Finding the Pitfalls in Query Performance

Transcription

1 Finding the Pitfalls in Query Performance M.L. Kersten P. Koutsourakis Y. Zhang CWI, MonetDB Solutions EU H2020 project ACTiCLOUD

2 The Challenge MonetDB Mar-18 Which system is relatively better? Postgres 10.6 MonetDB Jul-17 Postgres 11 SQLite

3 The Solution, TPC-H? Q4 Q5 Q6 MonetDB Mar-18 Q3 Q7 Q15 Q1 Q2 Q9 Q8 Q13 Q14 Q16 Postgres 10.6 Q10 Q12 Q17 Q11 Q19 Q18 Q21 Q22 MonetDB Jul-17 Q20 Postgres 11 SQLite

4 TPC-H is a collection of random points? MonetDB Mar-18 Postgres 10.6 Query space MonetDB Jul-17 Postgres 11 SQLite

5 TPC-H may miss clarifying queries Query tells a better story for MonetDB MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Postgres 11 SQLite

6 TPC-H may miss discriminative queries Postgres is relatively better than MonetDB MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Postgres 11 SQLite

7 TPC-H may miss discriminative queries MonetDB Mar-18 MonetDB Degrades on A query Postgres 10.6 MonetDB Jul-17 Postgres 11 SQLite

8 TPC-H may miss discriminative queries MonetDB Mar-18 Postgres 10.6 SQLite is superior MonetDB Jul-17 Postgres 11 SQLite

9 The Challenge Which system is better on a benchmark? MonetDB Mar-18 What queries perform relatively better on system A than system B? Postgres 10.6 MonetDB Jul-17 Postgres 11 SQLite

10 The Challenge MonetDB Mar-18 SQL scalpel Find the discriminative queries Postgres 10.6 MonetDB Jul-17 Postgres 11 SQLite

11 Assumptions SQL scalpel Find the discriminative queries The database schema and data distribution are given A collection of business inspired queries are available No direct access to the Database/DBMS/Platform

12 The Solution, TPC-H as a start Q1 Q2 Q10 Q3 Q9 Q11 Q4 Q8 Q12 Q19 Q5 Q7 Q13 Q18 Q6 Q14 Q17 Q21 Q15 Q16 Q22 MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Q20 Postgres 11 SQLite

13 The Solution, TPC-H as a start -- Query 6 Q3 select Q2 Q8 Q14 sum(l_extendedprice * l_discount) as revenue from Q1 Q9 Q13 lineitem Q10 Q12 Q17 where l_shipdate >= date ' Q18 Q11 and l_shipdate < date ' ' + interval '1' year and Q19 Q21 l_discount between and and l_quantity < 24; Q20 Q4 Q5 Q7 Q6 Q15 Q16 Q22 SQLite MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Postgres 11

14 The Solution, TPC-H as a start -- Query 6 Q3 Q4 Q5 Q7 select Q2 Q8 sum(l_extendedprice * l_discount) as revenue from Q1 Q9 Q13 lineitem Q10 Q12 where l_shipdate >= date ' Q18 Q11 Q19 Q6 Q14 Q17 Q21 Q15 Q16 Q22 MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Q20 Postgres 11 SQLite

15 The Solution, TPC-H as a start -- Query 6 Q3 Q4 Q5 Q7 select Q2 Q8 sum(l_extendedprice * l_discount) as revenue from Q1 Q9 Q13 lineitem Q10 Q12 where l_discount between Q18 Q and and l_quantity < 24; Q19 Q6 Q14 Q17 Q21 Q15 Q16 Q22 MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Q20 Postgres 11 SQLite

16 The Solution, TPC-H as a start -- Query 6 Q3 Q4 Q5 Q7 select Q2 Q8 sum(l_extendedprice * l_discount) as revenue from Q1 Q9 Q13 lineitem Q10 Q12 where l_shipdate < date ' ' Q18 Q11 + interval '1' year Q19 Q6 Q14 Q17 Q21 Q15 Q16 Q22 MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Q20 Postgres 11 SQLite

17 The Solution, TPC-H as a start -- Query 6 Q3 select Q2 Q8 Q14 sum(l_extendedprice * l_discount) as revenue from Q1 Q9 Q13 lineitem Q10 Q12 Q17 where l_shipdate >= date ' Q18 Q11 and l_shipdate < date ' ' + interval '1' year and Q19 Q21 l_discount between and and l_quantity < 24; Q20 Q4 Q5 Q7 Q6 Q15 Q16 Q22 SQLite MonetDB Mar-18 Postgres 10.6 MonetDB Jul-17 Postgres 11

18 SQLscalpel compiles the query into a grammar -- Query 6 select sum(l_extendedprice * l_discount) as revenue from lineitem where l_shipdate >= date ' and l_shipdate < date ' ' + interval '1' year and l_discount between and and l_quantity < 24; tpch_q6: SELECT ${projection} FROM ${table} WHERE ${pred} ${predlist}* projection: sum(l_extendedprice * l_discount) as revenue table: lineitem pred: l_shipdate >= date ' ' l_shipdate < date ' ' + interval '1' year l_discount between and l_quantity < 24 predlist: AND ${pred}

19 SQLscalpel enumerates templates SELECT ${projection} FROM ${table} WHERE ${pred} SELECT ${projection} FROM ${table} WHERE ${pred} AND ${pred} SELECT ${projection} FROM ${table} WHERE ${pred} AND ${pred} AND ${pred} SELECT ${projection} FROM ${table} WHERE ${pred} AND ${pred} AND ${pred} AND ${pred} tpch_q6: SELECT ${projection} FROM ${table} WHERE ${pred} ${predlist}* projection: sum(l_extendedprice * l_discount) as revenue table: lineitem pred: l_shipdate >= date ' ' l_shipdate < date ' ' + interval '1' year l_discount between and l_quantity < 24 predlist: AND ${pred}

20

21 Architecture Maintain a query pool Pick promising candidates Keep a workflow database SQL scalpel Support analysis

22 SQLscalpel architecture Maintain a query pool Pick promising candidates Keep a workflow database Support analysis SQL scalpel SQLscalpel server Script parser Grammar Template pool Task Generator SQLscalpel client Task Executor SQLscalpel database

23 SQLscalpel architecture Maintain a query pool Pick promising candidates Keep a workflow database Support analysis Strategies: Baseline Random Alter a lexical term Expand a query Prune a query Ditch manually

24 SQLscalpel architecture Maintain a query pool Pick promising candidates Keep a workflow database Support analysis Strategies: FCFS Manual steering Simulated annealing Biased terms

25

26

27

28

29

30

31 SQLscalpel pitfalls Handle SQL dialects Easily generate erroneous queries Brings the target down using Cartesian results tpch_q6: SELECT ${projection} FROM ${table} WHERE ${pred} ${predlist}* projection: sum(l_extendedprice * l_discount) as revenue table: lineitem lineitem, regions pred: l_shipdate >= date ' ' l_shipdate < date ' ' + interval '1' year l_discount between and l_quantity < 24 predlist: AND ${pred}

32 SQLscalpel prototype is up and running: Full-stack infrastructure Drivers for MonetDB, MonetDBlite, SQLite, Postgres Functional enhancements planned: Multi-party public/private projects Built-in forum to share the story

33

34 Divergence Q1: SELECT count(*) FROM nation WHERE nation.n_name='brazil' Q2: SELECT count(*) FROM nation WHERE nation.n_name='brazil AND nation.n_regionkey=1 T A (Q 2 ) T B (Q 2 ) T A (Q 1 ) T B (Q 1 ) T A (Q 2 ) T B (Q 2 ) T B (Q 1 ) T A (Q 1 )

35 Divergence Q1: SELECT count(*) FROM nation WHERE nation.n_name='brazil' Q2: SELECT count(*) FROM nation WHERE nation.n_name='brazil AND nation.n_regionkey=1 < 1 System A is better than B on Q 2 against Q1 T A (Q 2 ) T B (Q 2 ) T B (Q 1 ) T A (Q 1 ) = 1 > 1 System B is better than A on Q 2 against Q1