Serializability with Snapshot Isolation under the Hood...Serializability with Snapshot Isolation under the Hood Mihaela Bornea 1, S. Elnikety 2, O. Hodson 2, A Fekete 3 1IBM Research

Outline Motivation Concurrency Control Replication Model Readset Certification Evaluation Conclusions

Serializability with Snapshot Isolation underthe Hood

Mihaela Bornea 1, S. Elnikety 2, O. Hodson 2, A Fekete 3

1IBM Research 2Microsoft Research 3University of Sydney

M. Bornea, S. Elnikety, O. Hodson, A. Fekete IBM Research, Microsoft Research, Microsoft Research, University of Sydney

Serializability with Snapshot Isolation under the Hood


Motivation

Concurrency Control

Replication Model

Readset Certification

Evaluation

Conclusions




Transaction Processing in Replicated Databases

I Database Replication:I Higher availability & better performanceI Maintaining consistency is challenging

I State of the Art:I GSI Replicated Databases.I Each replica uses Snapshot Isolation (SI).

I Goal:I Global One Copy Serializability.I Overall Isolation level stronger than the one of individual

components.I The replicated system keeps its performance.




Transaction Isolation

I Isolation is a correctness criterion.I Concurency in the system.I Multiple levels of isolation:

I Snapshot Isolation.I Serializability.




Snapshot Isolation

I Multi-version concurency control technique.I Important

I Used by Oracle, SQL Server, Postgres.I Sometimes the strongest isolation level available.

I Attractive performanceI Read-only transactions never block or abort.I Read-only transactions do not block update transactions.I Updates might abort. Certification needed.

I checks for ww conflicts.




Anomaly under SI

Time

R(X,Y)

R(X,Y) W(X) C

CT2

T1

Y=50

X=50

X=50

Y=50

X=−40

Y=−40

W(Y)

I X,Y balance of two bank accounts.I T1 and T2 withdraw 90E from X and YI Logic: X + Y > 0




Serializability

I The strongest DB isolation level.I Illusion that transactions execute serially.I Programmers want it:

I As if there is no concurrency.I Commonly implemented with 2PL.

I expensive to achieve.




Serializability under SI

I Centralized DatabaseI Modify database engine, SSI.I Use Fekete’s work [SIGMOD 2008, best paper]

I Replicated DatabasesI Open question.I No modification of the database engine.




GSI Replicated Database




SQL Transaction ModelA. SELECT expr_list FROM Ri WHERE pred(Ri )B. INSERT INTO Ri VALUES (values)C. UPDATE Ri SET attr_values WHERE pred(Ri )D. DELETE FROM Ri WHERE pred(Ri )E. SELECT agg(attr) FROM Ri WHERE pred(Ri )GROUP BY group_attrHAVING pred(agg(attr))

F. SELECT attr_listFROM R1...Ri ...RnWHERE pred(R1) LOP ...LOP pred(Ri ) LOP ... LOP pred(Rn)

LOP pred(attri,j , attri,j )G. SELECT attr_list

FROM R1...Ri ...Rn,SQWHERE pred(R1) LOP ...LOP pred(Ri ) LOP ... LOP pred(Rn)

LOP pred(SQ)H. SELECT attr_list

FROM R1...Ri ...RnWHERE pred(R1) LOP ...LOP pred(Ri ) LOP ... LOP pred(Rn)

LOP pred(attri ,SQ)




1SR Needs Readsets

I Snapshot Isolation (SI) → Generalized Snapshot Isolation(GSI)

I Certify WritesetI Serializability → One Copy Serializability (1SR)

I Certify WritesetI Certify Readset

I Yes, we have a proof :) !




Writesets

I The Writeset contains modified tuplesI Introduced by UPDATE, INSERT and DELETEI Includes both new and old tuple valuesI All Writesets are managed at the Certifier.I Writeset certification is required by both GSI and 1SR

I checks if concurrent transactions modify the same item.I It is well knows how to manage the Writesets




Readsets

I The Readset contains read tuples.I Introduced by SELECT, UPDATE, INSERT and DELETE.I Readsets certification is required by 1SR.

I checks if a transaction reads data modified by concurrenttransactions.

I Readset identification is challenging:I never done in replicated setting.




So far ...

I We introduced SI.I Sometimes SI is not enough !I Serializability needed:

I Keep the nice properties of SI.I Open Problem for replicated databases:

I Readset management is difficult!




Main Contribution - Readset Management

I Framework to manage the ReadsetsI Observation: each SQL statement has a predicate.

I The Readset is a list of predicates.I Readset certification requires predicate evaluation.




Certifier Design

I The Certifier manages:I persistent log.I main memory database, CertDB.

I The log is used for durability.I CertDB is used to certify update transactions.I CertDB maintains the Writeset of recently committed

transactions.I CertDB schema:

I the replicated schema.I commit version attribute.




Readset Certification

I Intuition:I Ensures that if the transaction executes on the latest

version it would read the same values.I Implementation:

I Replica identifies the Readset:I Extracts the predicate of each SQL statement.

I Replica expresses the readset as certification queries.I The certification queries are evaluated on CertDBI Empty conflict set indicates serializable execution




Concurrent Transactions

I Snapshot versions at originating replicas.I Commit version of a transaction.I CertDB contains the writesets and committed version.I Consider a transaction T:

I version > snapshot(T )




Readset for SELECT Statements

Transaction Queries

A. SELECT expr_list FROM Ri WHERE pred(Ri)

Certification Queries

A. SELECT * FROM Ri WHERE pred(Ri) ANDversion > snapshot(T )




Readset for UPDATE Statements

Transaction Queries

B. INSERT INTO Ri VALUES (values)C. UPDATE Ri SET attr_values WHERE pred(Ri)D. DELETE FROM Ri WHERE pred(Ri)


B. SELECT * FROM Ri WHERE pk = @pk ANDversion > snapshot(T )

C. SELECT * FROM Ri WHERE pred(Ri) ANDversion > snapshot(T )

D. SELECT * FROM Ri WHERE pred(Ri) ANDversion > snapshot(T )

Certifying the Readset also detects ww conflicts.




Experimental Study

I Impact of providing 1SR vs. GSI:I Lower throughput and higher response timeI Higher abort rate

I Replicated system with 8 replicasI TPC-W




Workload

I TPC-W benchmark:I Web application (online book store).I Database schema consists of 10 tables.I Database size: 800 MB.I 13 transaction templates.I Ordering Mix(50% updates).I Browsing Mix (5% updates).

I Metrics:I Transactions per minute (TPM).I Response time.I Abort rate.




Scaling of SGSI with Replication Degree

0

10000

20000

30000

40000

50000

60000

0 150 300 450 600 750 900 1050

Th

rou

gh

pu

t, T

PM

Number of Clients

1-replica SGSI

1-replica GSI

2-replica SGSI

2-replica GSI

4-replica SGSI

4-replica GSI

8-replica SGSI

8-replica GSI

Throughput of TPC-WShopping Mix (20% updates)

0

500

1000

1500

2000

2500

3000

3500

4000

4500

0 150 300 450 600 750 900 1050

Re

sp

on

se

Tim

e,

ms

Number of Clients

1-replica SGSI

1-replica GSI

2-replica SGSI

2-replica GSI

4-replica SGSI

4-replica GSI

8-replica SGSI

8-replica GSI

Resp. Time of TPC-WShopping Mix (20% updates)




Comparing SGSI to GSI

0

50

100

150

200

250

300

350

400

10000 20000 30000 40000 50000

Re

sp

on

se

Tim

e,

ms

Throughput, TPM

8-replica GSI

8-replica SGSI

Scalability of TPC-W ShoppingMix (20% updates)

0

20

40

60

80

100

10000 20000 30000 40000 50000

Ce

rtif

ier

CP

U U

tili

za

tio

n,

%

Throughput, TPM

8-replica GSI

8-replica SGSI

Certifier CPU Utilization TPC-WShopping Mix (20% updates)




Sensitivity to Update Transaction Ratio

0

10000

20000

30000

40000

50000

60000

70000

80000

90000

0 250 500 750 1000 1250 1500

Th

rou

gh

pu

t, T

PM

Number of Clients

8-replica TPC-W Browsing Mix

8-replica TPC-W Shopping Mix

8-replica TPC-W Ordering Mix

SGSI Throughput of TPC-WMixes.

0

100

200

300

400

500

600

700

800

900

0 250 500 750 1000 1250 1500

Re

sp

on

se

Tim

e,

ms

Number of Clients

8-replica TPC-W Browsing Mix

8-replica TPC-W Shopping Mix

8-replica TPC-W Ordering Mix

SGSI Response Time ofTPC-W Mixes.




Abort Analysis via SmallBank

0

2

4

6

8

10

12

14

10 25 50 100 150 200 250 300

Co

nfl

icts

, %

Multi-Programming Level

ww conflictsrw conflicts

pivots




Conclusions

I We introduced SGSI:I 1SR in replicated databases.

I Built a replicated system prototype.I Evaluated SGSI performance:

I SGSI is practical.I Moderated cost for small degree of replication.I Performance and scaling is comparable with GSI.




Readset for Joins

Transaction Queries

F. SELECT attr_listFROM R1...Ri ...Rn

WHERE pred(R1) LOP ...LOP pred(Ri) LOP ... LOP pred(Rn)LOP pred(attri,j , attri,j)


for each relation Ri

F. SELECT * FROM Ri WHERE version > snapshot(T )

I An upper-set of the Readset is certified.I False aborts.




Data Managed at Certifier

I Accuracy depends the data maintained at the Certifier.I False aborts:

I not enough information to evaluate the ReadsetI Solution:

I manage a copy of relations at the Certifier.I physical design tuning problem.




Extended CertDB

I Each data item has several instances.I New instace: UPDATE,INSERT.I Expired: UPDATE, DELETE.I Each copy relation is augmented with VStart and VEnd .I VStart and VEnd determine:

I update predicate: upd(Ri).I visibility predicate: vis(Ri).




Extended Certification

Transaction Queries

F. SELECT attr_listFROM R1...Ri ...Rn

WHERE pred(R1) LOP ...LOP pred(Ri) LOP ... LOP pred(Rn)LOP pred(attri,j , attri,j)


SELECT * FROM R1C ...RiC ...RnC

WHERE (query_pred)AND (upd(R1C ) ...OR upd(RiC ) ... OR upd(RnC ))AND (vis(R1C ) ...AND vis(RiC ) ... AND vis(RnC ))



Documents

Serializability with Snapshot Isolation under the Hood...Serializability with Snapshot Isolation under the Hood Mihaela Bornea 1, S. Elnikety 2, O. Hodson 2, A Fekete 3 1IBM Research