Getting Started with Citus

Citus achieves order-of-magnitude faster execution compared to vanilla PostgreSQL through a combination of parallelism, keeping more data in memory, and higher I/O bandwidth.

Citus enables real-time interaction with large datasets that span billions of records—and is a good fit for customer-facing workloads that often require low-latency response times. Performance increases as you add nodes to a Citus database cluster. This 15-min performance demo from SIGMOD shows how Citus speeds up Postgres, using the HammerDB benchmark. Recently GigaOm published a benchmark performance report for Citus. Find out why benchmarking databases is so hard in this blog post by the lead engineer for Citus. Columnar storage can speed up analytics workloads that benefit from compression, too.

The easiest way to start is by utilizing schema-based sharding, which assumes assigning each tenant to a separate schema. Citus then automatically distributes these among the nodes in your cluster and routes queries accordingly. The only change you will need to do in your application is to SET search_path when switching tenants. In some cases like with microservices, even that change may not be necessary if every microservice uses a separate user matching their schema name.

If you want the best performance, row-based sharding, using a distribution column is the best approach. The first step in migrating an application from Postgres to Citus is to choose your distribution column (sometimes called a distribution key, or a sharding key.) You’ll want to understand your workload in order to pinpoint a “good” distribution column, e.g., a column that enables you to get the maximum performance from Citus.

The second step is to prepare the Postgres tables and SQL queries for migration. The amount of effort involved depends (you’ve heard that before, right?) on whether your application is already centered around that distribution column in terms of queries and schema. If not, you may have to update some of your queries and/or add the distribution column to some of your tables.

Alternately, you can now shard your database on a single node. So you can build your application on single-node Citus from the very start and be “scale-out-ready”, able to easily add nodes and rebalance your Citus cluster as your application grows.

If you are ready to delve deeper, the Migrating to Citus guide in the Citus documentation should be useful.

The Citus extension to Postgres is commonly used with customer-facing applications that are growing fast, have demanding performance requirements, are starting to experience slow queries, need to plan for future scale—or all of the above. Common use cases for Citus—both self-hosted and in the cloud as the Azure Cosmos DB for PostgreSQL managed service—include:

• Customer-facing analytics dashboards
• SaaS applications—usually multi-tenant
• Time series workloads
• Microservices
• IOT workloads—that need UPDATEs & JOINs
• Mixed transactional & analytical workloads

As you’ll learn in the Citus concepts section of the documentation there are two ways of sharding Citus—row-based and schema-based. In both sharding methods Citus divides Postgres tables into multiple smaller tables, called shards. The shards are then spread across the nodes in the Citus database cluster.

In the case of row-based sharding you decide how tables are split using the create_distributed_table() function.

SELECT create_distributed_table(
  'table_name',
  'distribution_column');

With schema-based sharding, the schema name acts as the grouping and you determine which schemas are distributed with citus_schema_distribute("name") function.

SELECT citus_schema_distribute('name');

When new data is ingested or when queries come in, the Citus coordinator routes them to the correct shards based on the value of the distribution column (row-based) or the schema name (schema-based) depending on which sharding method you chose.