Aggregation and Degradation in JetStream: Streaming Analytics in the Wide Area

Transcription

1 Aggregation and Degradation in JetStream: Streaming Analytics in the Wide Area Ariel Rabkin Princeton University Work done with Matvey Arye, Siddhartha Sen, Vivek S. Pai, and Michael J. Freedman

2 Today s Analytics Architectures MillWheel (Google) Storm Backhaul is inefficient and inflexible 2

3 Tomorrow s Architecture: JetStream JetStream Backhaul is inefficient and inflexible Goal: optimize use of WAN links by exposing them to streaming system. 3

4 Backhaul is Intrinsically Inefficient Bandwidth Buyer s remorse: wasted bandwidth Needed for backhaul Time [two days] Available Analyst s remorse: system overload or missing data 4

5 Stream Processing Basics Input Data Input Data Stream Operators Stream Operators Stream Operators Stream Operators Site A Site B Stream Operator Site C Filtering (count > 100) Sampling (drop 90% of data) Image Compression Some Operators in JetStream: Quantiles (95 th percentile) Query stored data 5

6 The JetStream System What: Streaming with aggregation and degradation as first-class primitives Where: Storage and processing at edge Why: Maximize goodput using aggregation and degradation How: Data cubes and feedback control 6

7 An Example Query CDN How popular is every URL? CDN 7

8 Mechanism 1: Storage with Aggregation CDN Local Aggregation and Storage Every minute, compute request counts by URL CDN Local Aggregation and Storage 8

9 Mechanism 2: Adaptive Degradation CDN Local Aggregation and Storage Adjustable Filtering Every minute, compute request counts by URL CDN Local Aggregation and Storage Adjustable Filtering 9

10 Requirements for Storage Abstraction Update-able (locally and incrementally) Stored Data += Data Data size is reducible (with predictable accuracy cost) Data Data Merge-able (without accuracy penalty) Data + = Data Merged Representation 10

11 The Data Cube Model Cube: A multidimensional array, indexed by a set of dimensions, whose cells hold aggregates. Counts by URL 12:00 12:01 12: Aggregation used for: Updates Roll-ups Merging cubes Summarizing cubes Cubes have aggregation function: Agg(, )à 11

12 Cubes can be Rolled Up Cube: A multidimensional array, indexed by a set of dimensions, whose cells hold aggregates. Counts by URL 12:00 12:01 12: Counts by URL * Counts by URL 12:00 12:01 12:02 *

13 Cubes Unify Storage and Aggregation Update Standing Query Update sent downstream Update Stored Data Update One-off query 13

14 Degradation: The Big Picture Local Data Dataflow Operators Summarized or Approximated Data Network Dataflow Operators Feedback control Level of degradation auto-tuned to match bandwidth. Challenge: Supporting mergeability and flexible policies 14

15 Mergeability Imposes Constraints Every ?????? Every Every Every 30?? Every Insight: Degradation may be discontinuous 15

16 There Are Many Ways to Degrade Data Can coarsen a dimension Can drop low-rank values 16

17 Coarsening Does Not Always Help Savings from Aggregation Domains URLs 5s minute 5 m hour day Aggregation time period 17

18 Degradations Have Trade-offs Name Fixed BW Savings Fixed Accuracy cost Parameter Dim. Coarsening Usually no Yes Dimension Scale Drop values (locally) Drop values (globally) Audiovisual downsampling Histogram Coarsening Yes No Cut-off No, multi-round protocol Yes Cut-off Yes Yes Sample rate Yes Yes Number of Buckets 18

19 A Simple Idea that Does Not Work Incoming data Coarsening Operator Sampled Data Network Sending 4x too much We have sensors that report congestion. Have operators read sensor and adjust themselves? 19

20 A Simple Idea that Does Not Work Incoming data Coarsening Operator Sampled Data Increase aggregation period up to 10 sec. If insufficient, use sampling Network Sending 4x too much We have sensors that report congestion. Have operators read sensor and adjust themselves? 20

21 Challenge: Composite Policies Incoming data Coarsening Operator Sampling Operator Network Sending 4x too much Chaos if two operators are simultaneously responding to the same sensor 21

22 Interfacing with Operators Incoming data Coarsening Operator Sampling Operator Network Controller Shrinking data by 50% Possible levels: [0%, 50%, 75%, 95%, ] Sending 4x too much Go to level 75% 22

23 Experimental Setup 80 nodes on VICCI testbed at three sites (Seattle, Atlanta, and Germany) Princeton Policy: Drop data if insufficient BW 23

24 Without Degradation BW (Mbits/sec) Drop BW Experiment time (minutes) Latency (sec) Maximum latency 95 th percentile latency Median Latency Elapsed time (minutes) 24

25 Degradation Keeps Latency Bounded BW (Mbits/sec) Latency (sec) Bandwidth Shaping Experiment time (minutes) 95 th percentile latency Median Latency Elapsed time (minutes) 25

26 Showing maximum latencies Latency (sec) Maximum Latency 95 th percentile latency Median Latency Elapsed time (minutes) 26

27 Programming Ease Scenario Lines of code Slow requests 5 by URL 5 Bandwidth by node 15 Bad referrers 16 Latency and size quantiles 25 Success by domain 30 Top 10 domains by period 40 Big 97 27

28 Conclusions and Future Work Useful to embed aggregation and degradation abstractions in streaming systems. Aggregation can be unified with storage. System must accommodate degradation semantics. Open questions: How to guide users to the right degradation policy? How to embed abstractions in higher-level language? 28