Toward timely, predictable and cost-effective data analytics. Renata Borovica-Gajić DIAS, EPFL

Transcription

1 Toward timely, predictable and cost-effective data analytics Renata Borovica-Gajić DIAS, EPFL

2 Big data proliferation Big data is when the current technology does not enable users to obtain timely, cost-effective, and quality answers to data-driven questions. [Steve Todd, Berkeley] Technology follows Moore s Law * The Digital Universe in 2020: Big Data, Bigger Digital Shadows, and Biggest Growth in the Far East, 2012, IDC Trends in big data analytics, 2014, Kambatla et al 2

3 Time Cost (million $) What business analysts want Timely, predictable, cost-effective queries 80% reuse within 3 hours Minimal data-toinsight time User frustration Wasted resources Expected Actual Predictable response time DW [WinterCorp, 2013] Development Administration System Low infrastructure cost 3

4 Thesis statement As traditional DBMS rely on predefined assumptions about workload, data and storage, changes cause loss of performance and unpredictability. Insight Query execution must adapt at three levels (to workload, data and hardware) to stabilize and optimize performance and cost. 4

5 Outline Minimize data-to-insight time Workload-driven adaptation [SIGMOD 12, VLDB 12, CACM 15] Improve predictability of response time Data-driven adaptation [DBTest 12, ICDE 15] Reduce analytics cost Cold storage & hardware-driven adaptation [VLDB 16] 5

6 Outline Minimize data-to-insight time Workload-driven adaptation Improve predictability of response time Data-driven adaptation Reduce analytics cost Cold storage & hardware-driven adaptation 6

7 Time Execution breakdown (%) insight data Data-to-insight time Traditional query stack querying 70 X loading Time to first insight too long Does not scale with data growth Current technology efficient exploration Raw data querying stack Raw data querying overheads Overheads too high Processing (Q1) Convert Tokenize Parse I/O 7

8 Execution breakdown (%) LOAD Response time Optimize raw data querying stack Raw data querying stack Processing (Q1) Convert Tokenize Parse I/O Not everything needed for Q1 NoDB Let users show by asking queries NoDB: Workload-driven data loading & tuning Q4 Q3 Q2 Q1 Raw data querying Q4 Q3 Q2 Q1 DBMS Q4 Q3 Q2 Q1 NoDB 8

9 Frequency PostgresRaw: NoDB from idea to practice 1. Positional indexing Pointers to end of tuples 1 Supplier# each slyly... 2 Supplier# slyly bold... 3 Supplier# blithely... 4 Supplier# riously eve... 5 Supplier# Slyly... 6 Supplier# final Cache NationKey Name 17 Supplier#01 Pointers to attributes scan Workload 5 Supplier#02 3. Statistics # Buckets Adjust to queries = progressively cheaper access 9

10 Execution time (sec) PostgresRaw in action Setting: 7.5M tuples, 150 attributes, 11GB file Queries: 10 arbitrary attributes per query, vary selectivity ~ 7000 ~ 4806 Q20 Q19 Q18 Q17 Q16 Q15 Q14 Q13 Q12 Q11 Q10 Q9 Q8 Q7 Q6 Q5 Q4 Q3 Q2 Q1 Load MySQL CSV Engine DBMS X DBMS X PostgreSQL PostgresRaw Data-to-insight time halved with PostgresRaw MySQL w/ external files Per query performance comparable to traditional DBMS 10

11 Summary of PostgresRaw Query processing engine over raw data files Uses user queries for partial data loading and tuning Comparable performance to traditional DBMS IMPACT Enables timely data exploration with 0 initialization Decouples user interest from data growth 11

12 Outline Minimize data-to-insight time Workload-driven adaptation Improve predictability of response time Data-driven adaptation Reduce analytics cost Cold storage & hardware-driven adaptation 12

13 Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12 Q13 Q14 Q16 Q18 Q19 Q21 Q22 Normalized exec. time (log scale) Index: with or without? Setting: TPC-H, SF10, DBMS-X, Tuning tool 5GB space for indexes Tuned With indexes Original Without indexes TPC-H Query Performance hurt after tuning 13

14 Execution time Performance cliff Access path selection problem Re-optimization [MID 98, POP 04, RIO 05, BOU 14] Index Scan Full Scan RISK Full Scan 0 Estimated Selectivity Actual 100% Statistics: unreliable advisor Re-optimization: risky 14

15 Execution time Quest for predictable execution Index Scan Predictable Execution RISK Full Scan 0 Selectivity 100% Removing variability due to (sub-optimal) choices 15

16 Smooth Scan Morph between Index and Sequential Scan based on observed result distribution 16

17 Modes: Morphing mechanism 1. Index Access: Traditional index access 2. Entire Page Probe: Index access probes entire page 3. Gradual Flattening Access: Probe adjacent region(s)... INDEX HEAP PAGES Mode 1 Mode 2 Mode 3 17

18 Morphing policy Selectivity Increase -> Mode Increase Selectivity Decrease -> Mode Decrease SEL_region >= SEL_global SEL_region < SEL_global INDEX HEAP PAGES X: Page with result SR: Region selectivity SG: Global selectivity X X X XX X XX X X X X X X XX SR:1 SR:1 SR:0.5 SR:0.75 SR:1 SR:1 SR:0.5 SG: Region snooping = Data-driven adaptation 18

19 Execution time (sec) (log scale) Smooth Scan in action Setting: Micro-benchmark, 25GB table, Order by, Selectivity 0-100% Selectivity(%) Full Scan Index Scan Sort Scan Smooth Scan Near-optimal over entire selectivity range 19

20 Summary of Smooth Scan Statistics-oblivious access path Uses region snooping to morph between alternatives Near-optimal performance for all selectivities IMPACT Removes access path selection decision Improves predictability by reducing variability in query execution 20

21 Outline Minimize data-to-insight time Workload-driven adaptation Improve predictability of response time Data-driven adaptation Reduce analytics cost Cold storage & hardware-driven adaptation 21

22 Proliferation of cold data 80% enterprise data is cold with 60% CAGR [Horison, 2015] cold data: incredibly valuable for analysis [Intel, 2013] Cold Storage Devices (CSD) to the rescue Active disks Latency ~10sec B A Power one disk Latency ~10ms Cool one disk PB-size storage at cost ~ tape and latency ~ disks 22

23 CSD in the storage tiering hierarchy DRAM SSD 15k RPM HDD 7200 RPM HDD Tiers Tape $$$ Performance $$ Capacity $ Archival ns µs ms sec min hour Data Access Latency 23

24 CSD in the storage tiering hierarchy DRAM SSD 15k RPM HDD 7200 RPM HDD Tiers CSD $$$ Performance $$ Capacity $ COLD Tape $ Archival? ns µs ms sec min hour Data Access Latency Can we shrink tiers to reduce cost? 24

25 CSD in the storage tiering hierarchy DRAM SSD 15k RPM HDD 7200 RPM HDD Tiers $$$ Performance $$ Capacity? CSD $ COLD $ Archival ns µs ms sec min hour Data Access Latency Can we shrink tiers to reduce cost? 25

26 Cost (x1000$) CSD in the storage tiering hierarchy DRAM SSD ns µs 15k RPM HDD ms Tiers CSD sec min Performance hour Data Access Latency $$$ $ COLD Storing 100TB of data [Horison, 2015] Trad. 3-tier $159,641 CSD 2-tier CSD offer significant cost savings (40%) But can we run queries over CSD? 26

27 Average execution time (x1000 sec) Query execution over CSD Setting: virtualized enterprise datacenter, clients: PostgreSQL, TPCH 50, Q12, CSD: shared, layout: one client per group 5 Postgre CSD SQL Ideal 4 HDD Number of clients (groups) Lost opportunity: CSD relegated to archival storage 27

28 Skipper to the rescue Virtualized enterprise data center VM1 VM2 VM3 DB1 DB2 Multi-way joins: VM Opportunistic execution triggered upon data arrival DB3 PostgreSQL 2. MJoin Network 1. I/O Scheduler object-group map. 3. Cache Management Hash Hash Hash Scan A Scan B Scan C Novel ranking algorithm: Balances access efficiency across groups and fairness across clients Cold Storage A1 B1 C1 A2 Progress driven caching: Favors caching of objects to maximize query progress 28

29 Average exec. time (x1000 sec) Skipper in action Setting: multitenant enterprise datacenter, clients: TPCH 50, Q12, CSD: shared, layout: one client per group PostgreSQL on CSD Ideal PostgreSQL on HDD Skipper Number of clients (groups) Approximates HDD-based capacity tier by 20% avg. 29

30 Summary of Skipper Efficient query execution over CSD with: 1. Rank-based I/O scheduling 2. Out-of-order execution based on multi-way joins 3. Progress based caching policy Approximates performance of HDD-based storage tier IMPACT Cold storage can reduce TCO by shrinking storage hierarchy Skipper enables data analytics-over-csd-as-a-service 30

31 Thesis contributions Minimize data-to-insight time Workload-driven adaptation Skip loading, tune as a byproduct of query execution Improve predictability of response time Data-driven adaptation Remove access decisions a priori, transform gradually Reduce analytics cost Cold storage & hardware-driven adaptation From plan pull-based to hardware push-based execution Uncertainty cured with adaptivity Thank you! 31