Tristan Robinson

Tristan Robinson's Blog

Getting Started with Databricks Cluster Pricing

The use of databricks for data engineering or data analytics workloads is becoming more prevalent as the platform grows, and has made its way into most of our recent modern data architecture proposals – whether that be PaaS warehouses, or data science platforms.

To run any type of workload on the platform, you will need to setup a cluster to do the processing for you. While the Azure-based platform has made this relatively simple for development purposes, i.e. give it a name, select a runtime, select the type of VMs you want and away you go – for production workloads, a bit more thought needs to go into the configuration/cost.  In the following blog I’ll start by looking at the pricing in a bit more detail which will aim to provide a cost element to the cluster configuration process.

For arguments sake, the work that we tend to deliver with databricks is based on data engineering usage – spinning up resource for an allocated period to perform a task. Therefore this is generally the focus for the following topic.

 

To get started in this area, I think it would be useful to included some definitions.

  • DBU – a databricks unit (unit of processing capability per hour billed on per second usage)
  • Data Engineering Workload - a job that both starts and terminates the cluster which it runs on (via the job scheduler)
  • Data Analytics Workload – a non automated workload, for example running a command manually within a databricks notebook. Multiple users can share the cluster to perform interactive analysis
  • Cluster – made up of instances of processing (VMs) and constitute of a driver, and workers. Workers can either be provisioned upfront, or autoscaled between a min no. workers / max no. workers.
  • Tier – either standard or premium. Premium includes role based access control, ODBC endpoint authentication, audit logs, Databricks Delta (unified data management system).

The billing for the clusters primarily works depending on the type of workload you initiate and tier (or functionality) you require. As you might of guessed data engineering workloads on the standard tier offer the best price.

I’ve taken the DS3 v2 instance (VM) pricing from the Azure Databricks pricing page.

image

 

The pricing can be broken down as follows:

  • Each instance is charged at £0.262/hour. So for example, the cost of a very simple cluster - 1 driver and 2 workers is £0.262/hour x 3 = £0.786/hour. The VM cost does not depend on the workload type/tier.
  • The DBU cost is then calculated at £0.196/hour. So for example, the cost of 3 nodes (as above) is £0.196/hour x 3 = £0.588/hour. This cost does change depending on workload type/tier.
  • The total cost is then £0.786/hour (VM Cost) + £0.588/hour (DBU Cost) = £1.374/hour. Also known as the pay as you go price. Discounts are then added accordingly for reserved processing power.

I thought this was worth simplifying since the pricing page doesn’t make this abundantly clear with the way the table is laid out and often this is overlooked. Due to the vast amount of options you can have to setup clusters, its worth understanding this element to balance against time.

The DBU count is merely to be used as reference to compare the different VMs processing power and is not directly included in the calculations. Its also worth mentioning that by default databricks services are setup as premium and can be downgraded to standard only by contacting support. In some cases, this can add some massive cost savings depending upon the type of work you are doing on the platform so please take this into account before spinning up clusters and don’t just go with the default.

With regards to configuration, clusters can either be setup under a High Concurrency mode (previously known as serverless) or as Standard. The high concurrency mode is optimised for concurrent workloads and therefore is more applicable to data analytics workloads and interactive notebooks which are used by multiple users simultaneously. This piece of configuration does not effect the pricing.

By using the following cost model, we can then assume for a basic batch ETL run where we have a driver and 8 worker nodes on relatively small DS3 instances, would cost £123.60/month given a standard 1 hour daily ETL window. Hopefully this provides as a very simple introduction into the pricing model used by Databricks.