1. Load Shedding
CS240B notes

2. Load Shedding in a DSMS
DSMS: online response on boundless and bursty data streams—How?
By using approximations and synopses and even
Shedding load when arrival rates become impossible
Approximations and Synopses are often used with normal load
Shedding is used for bursty streams and overload situations. 2 2

3. QoS and Load Schedding When input stream rate exceeds system capacity
a stream manager can shed load (tuples)
Load shedding affects queries and their answers: drop the tasks and the tuples that will cause least loss
Introducing load shedding in a data stream manager is a challenging problem
Random load shedding or semantic load shedding

4. Problems to Address When to shed load Where to shed load
Overload should be detected quickly
Where to shed load
Avoid wasted work
Upstream Drop Vs. Downstream Drop
How much to shed
The magnitude of the drop
Which tuples to shed

5. Loss-tolerance QoS function
Loss function is not linear:

6. Value-based QoS Value-based QoS
Show which values of the output tuple space are most important.
In a medical application that monitors patient heartbeats
Extreme values are certainly more interesting than normal ones
Corresponding higher utility

7. Load Shedding in Aurora
QoS for each application as a function relating output to its utility
– Delay based, drop based, value based
Techniques for introducing load shedding operators in a plan such that QoS isdisrupted the least
– Determining when, where and how much load to shed

8. Which Query to drop First?
Models and algorithms proposed include
Greedy algorithms or
Fractional Knapsack Problem
Other OR techniques
Must deal with nonlinearities

9. Load Shedding in STREAM
Formulate load shedding as an optimization problem for multiple sliding window aggregate queries
– Minimize inaccuracy in answers subject to output rate matching or exceeding arrival rate
Consider placement of load shedding operators in query plan
– Each operator sheds load uniformly with probability pi

10. Window-Oriented Load Shedding
Input stream divided into windows of size w
Use fewer Slides per windows to compute aggregates—tumbles is the extreme case.
Window-based Sampling
Reservoir sampling for incoming tuples
Expiring tuples pose a more difficult problem.

11. Load Shedding by Sampling for Continuous Aggregate Queries on Data Streams:
Only random samples are available for computing aggregate queries because of
Limitations of remote sensors, or transmission lines
Load Shedding policies implemented when overloads occur
When overloads occur (e.g., due to a burst of arrivals} we can
drop queries all together, or
sample the input---much preferable
Key objective: Achieve answer accuracy with sparse samples for complex aggregates on windows
Can we improve answer accuracy with minimal overhead?

12. Load Shedding To cope with bursty arrivals of high-volume data
DSMS has to shed load while minimizing the degradation of the Quality of Service (QoS)
The goal then becomes determining:
when, where and how much load to shed
An intelligent scheme, can improve the quality of our mining results under bursty arrivals

13. A first Architecture Basic Idea: [BDM04]
…... S1 Sn
Query Network ∑
Basic Idea: [BDM04]
Optimize sampling rates of load shedders for accurate answers.
Find an error bound for each aggregate query.
Determine sampling rates that minimize query inaccuracy within the limits imposed by resource constraints.
This approach works for SUM and COUNT
Generalization to other functions? Si
Data Stream
Load Shedder
Query Operator ∑
Aggregate

14. Query Network: arbitrary placement of aggregates and shedder after any aggregateSn L1 L4 Q1 Q4 L2 L5 Q2 Q3 Q5
Data Stream
Load Shedder
Aggregate Operator

15. Generalized Load Shedding in Stream Mill
A general framework that achieves optimal load shedding policies, while accommodating:
Different requirements for different users, different query sensitivities, and different penalties.
Applicability to a wide spectrum of aggregate functions:
We have formally characterized using a new notion, called reciprocal-error queries.
Proposing an extensible architecture that allows UDAs to benefit from the system provided load shedding functions.
Significant improvements (in absolute error, false positives, and false negatives) compared to the common uniform approach.
We propose an efficient (linear-time) algorithm to handle severe overloads without losing optimality.

16. Goals to Achieve Light-weight overhead handling
React to overload immediately
Minimizing QoS degradation
Delivering subset results
Only omitting tuples from the correct answer
Never produce incorrect answers

17. Prediction & Improvements
A larger class of queries was considered in [LZ08]
SUM, COUNT, AVG, Quantiles.
Temporal Correlation between answers can be used to improve answer
Example: sensor data
Current answer can be adjusted by the past answers so that:
Low sampling rate  current answer less accurate  more dependent on history.
High sampling rate  current answer more accurate  less dependent on history.
A Bayesian quality enhancement module which can achieve this objective automatically and reduce the uncertainty of the approximate answers.

18. Improved Model Using History
…...
The observed answer à is computed from random samples of the complete stream with sampling rate P.
A bayesian method to obtain the improved answer by combining
the observed answer
the error model
history of the answer
Query Network Sn ∑ ∑
…... ∑ Ã
…...
History P
Quality Enhancement Module
Query Operator
Load Shedder
Data Stream Si ∑
Improved answer
Aggregate

19. Summary
An error model
Works for ordered statistics and data mining functions as well as with traditional aggregates,
computationally very efficient
Bayesian quality enhancement method for approximate aggregates in the presence of sampling.
No correction when concept changes are suspected—a two-sample test used to detect suspected changes.

20. References—Sampling and load shedding
[Tabul03] Nesime Tatbul, Ugur Cetintemel, Stanley B. Zdonik, Mitch Cherniack, Michael Stonebraker: Load Shedding in a Data Stream Manager.VLDB2003, pp
[BDM04] Brian Babcock, Mayur Datar, Rajeev Motwani: Load Shedding for Aggregation Queries over Data Streams. ICDE 2004:
[Tabul07] Nesime Tatbul, Ugur Cetintemel, Stanley B. Zdonik: Staying FIT: Efficient Load Shedding Techniques for Distributed Stream Processing. VLDB 2007:
[LZ08] Yan-Nei Law and Carlo Zaniolo: Improving the Accuracy of Continuous Aggregates and Mining Queries on Data Streams under Load Shedding. International Journal of Business Intelligence and Data Mining, 2008.
[ICDE 2010] Barzan Mozafari and Carlo Zaniolo, Optimal Load Shedding with Aggregates and Mining Queries. In Proceedings of the 26th International Conference on Data Engineering (ICDE 2010), Long Beach, California, USA, March 1-6, 2010.