vitess-gh/doc/Sharding.md

Sharding is a method of horizontally partitioning a database to store data across two or more database servers. This document explains how sharding works in Vitess and the types of sharding that Vitess supports.

Overview

In Vitess, a shard is a partition of a keyspace. In turn, the keyspace might be a partition of the whole database. For example, a database might have one keyspace for product data and another for user data. The shard contains a subset of records within its keyspace.

For example, if an application's "user" keyspace is split into two shards, each shard contains records for approximately half of the application's users. Similarly, each user's information is stored in only one shard.

Note that sharding is orthogonal to (MySQL) replication. A Vitess shard typically contains one MySQL master and many MySQL slaves. The master handles write operations, while slaves handle read-only traffic, batch processing operations, and other tasks. Each MySQL instance within the shard should have the same data, excepting some replication lag.

Supported Operations

Vitess supports the following types of sharding operations:

• Horizontal sharding: Splitting or merging shards in a sharded keyspace
• Vertical sharding: Moving tables from an unsharded keyspace to a different keyspace.

With these features, you can start with a single keyspace that contains all of your data (in multiple tables). As your database grows, you can move tables to different keyspaces (vertical split) and shard some or all of those keyspaces (horizontal split) without any real downtime for your application.

Range-based Sharding

Vitess uses range-based sharding to manage data across multiple shards. (Vitess can also support a custom sharding scheme.)

In range-based sharding, each record in a keyspace is associated with a sharding key that is stored with the record. The sharding key value is also the primary key for sharded data. Records with the same sharding key are always collocated on the same shard.

Note: The API uses the term "keyspace ID" to refer to the sharding key.

The full set of shards covers the range of possible sharding key values. To guarantee a balanced use of shards, the sharding scheme should ensure an even distribution of sharding keys across the keyspace's shards. That distribution makes it easier to reshard the keyspace at a later time using a more granular division of sharding keys.

Vitess calculates the sharding key or keys for each query and then routes that query to the appropriate shards. For example, a query that updates information about a particular user might be directed to a single shard in the application's "user" keyspace. On the other hand, a query that retrieves information about several products might be directed to one or more shards in the application's "product" keyspace.

Key Ranges and Partitions

Vitess uses key ranges to determine which shards should handle any particular query.

• A key range is a series of consecutive sharding key values. It has starting and ending values. A key falls inside the range if it is equal to or greater than the start value and strictly less than the end value.
• A partition represents a set of key ranges that covers the entire space.

When building the serving graph for a keyspace that uses range-based sharding, Vitess ensures that each shard is valid and that the shards collectively constitute a full partition. In each keyspace, one shard must have a key range with an empty start value and one shard, which could be the same shard, must have a key range with an empty end value.

• An empty start value represents the lowest value, and all values are greater than it.
• An empty end value represents a value larger than the highest possible value, and all values are strictly lower than it.

Vitess always converts sharding keys to byte arrays before routing queries. The value [ 0x80 ] is the middle value for sharding keys. So, in a keyspace with two shards, sharding keys that have a byte-array value lower than 0x80 are assigned to one shard. Keys with a byte-array value equal to or higher than 0x80 are assigned to the other shard.

Several sample key ranges are shown below:

Start=[], End=[]: Full Key Range
Start=[], End=[0x80]: Lower half of the Key Range.
Start=[0x80], End=[]: Upper half of the Key Range.
Start=[0x40], End=[0x80]: Second quarter of the Key Range.

Two key ranges are consecutive if the end value of one range equals the start value of the other range.

Shard Names in Range-Based Keyspaces

In range-based, sharded keyspaces, a shard's name identifies the start and end of the shard's key range, printed in hexadecimal and separated by a hyphen. For instance, if a shard's key range is the array of bytes beginning with [ 0x80 ] and ending, noninclusively, with [ 0xc0], then the shard's name is 80-c0.

Using this naming convention, the following four shards would be a valid full partition:

• -40
• 40-80
• 80-c0
• c0-

Resharding

In Vitess, resharding describes the process of updating the sharding scheme for a keyspace and dynamically reorganizing data to match the new scheme. During resharding, Vitess copies, verifies, and keeps data up-to-date on new shards while the existing shards continue to serve live read and write traffic. When you're ready to switch over, the migration occurs with only a few seconds of read-only downtime. During that time, existing data can be read, but new data cannot be written.

The table below lists the sharding (or resharding) processes that you would typically perform for different types of requirements:

Requirement	Action
Uniformly increase read capacity	Add replicas or split shards
Uniformly increase write capacity	Split shards
Reclaim overprovisioned resources	Merge shards and/or keyspaces
Increase geo-diversity	Add new cells and replicas
Cool a hot tablet	For read access, add replicas or split shards. For write access, split shards.

Filtered Replication

The cornerstone of resharding is replicating the right data. Vitess implements the following functions to support filtered replication, the process that ensures that the correct source tablet data is transferred to the proper destination tablets. Since MySQL does not support any filtering, this functionality is all specific to Vitess.

1. The source tablet tags transactions with comments so that MySQL binlogs contain the filtering data needed during the resharding process. The comments describe the scope of each transaction (its keyspace ID, table, etc.).
2. A server process uses the comments to filter the MySQL binlogs and stream the correct data to the destination tablet.
3. A client process on the destination tablet applies the filtered logs, which are just regular SQL statements at this point.

Additional Tools and Processes

Vitess provides the following tools to help manage range-based shards:

• The vtctl command-line tool supports functions for managing keyspaces, shards, tablets, and more.
• Client APIs account for sharding operations.
• The MapReduce framework fully utilizes key ranges to read data as quickly as possible, concurrently from all shards and all replicas.

Custom Sharding

If your application already supports sharding or if you want to control exactly which shard handles each query, Vitess can support your custom sharding scheme. In that use case, each keyspace has a collection of shards, and the client code always specifies the shard to which it is directing a query.

One example of a custom sharding scheme is lookup-based sharding. In lookup-based sharding, one keyspace is used as a lookup keyspace, and it contains the mapping between a record's identifying key and the name of the record's shard. To execute a query, the client first checks the lookup table to locate the correct shard name and then routes the query to that shard.

In a custom sharding scheme, shards can use any name you choose, and they are always addressed by name. The vtgate API calls to use are ExecuteShard, ExecuteBatchShard, and StreamExecuteShard. None of the API calls for KeyspaceIds, KeyRanges, or EntityIds are compatible with a custom sharding scheme. Vitess' tools and processes for automated resharding also do not support custom sharding schemes.

If you use a custom sharding scheme, you can still use the MapReduce framework to iterate over the data on multiple shards.