research faculty summit 2018...through large-scale machine learning sigmod 2017 165 508 562 736 686...

Systems | Fueling future disruptions

ResearchFaculty Summit 2018

@andy_pavlo[Image Source]

Part #1Part #2Part #3

– Background– Engineering– Oracle Rant

AUTONOMOUS DBMSs 4SELF-ADAPTIVE DATABASES

1970-1990sSelf-Adaptive

Databases

SELECT * FROM A JOIN BON A.ID = B.ID

WHERE A.VAL > 123AND B.NAME LIKE 'XY%'

A.IDA.VALB.ID

B.NAME

+100 +200 +50

TuningAlgorithm

Admin

→Index Selection→Partitioning / Sharding→Data Placement

541

291

0

200

400

600

2000 2004 2008 2012 2016

Number of Knobs

AUTONOMOUS DBMSs 5SELF-TUNING DATABASES

1990-2000sSelf-TuningDatabases

SELECT * FROM A JOIN BON A.ID = B.ID

WHERE A.VAL > 123AND B.NAME LIKE 'XY%'

TuningAlgorithm

Admin

→Index Selection→Partitioning / Sharding→Data PlacementAutoAdmin

A.IDA.VALB.ID

B.NAME

→Knob Configuration

AUTONOMOUS DBMSs 6CLOUD MANAGED DATABASES

2010sCloud

Databases

→Initial Placement→Tenant Migration

Why is this Previous WorkInsufficient?

AUTONOMOUS DBMSs 8A BRIEF HISTORY

Problem #1Human

Judgements

Problem #2ReactionaryMeasures

Problem #3No Transfer

Learning

What is Different this Time?

CARNEGIE MELLON UNIVERSITY 10RESEARCH PROJECTS

OtterTuneExistingSystems

PelotonNew

System

OtterTune

Database Tuning-as-a-Service→Automatically generate

DBMS knob configurations.→Reuse data from previous

tuning sessions.

ottertune.cs.cmu.edu

OTTERTUNETPC-C TUNING

AUTOMATIC DATABASE MANAGEMENT SYSTEM TUNING THROUGH LARGE-SCALE MACHINE LEARNINGSIGMOD 2017

165

508562

736686

0

250

500

750

1000

426

845

714

843

946

0

250

500

750

1000Throughput (txn/sec)

Default RDS DBAScripts OtterTune

12

Peloton

Self-Driving Database System→ In-memory DBMS with

integrated planning framework.

→Designed for autonomous operations.

pelotondb.io

Design Considerations forAutonomous Operation

AUTONMOUS DBMS 15DESIGN CONSIDERATIONS

ConfigurationKnobs

InternalMetrics

ActionEngineering

Anything that requires a human value judgement should be marked as off-limits to autonomous components.

– File Paths / Network Info– Durability / Isolation Levels– Hardware Usage– Recovery Time

16UNTUNABLE KNOBSCONFIGURATION KNOBS

CONFIGURATION KNOBS

The autonomous components need hints about how to change a knob.

– Min/max ranges.– Separate knobs to enable/disable a feature.– Non-uniform deltas.

17HOW TO CHANGE

1 KB 1 MB 1 GB 1 TB+100 KB +100 MB +100 GB

INTERNAL METRICS

If the DBMS has sub-components that are tunable, then it must expose separate metrics for those components.

Bad Example:

18SUB-COMPONENTS

INTERNAL METRICS 19SUB-COMPONENTS

RocksDB Column Family Knobs

Column Family Metrics

Missing:ReadsWrites

Global Metrics

AggregatedMetrics

ACTION ENGINEERING

No action should ever require the DBMS to restart in order for it to take affect.The commercial systems are much better than this than the open-source systems.

20NO SHUTDOWN

ACTION ENGINEERING

Allow replica configurations to diverge from each other.

21REPLICA EXPLORATION

Master

Replicas

What About Oracle'sSelf-Driving DBMS?

ORACLE

Automatic Indexing

Automatic Recovery

Automatic Scaling

Automatic Query Tuning

SELF-DRIVING DBMS23

September 2017 January 2017

Problem #2React ionary

Measures

CONCLUSION

True autonomous DBMSs are achievable in the next decade.You should think about how each new feature can be controlled by a machine.

MAIN TAKEAWAYS24

@andy_pavlo

OTTERTUNE 27AUTOMATIC DBMS TUNING SERVICE

TARGETDATABASE

TUNING MANAGERCONTROLLERCOLLECTOR

Configuration Recommender

KnobAnalyzer

MetricAnalyzer

5001000150020002500

200400

600800

99th

%-ti

le(s

ec)

0.00.51.01.52.0

InternalRepository

INSTALL AGENT

"THE BRAIN"PELOTON 28THE SELF-DRIVING DBMS

FORECAST MODELS

Search Tree

ACTIONCATALOG

ACTION SEQUENCETARGET

DATABASE

WORKLOADHISTORY ?

??

QUERY-BASED WORKLOAD FORECASTING FOR SELF-DRIVING DATABASE MANAGEMENT SYSTEMSIGMOD 2018

PELOTONBUS TRACKING APP WITH ONE-HOUR HORIZON

QUERY-BASED WORKLOAD FORECASTING FOR SELF-DRIVING DATABASE MANAGEMENT SYSTEMSIGMOD 2018

0

15000

30000

45000

60000

9-Jan 11-Jan 13-Jan 15-Jan 17-Jan

Ensemble (LR+RNN)

29

Actual Predicted

Que

ries

Per H

our

ACTION ENGINEERING

Provide a notification callback to indicate when an action starts and when it completes.Harder for changes that can be used before the action completes.

30NOTIFICATIONS

Thank you!