Fundamental concepts for scaling in Azure Cosmos DB for PostgreSQL

APPLIES TO: Azure Cosmos DB for PostgreSQL (powered by the Citus database extension to PostgreSQL)

Before we investigate the steps of building a new app, it's helpful to see a quick overview of the terms and concepts involved.

Architectural overview

Azure Cosmos DB for PostgreSQL gives you the power to distribute tables and/or schemas across multiple machines in a cluster and transparently query them the same you query plain PostgreSQL:

In the Azure Cosmos DB for PostgreSQL architecture, there are multiple kinds of nodes:

• The coordinator node stores distributed table metadata and is responsible for distributed planning.
• By contrast, the worker nodes store the actual data, metadata and do the computation.
• Both the coordinator and workers are plain PostgreSQL databases, with the citus extension loaded.

To distribute a normal PostgreSQL table, like campaigns in the diagram above, run a command called create_distributed_table(). Once you run this command, Azure Cosmos DB for PostgreSQL transparently creates shards for the table across worker nodes. In the diagram, shards are represented as blue boxes.

To distribute a normal PostgreSQL schema, you run the citus_schema_distribute() command. Once you run this command, Azure Cosmos DB for PostgreSQL transparently turns tables in such schemas into a single shard colocated tables that can be moved as a unit between nodes of the cluster.

Note

On a cluster with no worker nodes, shards of distributed tables are on the coordinator node.

Shards are plain (but specially named) PostgreSQL tables that hold slices of your data. In our example, because we distributed campaigns by company_id, the shards hold campaigns, where the campaigns of different companies are assigned to different shards.

Distribution column (also known as shard key)

create_distributed_table() is the magic function that Azure Cosmos DB for PostgreSQL provides to distribute tables and use resources across multiple machines.

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

The second argument above picks a column from the table as a distribution column. It can be any column with a native PostgreSQL type (with integer and text being most common). The value of the distribution column determines which rows go into which shards, which is why the distribution column is also called the shard key.

Azure Cosmos DB for PostgreSQL decides how to run queries based on their use of the shard key:

Query involves	Where it runs
just one shard key	on the worker node that holds its shard
multiple shard keys	parallelized across multiple nodes

The choice of shard key dictates the performance and scalability of your applications.

• Uneven data distribution per shard keys (also known as data skew) isn't optimal for performance. For example, don’t choose a column for which a single value represents 50% of data.
• Shard keys with low cardinality can affect scalability. You can use only as many shards as there are distinct key values. Choose a key with cardinality in the hundreds to thousands.
• Joining two large tables with different shard keys can be slow. Choose a common shard key across large tables. Learn more in colocation.

Colocation

Another concept closely related to shard key is colocation. Tables sharded by the same distribution column values are colocated - The shards of colocated tables are stored together on the same workers.

Below are two tables sharded by the same key, site_id. They're colocated.

Azure Cosmos DB for PostgreSQL ensures that rows with a matching site_id value in both tables are stored on the same worker node. You can see that, for both tables, rows with site_id=1 are stored on worker 1. Similarly for other site IDs.

Colocation helps optimize JOINs across these tables. If you join the two tables on site_id, Azure Cosmos DB for PostgreSQL can perform the join locally on worker nodes without shuffling data between nodes.

Tables within a distributed schema are always colocated with each other.

Next steps

Classify application workload >