1. ICICLES: Self-tuning Samples for Approximate Query Answering
Haidong Wang
02/13/2007

2. Outline Introduction and background
Intuition and basic idea of Icicles
Icicle maintenance
Estimators for Aggregate Queries
Performance Evaluation
Conclusion and my comment

3. Introduction
Analysis of data in data warehouses is useful in decision support
OLAP—provide interactive response time s to aggregate queries
Approximate query answering (AQUA) system are being developed
Tolerate approximate answers to achieve response time

4. Introduction Various approaches to answering approximate queries
Sampling-based
Histogram-based
Clustering
Probabilistic
Wavelet-based

5. Uniform Random Sample Sales S_sales 50% Sample scale factor
Branch
State
Sales 1 CA
80K 2 TX
42K 3
40K 4 5
75K 6
48K 7
55K 8
38K 9 10
41K
50%
Sample
Branch
State
Sales 2 TX
42K 4 CA 6
48K 8
38K 10
41K
scale factor
SELECT SUM(sales) x 2 AS cnt
FROM s_sales
WHERE state = ‘TX’

6. Why Icicles? In practice, queries may follow a predictable pattern
Static sampling strategy treats all tuples uniformly, thus wasting memory on less required tuples.
Sample relation space better utilized if more samples from actual result set are present
Example: Manager of Walmart in Asia
location=”Asia” and year>=2000

7. Intuition of Icicles
Tuples being selected to the sample is proportional to the frequency with which it’s required to answer queries
location=”Asia” and year>=2000
We need a dynamic algorithm that tunes the sample with respect to the most recent knowledge of the workload

8. What is Icicles Icicle for Relation R is:
A uniform random sample of a multiset of tuples L, which is the union of R and all sets of tuples that were required to answer queries in the workload

9. What is Icicles
Uniform random
sample R
Icicle
R(Q1)
R(Q2)
R(Q3)

10. Icicle Maintenance Reservoir sampling algorithm
Each time we only need to access the new block of data
Uniform random
sample R
Icicle
R(Q1)
R(Q2)
R(Q3)

11. Icicle Maintenance

12. Icicle Maintenance Example
SELECT average(*)
FROM widget-tuners
WHERE date.month = ‘April’

13. Icicle Maintenance Maintaining the frequency relation
Keep frequency relation in main memory
Delay updating the frequency into the disk until the magnitude of the change crosses a threshold

14. Estimators for Aggregate Queries
Traditional estimators can’t be used due to selection bias and duplicates in icicle
Example: count
Maintain a set of frequencies one per tuple in the relation.

15. Estimators for Aggregate Queries
Avg: the average of distinct tuples in sample satisfying query
Doesn’t require frequency attribute
Count: the sum of expected contributions of all tuples in icicle that satisfy the query
Sum: Avg * Count

16. Performance Evaluation
Plots definition:
Static sample:
Uniform random sample on the relation
Icicle:
Icicle evolves with the workload
Icicle-complete
The tuned icicle again on the same workload

17. Performance Evaluation

18. Performance Evaluation

19. Conclusion
icicle are better than static sampling when workload focuses on relatively small subsets in relation
When the workload is randomized, icicle is at least the same as static sampling
Icicles adapt quickly to changing workload

20. My Comment Icicle is a trade-off between accuracy and cost
Icicle works well under certain restrictions
How to define a workload
Has to be exact query (not approximate)
The typical scenario of a analyst using Icicle:
First do a bunch of approximate query
Then do one or more exact query

21. My conclusion Icicle is useful when the following is true:
workload focuses on relatively small subsets in relation
Calls for high accuracy of approximate answer
Has to have exact query (the more the better)

22. Thanks
Questions?