1. H-store: A high-performance, distributed main memory transaction processing system
Robert Kallman, Hideaki Kimura, Jonathan Natkins, Andrew Pavlo, Alex Rasin, Stanley B. Zdonik, Evan P. C. Jones, Samuel Madden, Michael Stonebraker, Yang Zhang, John Hugg, Daniel J. Abadi

2. Paper highlights An experimental main-memory, parallel DBMS
Optimized for on-line transaction processing (OLTP) applications
Highly distributed, row-store-based relational database
Runs on a cluster on shared-nothing, main memory executor nodes.

3. Background

4. Relational DBMS Data are stored in tables
Each row corresponds to a record
No pointers or other links
Matching keys

5. Example PSID is a key Links rows among tables Must be unique Barbara
13339 CS CE
Student
PSID
Major
Alan
08887
PSID is a key
Links rows among tables
Must be unique
08887 89
13339 87 90
PSID
Mid
Final

6. Atomic Transactions: You buy a car

7. What are atomic transactions?
A mechanism used in databases and other financial systems.
System guarantees that an atomic transaction will either be executed properly or abort without leaving any trace
All or nothing semantics
Atomic transactions verify the four ACID properties

8. The ACID properties Atomicity: All or nothing property
Consistency: A transaction either brings the data in a new consistent state of data or returns them to their previous state
Isolation: A transaction in process and not yet committed has no effect on any other transaction.
Durability: Committed data are stored by the system in some kind of crash-proof storage

9. Importance of atomic transactions
Atomicity and consistency properties guarantee that either the transaction is correct or will leave no traces
No partial updates
No incorrect updates
Isolation property allows concurrent execution of transactions
Much faster than serial execution
Durability property ensures transactions will not be lost

10. Back to the paper

11. Motivation Legacy OLTP databases
Too many of their architectural components are old
Inherited from original System R
Mid-seventies!
Take advantage of recent trends
Multi-core architectures
Cheap abundant main memory
Dominant use of stored procedures

12. The focus Reject “one size fits all” approach
On-line transaction processing (OLTP) systems have specific proprieties
Repetitive short-lived transactions
Stored procedures
Sole focus of this work

13. Rely on replication to minimize the risk of data loss
Main issue
Poor I/O performance of RDBMS
Their solution
Scale system “horizontally”
Partition responsibilities among multiple shared nothing machines
Store entire DB in the memory of a large cluster of server machines
Rely on replication to minimize the risk of data loss

14. H-Store Next generation OLTP system
Operates on a distributed cluster of shared nothing machines
Coordinates the work of multiple single-threaded engines
All data are always kept in main memory

15. System Overview H-Store
Cluster containing two or more computational nodes
Nodes
Single physical component that holds multiple sites
Sites
Normally run on a dedicated core
Single-threaded
Do not share any data structure or memory with any collocated site

16. H-Store system architecture

17. System deployment Cluster deployment framework takes at inputs
A set of stored procedures
A database schema
A sample workload (used to optimize data layout)
A set of available sites in the cluster
Two-phase optimization
First optimize stored procedures as if the database was not distributed
Then come with distributed query plans

18. Run-time model All sites in the cluster are trusted
Any site is able to execute any OLTP application request
Execution plan is
Annotated with the locations of the target sites
Passed to a transaction manager
No shared data structures
Everything is single-threaded

19. Database properties
Physical layout of DB specifically optimized to execute precompiled transactions
Not ad hoc queries
Can still be executed but could be very slow

20. Transaction classes Two important special cases
Single-Site Transactions
Can be entirely executed on a single site
Easy to send the transaction to one of the target sites
One-Shot Transactions
Each of is individual queries executes on a single site
Output of these queries is not reused as inputs for other queries
Easy to execute in parallel

21. Physical layout
Replicate frequently-accessed or read-only tables on each site
Horizontal partition of tables
Partitions can be accessed in parallel
Collocate them with related data
Protect data against node failures
Important for in-memory DBs
k-safety
Number k of node failures DB must tolerate

22. DB layout loader Table Replication
Replicate all read-only tables on all sites
Data Partitioning
Divide horizontally each table into four disjoint partitions
Each partition is stored on two different sites
Accent is on parallelism
K-Safety
k = 2