Nested Loops? Sort-Merge Join?

1.27
 Original
Hash-Join Relation OUTPUT Partitions
1

1
INPUT 2
 Partition both relations 2
hash
...
function
on the join attributes
h
using hash function h. B-1
B-1
 R tuples in partition Ri
Disk B main memory buffers Disk
will only match S tuples
in partition Si.
Partitions
Join Result
For i= 1 to #partitions { of R & S
Hash table for partition
hash Ri (k < B-1 pages)
Read in partition Ri fn
h2
and hash it using
h2 (not h2
h). Input buffer Output
for Si buffer
Scan partition Si and
Disk B main memory buffers Disk
probe hash table
1.28
 Observations on Hash-Join

 #partitions k < B, and B-1 > size of smaller partition to be held in

memory. Assuming uniformly sized partitions, and maximizing k, we get:
k= B-1, and M/(B-1) < B-2, i.e., B must be >

 Since we build an in-memory hash table to speed up the matching of

tuples in the second phase, a little more memory is needed.

 If the hash function does not partition uniformly, one or more R partitions
may not fit in memory. Can apply hash-join technique recursively to do
the join of this R-partition with corresponding S-partition.

1.29
 Cost of Hash-Join

 In partitioning phase, read+write both relns; 2(M+N). In matching phase,

read both relns; M+N I/Os.

 In our running example, this is a total of 4500 I/Os.

 Sort-Merge Join vs. Hash Join:

 Given a minimum amount of memory (what is this, for each?) both have a cost
of 3(M+N) I/Os. Hash Join superior if relation sizes differ greatly (e.g., if one
reln fits in memory). Also, Hash Join shown to be highly parallelizable.
 Sort-Merge less sensitive to data skew; result is sorted.

1.30
 Set Operations
 Intersection and cross-product special cases of join.
 Union (Distinct) and Except similar; we’ll do union.

 Sorting based approach to union:

 Sort both relations (on combination of all attributes).
 Scan sorted relations and merge them.
 Alternative: Merge runs from Pass 0 for both relations.

 Hash based approach to union:

 Partition R and S using hash function h.
 For each S-partition, build in-memory hash table (using h2), scan corr. R-
partition and add tuples to table while discarding duplicates.

1.31
 General Join Conditions

 Equalities over several attributes

(e.g., R.sid=S.sid AND R.rname=S.sname):

 For Index NL, build index on <sid, sname> (if S is inner); or use existing
indexes on sid or sname.
 For Sort-Merge and Hash Join, sort/partition on combination of the two join
columns.

 Inequality conditions (e.g., R.rname < S.sname):

 For Index NL, need (clustered!) B+ tree index.
 Range probes on inner; # matches likely to be much higher than for equality
joins.
 Hash Join, Sort Merge Join not applicable!
 Block NL quite likely to be the best join method here.

1.32
 Review
 Implementation of Relational Operations as Iterators
 Focus largely on External algorithms (sorting/hashing)
 Choices depend on indexes, memory, stats,…
 Joins
 Blocked nested loops:
 simple, exploits extra memory
 Indexed nested loops:
 best if 1 rel small and one indexed
 Sort/Merge Join
 good with small amount of memory, bad with duplicates
 Hash Join
 fast (if enough memory), bad with skewed data
 Relatively easy to parallelize

1.33
Aggregation Operators

1.34
 Schema for Examples

Sailors (sid: integer, sname: string, rating: integer, age: real)

Reserves (sid: integer, bid: integer, day: dates, rname: string)

 Similar to old schema; rname added for variations.

 Reserves:
 Each tuple is 40 bytes long, 100 tuples per page, 1000 pages. So, M = 1000, p R
= 100.
 Sailors:
 Each tuple is 50 bytes long, 80 tuples per page, 500 pages.
 So, N = 500, pS = 80.

1.35
 Aggregate Operations (AVG, MIN,
etc.)
 Without grouping:
 In general, requires scanning the relation.
 Given a tree index whose search key includes all attributes in the SELECT or
WHERE clauses, can do index-only scan.
 With grouping:
 Sort on group-by attributes, then scan relation and compute aggregate for
each group. (Better: combine sorting and aggregate computation.)
 Similar approach based on hashing on group-by attributes.
 Given a tree index whose search key includes all attributes in SELECT,
WHERE and GROUP BY clauses, can do index-only scan; if group-by
attributes form prefix of search key, can retrieve data entries/tuples in group-
by order.

1.36
 Sort GROUP BY: Naïve Solution Aggregate

 The Sort iterator naturally permutes its input so that all tuples
are output in sequence Sort
 The Aggregate iterator keeps running info (“transition values” or
“transVals”) on agg functions in the SELECT list, per group:
Example transVals:
 For COUNT, it keeps count-so-far
 For SUM, it keeps sum-so-far
 For AVERAGE it keeps sum-so-far and count-so-far
 As soon as the Aggregate iterator sees a tuple from a new group:
ù It produces an output for the old group based on the agg
function
E.g. for AVERAGE it returns (sum-so-far/count-so-far)
ù It resets its running info.
ù It updates the running info with the new tuple’s info

1.37
Sort GROUP BY: Naïve Solution
A, 3
B, 2
C, 1
D, 1
<B,2>
<A, 3>
<A,2>
<A, 1>
<B,1>
Aggregate

A
B
C
A
AB

Sort

A
B
D
C
B
A
A C A A
B D B

1.38
Hash GROUP BY: Naïve Solution
Aggregate
(similar to the Sort GROUPBY)
Hash
 The Hash iterator permutes its input so that all tuples are output in
groups.
 The Aggregate iterator keeps running info (“transition values” or
“transVals”) on agg functions in the SELECT list, per group
 E.g., for COUNT, it keeps count-so-far
 For SUM, it keeps sum-so-far
 For AVERAGE it keeps sum-so-far and count-so-far
 When the Aggregate iterator sees a tuple from a new group:
ù It produces an output for the old group based on the agg
function
E.g. for AVERAGE it returns (sum-so-far/count-so-far)
ù It resets its running info.
ù It updates the running info with the new tuple’s info

1.39
 External Original
Hashing Relation OUTPUT
1
Partitions
 Partition:
1
Each group will INPUT 2
hash 2
be in a single
...
function
disk-based partition file. But hp B-1
those files have many B-1
groups inter-mixed.
 Rehash: Disk B main memory buffers Disk

For Each Partition i:

hash i into an Partitions Result
in-memory hash table Hash table for partition
Ri (k < B pages)
Return results until fn INPUT
records exhuasted then i++
hash
function
h2p

Disk B main memory buffers

1.40
 We Can Do Better!
HashAgg
 Put summarization into the hashing process
 During the ReHash phase, don’t store tuples, store pairs of the
form <GroupVals, TransVals>
 When we want to insert a new tuple into the hash table
 If we find a matching GroupVals, just update the TransVals
appropriately
 Else insert a new <GroupVals,TransVals> pair
 What’s the benefit?
 Q: How many pairs will we have to maintain in the rehash phase?
 A: Number of distinct values of GroupVals columns
 Not the number of tuples!!
 Also probably “narrower” than the tuples

1.41
 SELECT DISTINCT
Projection R.sid, R.bid

(DupElim) FROM Reserves R

 Issue is removing duplicates.

 Basic approach is to use sorting
 1. Scan R, extract only the needed attrs (why do this 1st?)
 2. Sort the resulting set
 3. Remove adjacent duplicates
 Cost: Reserves with size ratio 0.25 = 250 pages. With 20 buffer pages
can sort in 2 passes, so
1000 +250 + 2 * 2 * 250 + 250 = 2500 I/Os
 Can improve by modifying external sort algorithm:
 Modify Pass 0 of external sort to eliminate unwanted fields.
 Modify merging passes to eliminate duplicates.
 Cost: for above case: read 1000 pages, write out 250 in runs of 20
pages, merge runs = 1000 + 250 +250 = 1500.

1.42
 DupElim & Indexes

 If an index on the relation contains all wanted attributes in its search key,
can do index-only scan.
 Apply projection techniques to data entries (much smaller!)
 If an ordered (i.e., tree) index contains all wanted attributes as prefix of
search key, can do even better:
 Retrieve data entries in order (index-only scan), discard unwanted fields,
compare adjacent tuples to check for duplicates.

 Same tricks apply to GROUP BY/Aggregation

1.43
 Summary of Query Evaluation
 Queries are composed of a few basic operators;
 The implementation of these operators can be carefully tuned (and it is
important to do this!).
 Operators are “plug-and-play” due to the Iterator model.

 Many alternative implementation techniques for each operator; no

universally superior technique for most.

 Must consider alternatives for each operation in a query and

choose best one based on statistics, etc.

 This is part of the broader task of Query Optimization, which we

will cover next!

1.44