Centiman : Scalable High Performance Transaction Processing

1. Centiman: Scalable High Performance Transaction Processing Bailu Ding Cornell University 2013

2. Why Large Scale TP is Important? • Internet scale web applications • Multi-tenancy in the cloud

3. What Is a Transaction? • Transaction is a sequence of read / write operations executed in a database • ACID properties • Atomicity: no partial completion • Consistency: bring the database from one valid state to another • Isolation: concurrent execution of transactions results in a database state as they were executed serially (serializability) • Durability: committed transactions will remain in the database in case of failures

4. Why Is Large Scale TP Hard • Lots of data • Scale beyond a single machine • Lots of transactions • High throughput • Strong consistency • Serializability • High availability

5. Current Solution: Key Value Store • Scalability and high availability • Not transactional • Weak consistency

6. Current Solution: Data Partitioning • Scalable • Strong consistency within a partition • Weak consistency across partitions • Strong consistency across partitions is expensive • Performance sensitive to choice of partition Jones, Evan et al. Low Overhead Concurrency Control for Partitioned Main Memory Databases. SIGMOD’10 Jason Baker et al. Megastore: Providing Scalable, Highly Available Storage for Interactive Services. CIDR’11

7. Current Solution: Google Spanner • Scalable • Strong consistency • Require special hardware: GPS or atomic clocks James Corbett et al. Spanner: Google's Globally-Distributed Database. OSDI’12

8. Our Goals • Large scale high performance transaction processing • Support strong consistency • No special hardware needed

9. Concurrency Control • Concurrency control: keep transactions isolated • Locking-based concurrency control • Optimistic concurrency control

10. Locking-based Concurrency Control Start T1(W(X)) T2(R(Y, X), W(X)) Obtain Locks Write Lock (X) Read Lock (Y) Release Locks and Commit Read / Execute / Write Read Lock (X) Release Locks and Commit Write Lock (X) Release Locks and Commit

11. Optimistic Concurrency Control Start T1(W(X)) T2(R(Y, X), W(X)) Read Local Write (X) Read (Y, X) Local Write (X) Validate Execute / Local Write Update and Commit Validate Validation Abort Update and Commit / Abort Retry (optional)

12. Centiman: Large Scale TP • Approach: optimistic concurrency control • Avoid overhead of locking • Centralized validation may become a bottleneck • Contribution: scalable validation • Parallel validation • Local timestamp generation • Watermark-based garbage collection

13. OCC: Centralized Validation Start Validator Read Processor Processor Execute / Local Write Processor Validation Storage Storage Storage Update and Commit / Abort

14. Parallel Validation Validator Validator Validator Client Client Processor Read / Update Database Read / Update Database Storage

15. Parallel Validation Vblue Vred Blue Items Red Items Processor T Storage

16. Experiment: TPC-C

17. Experiment: TATP

18. Future Work • Fault tolerance • Elasticity