Permalink
Fetching contributors…
Cannot retrieve contributors at this time
215 lines (166 sloc) 10.3 KB

Vitess Query Features

Overview

Client applications connect to Vitess through VTGate. VTGate is a proxy that routes queries to the right vttablets. It is basically a fat client, but provided as a service. Over time, we moved away from a fat client, which had to be incorporated into the application. This way, the routing logic is more decoupled from the application and easier to maintain separately.

The previous versions of the VTGate API existed as transition points while the logic was gradually moved away from the client into this new server:

  • V0: This version had no VTGate. The application was expected to fetch the sharding info and tablet locations from a 'toposerver' and use it to talk directly to the tablet servers.
  • V1: This was the first version of VTGate. In this version, the app only needed to know the number of shards, and how to map the sharding key to the correct shard. The rest was done by VTGate. In this version, the app was still exposed to resharding events.
  • V2: In this version, the keyspace id was required instead of the shard. This allowed the app to be agnostic of the number of shards.

With V3, the app does not need to specify any routing info. It just sends the query to VTGate as if it's a single database. Apart from simplifying the API, there are some additional benefits:

  • Database compliant drivers can be built for each client language. This will also allow for integration with third party tools that work with such drivers.
  • V3 can aspire to satisfy the full SQL syntax. This means that it will be able to perform cross-shard joins, aggregations and sorting.
  • Easier migration: an application that was written to use a single database can be trivially changed to use Vitess, and then the database can be scaled from underneath without changing much of the app.

The V3 design is quite elaborate. If necessary, it will allow you to plug in custom indexes and sharding schemes. However, it comes equipped with some pre-cooked recipes that satisfy the immediate needs of the real-world:

Sharding schemes

At its core, vitess uses range-based sharding, where the sharding column is typically a number or a varbinary. However, allowing data to be accessed only by the sharding key limits the flexibility of an application. V3 comes with a set of new indexing schemes that are built on top of range-based sharding.

Basic sharding key

If the application already has a well-distributed sharding key, you just have to tell VTGate what those keys are for each table. VTGate will correctly route your queries based on input values or the WHERE clause.

Hashed sharding key

If the application's sharding key is a monotonically increasing number, then you may not get well-balanced shards. In such cases, you can ask V3 to route queries based on the hash of the main sharding key.

With V3, it is no longer necessary to store the sharding key as an extra column in the table. Even during resharding, filtered replication can now compute the hashed values.

Auto-increment columns

When a table gets sharded, you are no longer able to use MySQL's auto increment functionality. V3 allows you to designate a table in an unsharded database as the source of auto-increment ids. Once you've specified this, V3 will transparently use generated values from this table to keep the auto-increment going. The auto-increment column can in turn be a basic or hashed sharding key. If it's a hashed sharding key, the newly generated value will be hashed before the query is routed.

Cross-shard indexes

As your application evolves, you'll invariably find yourself wanting to fetch rows based on columns other than the main sharding key. For example, if you've sharded your database by user id, you may still want to be able find users by their username. If you only had the sharding key, such queries can only be answered by sending it to all shards. This could become very expensive as the number of shards grow.

The typical strategy to address this problem is to build a separate lookup table and keep it up-to-date. In the above case, you may build a separate username->user_id relationship table. Once you've informed V3 of this table, it will know what to do with a query like 'select * from user where username=:value'. You can also configure V3 to keep this table up-to-date as you insert or delete data. In other words, the application can be completely agnostic of this table's existence.

Non-unique indexes

Cross-shard indexes don't have to be unique. It is possible that rows may exist in multiple shards for a given where clause. V3 allows you to specify indexes as unique or non-unique, and accordingly enforces such constraints during changes. This flexibility is currently available for lookup indexes. For example, you may want to index users by their last name. If so, there would be a last_name->user_id index, which could result in multiple user ids being returned for a given last_name.

Shared indexes

There are situations where multiple tables share the same foreign key. A typical use case is a situation where there is a customer table, an order table and an order_detail table. The order table would have a customer_id column. In order to efficiently access all orders of a customer, it would be beneficial to shard this table by customer_id. This will co-locate order rows with their corresponding customer row.

The order table would also need an order_id column. As mentioned above, you can create an order_id->customer_id cross-shard index for this table. This will allow you to efficiently access orders by their order_id.

In the case of an order_detail table, it may only need an order_id foreign key. Since this foreign key means the same thing as the order_id in order, creating a cross-shard index for it will result in a duplication of the order_id->customer_id index. In such situations, V3 allows you to just reuse the existing index for the order_detail table also. This saves disk space and also reduces the overall write load.

Knowing where tables are

As your database grows, you will not only be sharding it, you will also be splitting it vertically by migrating tables from one database to another. V3 will be able to keep track of this. The app will only have to refer a table by name, and VTGate will figure out how to route the query to the correct database.

The vitess workflow also ensures that such migrations are done transparently with virtually no downtime.

Consistency

Once you add multiple indexes to tables, it's possible that the application could make inconsistent requests. V3 makes sure that none of the specified constraints are broken. For example, if a table had both a basic sharding key and a hashed sharding key, it will enforce the rule that the hash of the basic sharding key matches that of the hashed sharding key.

Some of the changes require updates to be performed across multiple databases. For example, inserting a row into a table that has a cross-shard key requires an additional row to be inserted into the lookup table. This results in distributed transactions. Currently, this is a best effort update. It is possible that partial commits happen if databases fail in the middle of a distributed commit.

The 2PC transactions feature is coming very soon to overcome this limitation.

Query Diversity

V3 does not support the full SQL feature set. The current implementation supports the following queries:

  • Single table DML statements: This is a vitess-wide restriction where you can affect only one table and one sharding key per statement. This restriction may be removed in the future.
  • Single table SELECT statements:
    • All constructs allowed if the statement targets only a single sharding key
    • Aggregation and sorting not allowed if the statement targets more than one sharding key. Selects are allowed to target multiple sharding keys as long as the results from individual shards can be simply combined together to form the final result.
  • Joins that can be served by sending the query to a single shard.
  • Joins that can be served by sending the query to multiple shards, and trivially combined to form the final result.

Work is underway to support the following additional constructs:

  • Cross-shard joins that can be served through a simple nested lookup.
  • Sorting that can be trivially merged from the results of multiple shards (merge-sort).
  • Aggregations (and grouping) that can be trivially combined from multiple shards.
  • A combination of the above constructs as long as the results remain trivially combinable.

SQL is a very powerful language. You can build queries that can result in large amount of work and memory consumption involving big intermediate results. Such constructs where the scope of work is open-ended will not be immediately supported. In such cases, it's recommended that you use map-reduce techniques for which there is a separate API.

On-the-fly map-reducers can be built to address the more complex needs in the future.

Special Queries

MySQL supports a number of tables that are not storing data, but rather expose meta-data about the database. For instance, the information_schema database.

The V3 API is meant to resemble a single database instance (using the Execute API calls). But it can be backed by multiple keyspaces and each can have multiple shards. In the long term, we want to behave as if all keyspaces and tables are in one single instance: each keyspace would be shown as a database, each table in a sharded keyspace would only be shown once (and not once per shard).

We do not support these queries just yet.

For administrators who want to target individual shards, it is possible to use the V1 API (with the ExecuteShards API). A deeper knowledge of the existing shards is then required, but administrators have that knowledge.

VSchema

The above features require metadata like configuration of sharding key and cross-shard indexes to be configured and stored in some place. This is known as the vschema.

A way to edit the VSchema using SQL-like commands will be built to integrate VSchema changes with regular SQL schema changes. For now, the VSchema needs to be specified as a JSON file.

Under the covers, the vschema is a JSON file. There are low level vtctl commands to upload it also. This will allow you to build workflows for tracking and managing changes.