Database Replication Using Generalized Snapshot Isolation

1
Database Replication Using Generalized Snapshot
Isolation

• Sameh Elnikety, EPFL
• Fernando Pedone, USI
• Willy Zwaenepoel, EPFL

2
Snapshot Isolation (SI)

• Snapshot committed state of database
• On begin
• Snapshot(T) latest snapshot at start(T)
• On read or write operation
• T reads from and writes to its snapshot
• On commit
• Read-only T commits immediately
• Update T commits if no conflicting writes between
  its start commit times

3
Advantages of SI

• Read-only Ts never block or abort
• Read-only Ts never cause update Ts to block or
  abort
• Compare to 2PL
• No read-locks are used in SI
• Important for read-dominated workloads

4
Drawbacks of SI

• Not serializable
• Permits certain anomalies
• But
• Anomalies are rare in practice
• Conditions on workload can identify and avoid
  them
• Developers use SI serializably

5
Summary of SI

• SI is here to stay
• Used in several databases, e.g.,
• Oracle
• PostgreSQL
• Microsoft SQL Server ( 2PL SI )
• Borland InterBase

6
SI Replication

• Replicate SI to scale performance for dynamic
  content Web servers
• E.g., E-commerce, bulletin boards
• Workload is suitable for SI
• Read-only Ts dominate workload
• Update Ts are short few
• How to maintain SI properties?

7
SI in Replicated Database

• On begin
• Snapshot(T) latest snapshot at start(T)
• On read or write operation
• T reads from and writes to its snapshot
• On commit
• Read-only T commits immediately
• Update T commits if no conflicting writes between
  its start commit times

8
Strict SI in Replicated Database

• On begin
• Snapshot(T) latest snapshot at start(T)
• On read or write operation
• T reads from and writes to its snapshot
• On commit
• Read-only T commits immediately
• Update T commits if no conflicting writes between
  its start commit times

9
Generalized Snapshot Isolation (GSI)

• On begin
• Snapshot(T) (latest) older snapshot
• At replica, use latest local snapshot
• On read or write operation
• T reads from and writes to its snapshot
• On commit
• Read-only T commits immediately
• Update T commits if no conflicting writes between
  its (start) snapshot commit times

10
Generalized Snapshot Isolation (GSI)

• On begin
• Snapshot(T) (latest) older snapshot
• At replica, use latest local snapshot
• On read or write operation
• T reads from and writes to its snapshot
• On commit
• Read-only T commits immediately
• Update T commits if no conflicting writes between
  its (start) snapshot commit times

Certification for update T
11
Advantages of GSI

• All Ts reads and writes are local
• Important for replicated databases
• Read-only Ts never block or abort
• Read-only Ts never cause update Ts to block or
  abort
• Important for read-dominated workloads

12
A - GSI Serializability

• Not serializable
• Permits certain anomalies as in SI
• But
• Anomalies are rare in practice
• Two serializability conditions (in the paper)
• Static examine transaction templates
• Dynamic at run time
• Easy to verify workload is serializable
• Easy to modify workload to be serializable

13
A - GSI Serializability

• Not serializable
• Permits certain anomalies as in SI
• But
• Anomalies are rare in practice
• Two serializability conditions (in the paper)
• Static examine transaction templates
• Dynamic at run time
• Easy to verify workload is serializable
• Easy to modify workload to be serializable

Similar to what many Oracle DBAs already do
14
B - GSI Older Snapshots

• GSI uses older snapshots
• But
• Clear definition, always consistent data
• No new anomalies ( same as in SI )
• In replicated database
• Transparent db appears as running SI
• Efficient reads are non-blocking
• Staleness can be bounded

1- On begin Snapshot(T) (latest) older
snapshot
15
C - GSI Abort Rates
3- On commit - Read-only T commits
immediately - Update T commits if no
conflicting writes between its (start)
snapshot commit times

• Potentially higher abort rate for updates
• But
• Abort rates are small in target workloads
• GSI Abort rates can be higher or lower

Certification for update T
16
GSI in Replicated Databases

• System consists of
• Many SI replicas, full replication
• Centralized certifier ( distributed in the paper
  )
• A client connects to one replica
• Issues read and update transactions
• Algorithm implements an instance GSI
• Snapshot(T) latest local snapshot at replica

17
Algorithm at Replica

• On begin
• Provide T with a local Snapshot
• Record T.version Snapshot.version
• On read or write operation
• Run transaction (reads/writes) locally
• Record T.writeset
• On commit
• IF ( T is read-only ) THEN commit
• ELSE Invoke certification ( T.version,
  T.writeset ) . . .

18
Algorithm at Certifier

• Check for conflicting writes from committed Ts
  with larger version number
• IF ( yes ) THEN Reply ( abort )
• ELSE Advance certifier-version
• Record (writeset, certifier-version) to
  log Reply ( 1 - commit,
  2 - certifier-version, 3 - missing
  writesets )

19
Algorithm at Replica (cont.)

• On begin
• . . .
• On read or write operation
• . . .
• On commit
• IF ( T is read-only ) THEN commit
• ELSE Invoke certification (T.version,
  T.writeset ) 1- Apply missing writesets 2-
  Commit locally 3- Advance local version

20
Performance Tradeoff GSI SI

• GSI
• better response time
• SI
• fresher data (latest snapshot in the system)
• lower abort rate for updates (?)
• Analytical performance model
• Model used by Jim Gray
• Replicated database over WAN

21
Analytical Model

• GSI
• Execute T immediately
• Updates are certified remotely (communication)
• SI
• Block T to obtain latest version (communication)
• Updates are certified remotely (communication)
• Objective is to compare GSI SI
• Response time
• Abort rate

22
Analytical Equations

• Parameters
• x round trip delay / transaction length
• Response time ratio (GSI SI)
• Read-only update

23
Analytical Equations

• Parameters
• x round trip delay / transaction length
• t snapshot age / transaction length
• Response time ratio (GSI SI)
• Read-only update
• Abort rate ratio (GSI SI)
• Read-only (never aborted!) update

24
Analytical Results

• Parameters
• x round trip delay / transaction length
• t snapshot age / transaction length
• X-axis
• x round trip delay / transaction length
• x 0 ? centralized database
• x is increasing as technology advances
• Y-axis
• Response time ratio (for reads updates)
• Abort ratio (updates)

25
Response Time Ratio of GSI SI

• .

GSI is better
26
Abort Ratio of GSI SI for Updates

• .

SI better
GSI better
Parameter t ( snapshot age / transaction length
)
27
Abort Ratio of GSI SI for Updates

• .

t decreasing fresher snapshot
SI better
GSI better
Parameter t ( snapshot age / transaction length
)
28
GSI SI - Summary

• GSI response times are better
• Read-only Ts ratio significantly better
• Update Ts ratio reaches ½
• GSI abort rate
• maybe higher or lower
• COST observing older data in GSI
• Favorable trade-off
• Distributed environments
• Read-dominated workloads

29
Conclusions

• GSI is appealing for replication
• All Ts read write operations are local
• Read-only Ts never block or abort
• GSI can be made serializable
• Algorithm for GSI in replicated databases
• Analytical results are encouraging