Using SQL and PLSQL for Mid-Tier Database Caching · 2019. 5. 20. · –Standard SQL / PLSQL –Entire database in DRAM Persistent and Recoverable –Database and Transaction logs

Copyright©2016, Oracleand/oritsaffiliates.Allrightsreserved.|

UsingSQLandPLSQLforMid-TierDatabaseCaching

DougHood@ScalableDBDougConsultingMemberofTechnicalStaffProductManagerTimesTen In-MemoryDatabaseMay16,2019


Agenda

BusinessProblem

DevOpschallenge

Architecturesandtechnicalchallenges

Howtodoit

Demo,summaryandQ&A

2

1

2

3

4

5


Latency,ThroughputandScalabilityLatency Howquicklycanoneoperationcomplete

Onesprinterin9.58 seconds~40km/hfor100M[2009]

Throughput Howquicklycanmanyoperationscomplete

Tensprintersinunder11seconds~40km/hfor100m[2009]

Scalability Byaddingmoreresourcescanthroughput keepincreasing

33carson2.5mileovaltrack~250km/hfor804km[Indy500,2017]

Confidential– OracleInternal/Restricted/HighlyRestricted 3


LowerLatencywithTimesTen Cache


Oracle11.2.0.4RACRACnodeswereOracleSunX7-2LNVMe StorageOver50MillionUsers

LatencyisinMicroSeconds…

ApplicationTierDatabaseCache(TimesTen)RanonthesamenodesastheproductionRAC5tablejoinsfor100sofmillionsofrowsofdata

Query Oracle Cache

Q1 43 3

Q2 69 6

Q3 105 8

Q4 121 20

Q5 140 18

Q6 163 19

Q7 231 18


WhyCacheData?



WhyCacheData?

• TogetlowerlatencyforSQLstatements• Togetmorethroughput• Togetmorescalability



CachingChallenges



TheBusinessProblem

• DBapps,makethemfasterandcheaper• Domorewithless



TheDevOpsChallenge

• Makeitboring,nosurprises• Standardized,runinVMsorcontainers• Everything isautomated• Figureouteverythingthatcangowrongandbewellbehaved




• Howmanyusers?Howmanyconcurrentusers?• HowmanyconcurrentOracleDBconnections?• HowmanyconcurrentApplicationServerconnections?• WhataboutHA?WhataboutDR?• Whataboutlatency?





Canwedobetter?




Connectionpool

Canwedobetter?




Connectionpool

SharedServers(MTS)

Canwedobetter?




Connectionpool

Canwedobetter?

Read/writecaches• Offloadwork fromOracle• Offloadconnections fromOracle

SharedServers(MTS)


Architecturesandtechnicalchallenges• Gofaster thanNoSQLsolutions&minimize thehardwarecost– Lowerlatencyatthe99th percentile



Architecturesandtechnicalchallenges• Gofaster thanNoSQLsolutions&minimize thehardwarecost– Lowerlatencyatthe99th percentile


• Simpleandscalableisgood,buthowmanymachinesdoyouneed?• Whataboutthe95th and99th percentileforlatency?• Whataboutcachemisses?• Howisthecachekeptuptodate?

99th percentilelatencycanbeashighas324ms…

100th percentilelatencyismanyseconds,timeout>100ms


YCSBWorkloadB(95%read,5%write)


Database TPS Nodes

Cassandra 221K 32

MongoDB 260K 2

CouchBase 454K 9

Redis 1M 3

VoltDB 1.5M 6

AeroSpike 1.6M 1


YCSBWorkloadB(95%read,5%write)


Database TPS Nodes

Cassandra 221K 32

MongoDB 260K 2

CouchBase 454K 9

Redis 1M 3

VoltDB 1.5M 6

AeroSpike 1.6M 1

TimesTen 2.8M 1 Letstalk




Connectionpool

Canwedobetter?



CustomermeasuredTimesTenCachelatency

• 97%cachehitratio– 3%ofthetimethedatawasnotinthecache,soneededaroundtriptoOracle

• 99th percentilelatency=1ms







Percentile Latencyinms

87 0.016

98 0.125

99.7 1

99.99 8

100 423

NoSQLwas324ms






Percentile Latencyinms

87 0.016

98 0.125

99.7 1

99.99 8

100 423Workingtoimprovethis• Goalis<50msfor100th percentile• Alsodefinea300ms timeout


Improvingthe100th percentile[maxlatency]


SQLNetConnection

pool

Canwedobetter?


TimesTenConnection

pool

Forcachemisses,createanewSQLNetconnection!

Thiscanbereallyslow…


Improvingthe100th percentile[maxlatency]


SQLNetConnection

pool

Canwedobetter?


SQLNetConnection

pool

TimesTenConnection

pool

Forcachemisses,getconnectionfromTTOCICPratherthanthesharedpool

Tunethepoolfortheexpectedmin/max


Howisitpossible



Howisitpossible

• Useareallyfast In-memorySQLRDBMS



Howisitpossible

• Useareallyfast In-memorySQLRDBMS• UseaCacheDB thatjustrequiresconfiguration


T1

T3

T2


Howisitpossible

• Useareallyfast In-memorySQLRDBMS• UseaCacheDB thatjustrequiresconfiguration• Somehardwareandsoftwaretuning


T1

T3

T2

• Prepare&bindSQLstatements• Usetherelevantindexes• Dosensiblejoins• UpdatestatisticsforSQLoptimizer• CheckTimesTenSQLstatsforSQLandIObottlenecks


MostWidelyUsedRelationalIn-MemoryDatabaseDeployedbyThousandsofCompanies

29


RelationalDatabase– Purein-memory– ACIDcompliant– StandardSQL/PLSQL– EntiredatabaseinDRAM

PersistentandRecoverable– DatabaseandTransactionlogspersistedonlocaldiskorflashstorage

– ReplicationtostandbyandDRsystems

30

ExtremelyFast– Microsecondsresponsetime– Veryhighthroughput

HighlyAvailable– Active-Standbyandmulti-masterreplication

– Veryhighperformanceparallelreplication

– HAandDisasterRecovery

OracleTimesTenIn-MemoryDatabase


LowLatency- Microseconds ResponseTimePerformance– ResponseTime

Millionthsofa

Second

Millionthsofa

Second

31

[email protected]

2socket,22cores/socket,2threads/core

TimesTen11.2.2.8.0(100Mrows,17GBdata)

1.64

5.06

0.00

1.00

2.00

3.00

4.00

5.00

6.00

SELECTQuery UPDATETransaction

Microsecond

s


HLRMobileTransactionsResponseTimeResponseTimeImprovementWithTimesTenApplication-TierDatabaseCache

Intel®[email protected]/socket2hyper-threads/core32vCPU

OracleDatabaseWithTimesTenCacheOracleDatabaseWithoutTimesTenCache

Speedup

17x

10x

17x

44x

69x

31x

29x


TimesTenIn-MemoryDatabasePersistenceandRecovery

• TimesTendatabasepersistencecanbeconfiguredon– Flash,SSD,harddiskstorage

• Alltransactionsareloggedandpersisted– Redo,undo,andrecovery

• DualdatabaseCheckpointfiles– Databaserestart– Databaserecoveryafterfailures

TimesTen Checkpoint

Files

TimesTen Transaction

Log Files


CacheDB =OracleApplicationTierDatabaseCacheAfeatureofOracleEnterpriseEdition

34

T1

T3

T2T1

T3

T2

CacheAgent

RepAgent

Metadata• Cachegroups

Batchreadstorefreshthecache

Parallelbatchwrites


FlexibleCacheGroupConfigurations

• CacheGroupdescribestheOracleDatabasetablestocache– Allorsubsetofrowsandcolumns– DefinedusingSQLCREATECACHEGROUPPremierUsersFROMOE.CUSTOMER(NAMEVARCHAR2(100)NOTNULL,ADDRVARCHAR2(100)

)WHEREOE.CUSTOMER.ORDER>500;

• CachetablesareregulartablesinTimesTen– Joins/search,insert/update/delete

ApplicationTransactions CacheGroups

AutomaticDataSynchronization







createdynamicreadonly cachegroupmyReadCache1autorefreshmodeincrementalinterval 100millsecondsstateonfrom hr.employees (employee_id number(6)notnull,

first_name varchar2(20),last_name varchar2(25)notnull,hire_datedatenotnull,job_id varchar2(10)notnull,salary number(8,2),manager_id number(6),department_id number(4),

primarykey(employee_id)),hr.job_history (employee_idnumber(6)notnull,

start_datedatenotnull,end_datedatenotnull,job_idvarchar2(10)notnull,department_idnumber(4),

primarykey (employee_id,start_date),foreignkey (employee_id)references hr.employees(employee_id));


OracleTimesTen• Onetechnology,twoproducts,threedeploymentmodes

ApplicationTierDatabaseCache

StandaloneIn-MemoryDatabaseforOLTP

andanalytics

Application

TimesTenScaleout

SingleSystemImageIn-MemoryDatabase

OracleTimesTenIn-MemoryDatabase

Application


Summary• CloudScaleread/writecachingwith1mslatencyat99th percentile• OracleTimesTenisfasterthananyNoSQLDB• Cachingisconfiguration,notcoding• UseTimesTen PLSQLstoredprocedurestominimizenetworkhops

• WriteTimesTenappsasifitwereanOracleDB*– SQL,JDBC,PLSQL,OCI,ODBC,ODP.Net,Pro*C,Pro*COBOL– R,Python,Node.js,Go,Ruby&PHP

Subset oftheOracleSQLandPLSQL

Using SQL and PLSQL for Mid-Tier Database Caching · 2019. 5. 20. · –Standard SQL / PLSQL –Entire database in DRAM Persistent and Recoverable –Database and Transaction logs

Documents