Nodes and tables in Azure Database for PostgreSQL – Hyperscale (Citus)

Nodes

The Hyperscale (Citus) hosting type allows Azure Database for PostgreSQL servers (called nodes) to coordinate with one another in a "shared nothing" architecture. The nodes in a server group collectively hold more data and use more CPU cores than would be possible on a single server. The architecture also allows the database to scale by adding more nodes to the server group.

Coordinator and workers

Every server group has a coordinator node and multiple workers. Applications send their queries to the coordinator node, which relays it to the relevant workers and accumulates their results. Applications are not able to connect directly to workers.

Hyperscale (Citus) allows the database administrator to distribute tables, storing different rows on different worker nodes. Distributed tables are the key to Hyperscale (Citus) performance. Failing to distribute tables leaves them entirely on the coordinator node and cannot take advantage of cross-machine parallelism.

For each query on distributed tables, the coordinator either routes it to a single worker node, or parallelizes it across several depending on whether the required data lives on a single node or multiple. The coordinator decides what to do by consulting metadata tables. These tables track the DNS names and health of worker nodes, and the distribution of data across nodes.

Table types

There are three types of tables in a Hyperscale (Citus) server group, each stored differently on nodes and used for different purposes.

Type 1: Distributed tables

The first type, and most common, is distributed tables. They appear to be normal tables to SQL statements, but they're horizontally partitioned across worker nodes. What this means is that the rows of the table are stored on different nodes, in fragment tables called shards.

Hyperscale (Citus) runs not only SQL but DDL statements throughout a cluster. Changing the schema of a distributed table cascades to update all the table's shards across workers.

Distribution column

Hyperscale (Citus) uses algorithmic sharding to assign rows to shards. The assignment is made deterministically based on the value of a table column called the distribution column. The cluster administrator must designate this column when distributing a table. Making the right choice is important for performance and functionality.

Type 2: Reference tables

A reference table is a type of distributed table whose entire contents are concentrated into a single shard. The shard is replicated on every worker. Queries on any worker can access the reference information locally, without the network overhead of requesting rows from another node. Reference tables have no distribution column because there's no need to distinguish separate shards per row.

Reference tables are typically small and are used to store data that's relevant to queries running on any worker node. An example is enumerated values like order statuses or product categories.

Type 3: Local tables

When you use Hyperscale (Citus), the coordinator node you connect to is a regular PostgreSQL database. You can create ordinary tables on the coordinator and choose not to shard them.

A good candidate for local tables would be small administrative tables that don't participate in join queries. An example is a users table for application sign-in and authentication.

Shards

The previous section described how distributed tables are stored as shards on worker nodes. This section discusses more technical details.

The pg_dist_shard metadata table on the coordinator contains a row for each shard of each distributed table in the system. The row matches a shard ID with a range of integers in a hash space (shardminvalue, shardmaxvalue).

SELECT * from pg_dist_shard;
 logicalrelid  | shardid | shardstorage | shardminvalue | shardmaxvalue
---------------+---------+--------------+---------------+---------------
 github_events |  102026 | t            | 268435456     | 402653183
 github_events |  102027 | t            | 402653184     | 536870911
 github_events |  102028 | t            | 536870912     | 671088639
 github_events |  102029 | t            | 671088640     | 805306367
 (4 rows)

If the coordinator node wants to determine which shard holds a row of github_events, it hashes the value of the distribution column in the row. Then the node checks which shard's range contains the hashed value. The ranges are defined so that the image of the hash function is their disjoint union.

Shard placements

Suppose that shard 102027 is associated with the row in question. The row is read or written in a table called github_events_102027 in one of the workers. Which worker? That's determined entirely by the metadata tables. The mapping of shard to worker is known as the shard placement.

The coordinator node rewrites queries into fragments that refer to the specific tables like github_events_102027 and runs those fragments on the appropriate workers. Here's an example of a query run behind the scenes to find the node holding shard ID 102027.

SELECT
    shardid,
    node.nodename,
    node.nodeport
FROM pg_dist_placement placement
JOIN pg_dist_node node
  ON placement.groupid = node.groupid
 AND node.noderole = 'primary'::noderole
WHERE shardid = 102027;
┌─────────┬───────────┬──────────┐
│ shardid │ nodename  │ nodeport │
├─────────┼───────────┼──────────┤
│  102027 │ localhost │     5433 │
└─────────┴───────────┴──────────┘

Next steps