Reducing Order Enforcement Cost in Complex Query Plans

Reducing Order Enforcement Cost in Complex Query Plans

Ravindra Guravannavar and S. Sudarshan(To appear in ICDE 2007)

2

Background Sort-based query processing algorithms

Sort-merge Join (also Union/Intersection) Sort-based grouping and duplicate elimination

Explicit “order by” Notion of “Interesting Sort Orders” (System-R)

Find and remember the best plan for each sort order that may be useful

Optimization goal in Volcano : (expr, sort-order)

3

The Problem Interesting orders can be too many!

Factorial in number of attributes involved Plan cost can vary substantially with the

choice of interesting order Clustering and covering indices Other operators in the input sub-expressions Possibility of partial sorting

G Group By {a2,a4,a5,…}

R S

R.a1=S.a1 and R.a2=S.a2 … R.an=S.an

4

Motivation Joins in data integration and decision support

involve large number of attributes Increasing use of covering indices

Several alternative sort orders Partial sorting

Query patterns Attributes common to multiple operators

Known techniques Work only for unary operators like group-by

5

Outline of the Talk Partial sorting

Changes to external sort Optimizer changes to handle partial sort orders

Interesting orders for a join tree : A special case Problem is NP-Hard A 2-approximation for the special case

The general problem Notion of favorable orders Plan generation using favorable orders Post-optimization phase

Experimental results

6

Exploiting Partial Sort Orders

Sort on (a1, a2) given (a1) Standard external-sort

Cost is independent of input sort order Replacement-selection

Produces single run but incurs I/O Both methods break the pipeline – first o/p tuple

after reading all i/p

R S

R.a1=S.a1 and R.a2=S.a2

C. Index on (R.a1)

(a1) (a1,a2) () (a1,a2)

7

A Minor Change to External Sorting

Multiple “partial sort segments” Hold only one segment at any given

time When a new segment starts

Sort the current segment and output

No run generation I/O if each segment fits in memory

Early output (good for Top-K) Reduced comparisons

O(n log n/k) Vs. O(n log n), k = # segments

a1 a2

1 2

1 1

1 5

1 3

2 4

2 1

2 6

6 3

… …

8

Optimizer Changes to Handle Partial Sort Orders Cost Model for Partial Sort:

Let the input order be o1

Required (output) order be o2

Let os=Longest common prefix between o1 and o2

Let or=o2 – os (i.e, os + or = o2) A(o) = Attribute set of order o Є : Empty (no) sort order

coe(e, o1,o2) = D(e, A(os)) X coe(e’, Є, or), where e’=p(e) and p equates A(os) to a constant.

9

Optimizer Changes to Handle Partial Sort Orders Cost Model for Partial Sort:

coe(e, o1,o2) = D(e, A(os)) X coe(e’, Є, or), where e’=p(e) and p equates A(os) to a constant.

o1=(a,b)

o2=(a,c)

os=(a), or=(c), e’=(a=k)(e)

e

10

Flexible Order Requirements Most operators have interest in any order on the

attributes involved Merge-Join, Merge-Union, Group By, Duplicate Elimination Binary operators demand the same order from inputs

G {a1, a2}

{a1,a2,a3,a4}

{a4,a7}{a3,a5,a6}

11

Finding Optimal is NP-Hard A special case:

All relations/intermediate results of the same size

All attribute cardinalities same

We try to maximize the length of common prefixes

Maximize LCP(pi, pj)

Reduction from graph layout problem SUM-CUT Optimal algorithm for paths and 2-approximation for binary trees

12

A 2-Approximation Algorithm Optimal algorithm for paths

s2s1 sns3 Sn-1

OPT(i,j) = max {OPT(i,k) + OPT(k+1,j) + c(i,j)}, i ≤ k < j

2-Approximation for binary trees

- OPT ≤ OPT-EVEN + OPT-ODD- Take the one with higher benefit

Even levelsOdd levels

13

General Case Logical plan space for inputs not expanded

(i.e, Join order not fixed)

Varying sizes of relations and intermediate results

All orders on base relations do not have the same cost (due to clustering and covering indices)

14

Overview of the Approach Identify a small set of favorable orders

Orders that are relatively inexpensive Should not require expanding the input plan space

Plan generation (Phase-1) Deduce the interesting orders from the favorable

orders Try each of the interesting order, retain the best

Plan refinement (Phase-2) Use the 2-approximation algorithm and refine the

sort orders further

15

Favorable Orders Benefit of an order:

benefit(o, e) = cbp(e, Є) + coe (e, Є, o) – cpb(e,o)Positive benefit The order can be obtained at cost

less than the full sort of unordered result (e.g., the

clustering order)

Favorable orders:ford(e)={ o : benefit(o,e) > 0 } Can be a huge set E.g., Every order having the clustering order as its

prefix is a favorable order.

16

Minimal Favorable Orders A favorable order o that satisfies:

1. o’ ≤ o s.t. cbp(e, o’) + coe(e, o’, o) = cbp(e,o)

2. o” s.t. o ≤ o” and cbp(e, o”) = cbp(e,o)E.g., Relation R with clustering index on (a1,a2)

(a1,a2) is a minimal favorable order

(a1 ), (a1,a2,a3) are not

ford-min(e): Set of all minimal favorable orders for expression e

For base relations size of ford-min limited to the number of covering indices

E

E

17

Computing Favorable Orders: Issues Defined in terms of cost of best plan

Need them before optimizing input sub-expressions Even ford-min can get prohibitively large for

join, group-by expressions

R S

J1

J2

ford-min contains every permutation of

the join attributes

18

Heuristics for Computing ford-min

e=R : {o: o is clustering or covering index order}

e=p(e1) : {o: o ford-min(e1)}

e=L(e1) :{o: o’ ford-min(e1) and o=o’ ^ L} a,b(e1), ford-min(e1)={(a,c,b)} ford-min(e)={(a)}

e=e1 e2 : Let T=ford-min(e1) U ford-min(e2) T U {o: o’ T and o=((o’ ^ S) permute(S – A(o’ ^ S)))

UU

U

19

Heuristics for Computing ford-minS={a,b,c,d}

ford-min={(a,b,e),(b)} ford-min={(a)}

T = {(a,b,e), (b), (a)}

Input F.Order (o) o ^ {a,b,c,d} Extended Order

(a,b,e) (a,b) (a,b,c,d)

(b) (b) (b,a,c,d)

(a) (a)

20

Plan Generation (Phase-1) Form the set I of interesting orders to try

Collect input favorable orders and rqd. o/p order Take LCP with the set of join attributes Extend the orders (arbitrarily) to include remaining

attributes For each order o in I, generate optimization

sub-goals for input sub-expressions

21

Plan Refinement (Phase-2)

Identify the suffix that can be freely reordered Use the 2-approximation algorithm to reorder

the suffix

R2(a)

(a,b,c,h)

(a,d,h)

R4(a)

R3(a)

R1(a)

(a,e,h){a,d,h} {a,e,h}

{a,e,h}

(a,h,e)

(a,h,b,c)

(a,h,d)

22

Experiments1. Benefits of exploiting partial sort orders

2. Evaluate the plans produced by our optimizer extensions

Systems Compared

PostgreSQL 8.1.3, SQLServer 2005,

DB2 8.2, PYRO

Test Machine Intel P4 (HT) PC, 512 MB

Dataset TPC-H 1GB and synthetic

Queries Synthetic and from a real application

23

Experiment 1SELECT suppkey, partkey FROM lineitem

ORDER BY suppkey, partkey;

(suppkey) (suppkey, partkey)

24

Experiment 2

R(c1,c2,c3), 10 M records, (c1)(c1,c2), card(c1)=10,000

25

Experiment 3

26

Experiment 4

SELECT ps_suppkey, ps_partkey, ps_availqty,

sum(l_quantity) AS total_required

FROM partsupp, lineitem

WHERE ps_suppkey=l_suppkey AND ps_partkey=l_partkey

AND l_linestatus='O'

GROUP BY ps_partkey, ps_suppkey, ps_availqty,

HAVING sum(l_quantity) > ps_availqty

ORDER BY ps_partkey;

Parts running out of stock:

27

Experiment 4 - Plans

Merge-Join Plan on SYS1 and SYS2 Plan Generated by PYRO-O

28

Experiment 4 & 5 - Timings

29

Experiments with Variants of PYRO

PYRO : Baseline PYROPYRO-O-: No partial sortPYRO-P : Postgres HeuristicPYRO-O : Our ApproachPYRO-E : Exhaustive

30

Optimization Overheads

31

Questions?