Running in production

Cube makes use of two different kinds of cache:

• In-memory storage of query results
• Pre-aggregations

Cube Store is enabled by default when running Cube in development mode. In production, Cube Store must run as a separate process. The easiest way to do this is to use the official Docker images for Cube and Cube Store.

You can run Cube Store with Docker with the following command:

docker run -p 3030:3030 cubejs/cubestore

Next, run Cube and tell it to connect to Cube Store running on localhost (on the default port 3030):

docker run -p 4000:4000 \
  -e CUBEJS_CUBESTORE_HOST=localhost \
  -v ${PWD}:/cube/conf \
  cubejs/cube

In the command above, we're specifying CUBEJS_CUBESTORE_HOST to let Cube know where Cube Store is running.

You can also use Docker Compose to achieve the same:

version: "2.2"
services:
  cubestore:
    image: cubejs/cubestore:latest
    environment:
      - CUBESTORE_REMOTE_DIR=/cube/data
    volumes:
      - .cubestore:/cube/data
 
  cube:
    image: cubejs/cube:latest
    ports:
      - 4000:4000
    environment:
      - CUBEJS_CUBESTORE_HOST=localhost
    depends_on:
      - cubestore
    links:
      - cubestore
    volumes:
      - ./model:/cube/conf/model

Architecture

A Cube Store cluster consists of at least one Router and one or more Worker instances. Cube sends queries to the Cube Store Router, which then distributes the queries to the Cube Store Workers. The Workers then execute the queries and return the results to the Router, which in turn returns the results to Cube.

Scaling

Scaling Cube Store for a higher concurrency is relatively simple when running in cluster mode. Because the storage layer is decoupled from the query processing engine, you can horizontally scale your Cube Store cluster for as much concurrency as you require.

In cluster mode, Cube Store runs two kinds of nodes:

• one or more router nodes handle incoming client connections, manage database metadata and serve simple queries.
• multiple worker nodes which execute SQL queries

Cube Store querying performance is optimal when the count of partitions in a single query is less than or equal to the worker count. For example, you have a 200 million rows table that is partitioned by day, which is ten daily Cube partitions or 100 Cube Store partitions in total. The query sent by the user contains filters, and the resulting scan requires reading 16 Cube Store partitions in total. Optimal query performance, in this case, can be achieved with 16 or more workers. You can use EXPLAIN and EXPLAIN ANALYZE SQL commands to see how many partitions would be used in a specific Cube Store query.

Resources required for the main node and workers can vary depending on the configuration. With default settings, you should expect to allocate at least 4 CPUs and up to 8GB per main or worker node.

The configuration required for each node can be found in the table below. More information about these variables can be found in the Environment Variables reference.

CUBESTORE_WORKERS and CUBESTORE_META_ADDR variables should be set with stable addresses, which should not change. You can use stable DNS names and put load balancers in front of your worker and router instances to fulfill stable name requirements in environments where stable IP addresses can't be guaranteed.

A sample Docker Compose stack for the single machine setting this up might look like:

version: "2.2"
services:
  cubestore_router:
    restart: always
    image: cubejs/cubestore:latest
    environment:
      - CUBESTORE_SERVER_NAME=cubestore_router:9999
      - CUBESTORE_META_PORT=9999
      - CUBESTORE_WORKERS=cubestore_worker_1:9001,cubestore_worker_2:9001
      - CUBESTORE_REMOTE_DIR=/cube/data
    volumes:
      - .cubestore:/cube/data
  cubestore_worker_1:
    restart: always
    image: cubejs/cubestore:latest
    environment:
      - CUBESTORE_SERVER_NAME=cubestore_worker_1:9001
      - CUBESTORE_WORKER_PORT=9001
      - CUBESTORE_META_ADDR=cubestore_router:9999
      - CUBESTORE_WORKERS=cubestore_worker_1:9001,cubestore_worker_2:9001
      - CUBESTORE_REMOTE_DIR=/cube/data
    depends_on:
      - cubestore_router
    volumes:
      - .cubestore:/cube/data
  cubestore_worker_2:
    restart: always
    image: cubejs/cubestore:latest
    environment:
      - CUBESTORE_SERVER_NAME=cubestore_worker_2:9001
      - CUBESTORE_WORKER_PORT=9001
      - CUBESTORE_META_ADDR=cubestore_router:9999
      - CUBESTORE_WORKERS=cubestore_worker_1:9001,cubestore_worker_2:9001
      - CUBESTORE_REMOTE_DIR=/cube/data
    depends_on:
      - cubestore_router
    volumes:
      - .cubestore:/cube/data
  cube:
    image: cubejs/cube:latest
    ports:
      - 4000:4000
    environment:
      - CUBEJS_CUBESTORE_HOST=cubestore_router
    depends_on:
      - cubestore_router
    volumes:
      - .:/cube/conf

Replication and High Availability

The open-source version of Cube Store doesn't support replicating any of its nodes. The router node and every worker node should always have only one instance copy if served behind the load balancer or service address. Replication will lead to undefined behavior of the cluster, including connection errors and data loss. If any cluster node is down, it'll lead to a complete cluster outage. If Cube Store replication and high availability are required, please consider using Cube Cloud.

Storage

Cube Store cluster uses both persistent and scratch storage.

Persistent storage

Cube Store makes use of a separate storage layer for storing metadata as well as for persisting pre-aggregations as Parquet files.

Cube Store can be configured to use either AWS S3 or Google Cloud Storage (GCS) as persistent storage. If desired, local path on the server can also be used in case all Cube Store cluster nodes are co-located on a single machine.

A simplified example using AWS S3 might look like:

version: "2.2"
 
services:
  cubestore_router:
    image: cubejs/cubestore:latest
    environment:
      - CUBESTORE_SERVER_NAME=cubestore_router:9999
      - CUBESTORE_META_PORT=9999
      - CUBESTORE_WORKERS=cubestore_worker_1:9001
      - CUBESTORE_S3_BUCKET=<BUCKET_NAME_IN_S3>
      - CUBESTORE_S3_REGION=<BUCKET_REGION_IN_S3>
      - CUBESTORE_AWS_ACCESS_KEY_ID=<AWS_ACCESS_KEY_ID>
      - CUBESTORE_AWS_SECRET_ACCESS_KEY=<AWS_SECRET_ACCESS_KEY>
 
  cubestore_worker_1:
    image: cubejs/cubestore:latest
    environment:
      - CUBESTORE_SERVER_NAME=cubestore_worker_1:9001
      - CUBESTORE_WORKER_PORT=9001
      - CUBESTORE_META_ADDR=cubestore_router:9999
      - CUBESTORE_WORKERS=cubestore_worker_1:9001
      - CUBESTORE_S3_BUCKET=<BUCKET_NAME_IN_S3>
      - CUBESTORE_S3_REGION=<BUCKET_REGION_IN_S3>
      - CUBESTORE_AWS_ACCESS_KEY_ID=<AWS_ACCESS_KEY_ID>
      - CUBESTORE_AWS_SECRET_ACCESS_KEY=<AWS_SECRET_ACCESS_KEY>
    depends_on:
      - cubestore_router

Note that you can’t use the same bucket as an export bucket and persistent storage for Cube Store. It's recommended to use two separate buckets.

Scratch storage

Separately from persistent storage, Cube Store requires local scratch space to warm up partitions by downloading Parquet files before querying them.

By default, this folder should be mounted to .cubestore/data inside the container and can be configured by CUBESTORE_DATA_DIR environment variable. It is advised to use local SSDs for this scratch space to maximize querying performance.

AWS

Cube Store can retrieve security credentials from instance metadata automatically. This means you can skip defining the CUBESTORE_AWS_ACCESS_KEY_ID and CUBESTORE_AWS_SECRET_ACCESS_KEY environment variables.

Garbage collection

Cleanup isn’t done in export buckets; however, it's done in the persistent storage of Cube Store. The default time-to-live (TTL) for orphaned pre-aggregation tables is one day.

Refresh worker should be able to finish pre-aggregation refresh before garbage collection starts. It means that all pre-aggregation partitions should be built before any tables are removed.

Security

Cube Store currently does not have any in-built authentication mechanisms. For this reason, we recommend running your Cube Store cluster on a network that only allows requests from the Cube deployment.