Horizontal partitioning is a database design principle whereby rows of a database table are held separately, rather than being split into columns (which is what normalization and vertical partitioning do, to differing extents). Each partition forms part of a shard, which may in turn be located on a separate database server or physical location.
There are numerous advantages to this partitioning approach. Since the tables are divided and distributed into multiple servers, the total number of rows in each table in each database is reduced. This reduces index size, which generally improves search performance. A database shard can be placed on separate hardware, and multiple shards can be placed on multiple machines. This enables a distribution of the database over a large number of machines, which means that the database performance can be spread out over multiple machines, greatly improving performance. In addition, if the database shard is based on some real-world segmentation of the data (e.g., European customers v. American customers) then it may be possible to infer the appropriate shard membership easily and automatically, and query only the relevant shard.
In practice, sharding is far more complex. Although it has been done for a long time by hand-coding (especially where rows have an obvious grouping, as per the example above), this is often inflexible. There is a desire to support sharding automatically, both in terms of adding code support for it, and for identifying candidates to be sharded separately. Consistent hashing is one form of automatic sharding to spread large loads across multiple smaller services and servers.
Where distributed computing is used to separate load between multiple servers (either for performance or reliability reasons), a shard approach may also be useful.
Horizontal partitioning splits one or more tables by row, usually within a single instance of a schema and a database server. It may offer an advantage by reducing index size (and thus search effort) provided that there is some obvious, robust, implicit way to identify in which table a particular row will be found, without first needing to search the index, e.g., the classic example of the 'CustomersEast' and 'CustomersWest' tables, where their zip code already indicates where they will be found.
Sharding goes beyond this: it partitions the problematic table(s) in the same way, but it does this across potentially multiple instances of the schema. The obvious advantage would be that search load for the large partitioned table can now be split across multiple servers (logical or physical), not just multiple indexes on the same logical server.
Splitting shards across multiple isolated instances requires more than simple horizontal partitioning. The hoped-for gains in efficiency would be lost, if querying the database required both instances to be queried, just to retrieve a simple dimension table. Beyond partitioning, sharding thus splits large partitionable tables across the servers, while smaller tables are replicated as complete units.
This is also why sharding is related to a shared nothing architecture—once sharded, each shard can live in a totally separate logical schema instance / physical database server / data center / continent. There is no ongoing need to retain shared access (from between shards) to the other unpartitioned tables in other shards.
This makes replication across multiple servers easy (simple horizontal partitioning does not). It is also useful for worldwide distribution of applications, where communications links between data centers would otherwise be a bottleneck.
There is also a requirement for some notification and replication mechanism between schema instances, so that the unpartitioned tables remain as closely synchronized as the application demands. This is a complex choice in the architecture of sharded systems: approaches range from making these effectively read-only (updates are rare and batched), to dynamically replicated tables (at the cost of reducing some of the distribution benefits of sharding) and many options in between.
eXtreme Scale is a cross-process in-memory key/value datastore (a variety of NoSQL datastore). It uses sharding to achieve scalability across processes for both data and MapReduce-style parallel processing.
Hibernate Shards provides support for shards, although there has been little activity since 2007.
IBM supports sharding in Informix since version 12.1 xC1 as part of the MACH11 technology. Informix 12.10 xC2 added full compatibility with MongoDB drivers, allowing the mix of regular relational tables with NoSQL collections, still supporting sharding, failover and ACID properties.
Solr enterprise search server provides sharding capabilities.
Spanner is Google's global scale distributed database that shards data across multiple Paxos state machines to scale to "millions of machines across hundreds of datacenters and trillions of database rows".
SQLAlchemy is an object-relational mapper for the Python programming language that provides sharding capabilities.
Microsoft supported sharding in SQL Azure through "Federations". In 2014, support for the under-used Federations was dropped. Although Azure users were using sharding, they were choosing to implement it through application-specific custom approaches, rather than the automatic approach of Federations.
Sharding a database table before it has been optimized locally causes premature complexity. Sharding should be used only when all other options for optimization are inadequate. The introduced complexity of database sharding causes the following potential problems:
Increased complexity of SQL - Increased bugs because the developers have to write more complicated SQL to handle sharding logic.
Sharding introduces complexity - The sharding software that partitions, balances, coordinates, and ensures integrity can fail.
Single point of failure - Corruption of one shard due to network/hardware/systems problems causes failure of the entire table.
Failover servers more complex - Failover servers must themselves have copies of the fleets of database shards.
Backups more complex - Database backups of the individual shards must be coordinated with the backups of the other shards.
Operational complexity added - Adding/removing indexes, adding/deleting columns, modifying the schema becomes much more difficult.
These historical complications of do-it-yourself sharding are now being addressed by independent software vendors who provide autosharding solutions.