Database Partitioning Options DATABASE SHARDING

w主写从读、集群节点间时时内存复制、单表横切纵切、分析报表系统通过服务器联表

http://www.agildata.com/database-sharding/

Database Partitioning Options

It has long been known that database partitioning is the answer to improving the performance and scalability of relational databases. Many techniques have been evolved, including:

• Master/Slave: This is the simplest option used by many organizations, with a single Master server for all write (Create Update or Delete, or CRUD) operations, and one or many additional Slave servers that provide read-only operations. The Master uses standard, near-real-time database replication to each of the Slave servers. The Master/Slave model can speed overall performance to a point, allowing read-intensive processing to be offloaded to the Slave servers, but there are several limitations with this approach:
  • The single Master server for writes is a clear limit to scalability, and can quickly create a bottleneck.
  • The Master/Slave replication mechanism is “near-real-time,” meaning that the Slave servers are not guaranteed to have a current picture of the data that is in the Master. While this is fine for some applications, if your applications require an up-to-date view, this approach is unacceptable.
  • Many organizations use the Master/Slave approach for high-availability as well, but it suffers from this same limitation given that the Slave servers are not necessarily current with the Master. If a catastrophic failure of the Master server occurs, any transactions that are pending for replication will be lost, a situation that is highly unacceptable for most business transaction applications.
• Cluster Computing: Cluster computing utilizes many servers operating in a group, with shared messaging between the nodes of the cluster. Most often this scenario relies on a centralized shared disk facility, typically a Storage Area Network (SAN). Each node in the cluster runs a single instance of the database server, operating in various modes:
  • For high-availability, many nodes in the cluster can be used for reads, but only one for write (CRUD) operations. This can make reads faster, but write transactions do not see any benefit. If a failure of one node occurs, then another node in the cluster takes over, again continuing to operating against the shared disk facility. This approach has limited scalability due to the single bottleneck for CRUD operations. Even the reads will ultimately hit a performance limit as the centralized shared disk facility can only spread the load so much before diminishing returns are experienced. The read limitations are particularly evident when an application requires complex joins or contains non-optimized SQL statements.
  • More advanced clustering techniques rely on real-time memory replication between nodes, keeping the memory image of nodes in the cluster up to date via a real-time messaging system. This allows each node to operate in both read or write mode, but is ultimately limited by the amount of traffic that can be transmitted between nodes (using a typical network or other high-speed communication mechanism). Therefore, as nodes are added, the communication and memory replication overhead increases geometrically, thus hitting severe scalability limits, often with a relatively small number of nodes. This solution also suffers from the same shared disk limitations of a traditional cluster, given that a growing, single large database has increasingly intensive disk I/O.
• Table Partitioning: Many database management systems support table partitioning, where data in a single large table can be split across multiple disks for improved disk I/O utilization. The partitioning is typically done horizontally (separating rows by range across disk partitions), but can be vertical in some systems as well (placing different columns on separate partitions). This approach can help reduce the disk I/O bottleneck for a given table, but can often make joins and other operations slower. Further, since the approach relies on a single server instance of the database management system, all other CPU and memory contention limitations apply, further limiting scalability.
• Federated Tables: An offshoot of Table Partitioning is the Federated Table approach, where tables can be accessed across multiple servers. This approach is necessarily highly complex to administer, and lacks efficiency as the federated tables must be accessed over the network. This approach may work for some reporting or analytical tasks, but for general read/write transactions it is not a very likely choice.

The common drawback with each of these approaches is the reliance on shared facilities and resources. Whether relying on shared memory, centralized disk, or processor capacity they each suffer with scalability limitations, not to mention many other drawbacks, including complex administration, lack of support for critical business requirements, and high availability limitations.