Druid is an OLAP engine designed to provide fast real time analytics. Druid adopts a clustered architecture with servers that host various role specific processes. These processes address real time and batch ingestion, indexing, querying of historical and real time data. Apache Druid can be deployed as a virtual machine or a Kubernetes based cluster. Druid does not support a decoupled compute & storage architecture. Deep storage in the form of object storage is used to replicate data to.
Athena is serverless and built on a decoupled storage and compute architecture that queries data directly in S3, without the need to ingest/copy the data. It runs in multi-tenancy with shared resources. Users do not have control over the compute resources Athena chooses to allocate per query from the shared resource pool. For folks requiring additional or dedicated resources, they can reserve dedicated processing capacity in the form of Data Processing Units (DPU), with each DPU providing 4 vCPU and 16 GB RAM. RPU allocation ranges from 24 - 1000 per region.
Druid provides the ability to handle fast ingest and high concurrency. Custom sizing and cluster tuning are required to balance the compute, memory, storage needs of each process within Druid and to provide high concurrency. Druid clusters can be grown by adding nodes with automatic rebalancing of storage segments assigned to nodes. Self hosted Druid on Kubernetes is an option that users leverage to simplify scaling. Additionally, Cloud based managed Druid offerings are being rolled out. However, these managed offerings are limited in scale and scaling is not granular.
Athena is a shared multi-tenant resource, with no guarantees on the amount or availability of the resources allocated for your queries. From a data volume perspective, it can scale to large volumes, but large data volumes can suffer from very long run times and frequent timeouts. Query concurrency is maxed at 20. If scalability is a top priority, Athena is probably not the best choice.
Druid provides high performance through columnar storage format, parallel processing, bitmap indexes and roll-ups. Druid, however, recommends a denormalized data model for performance needs. Join operations in Druid are a relatively new feature with various limitations, especially if there is a need to join large datasets.
Athena (and Presto) are designed to query data where it is, sacrificing storage-compute optimizations. This makes it very convenient for easy and immediate querying but at the expense of performance. This typically puts Athena behind cloud data warehouses in terms of performance. But Athena still does relatively well in performance benchmarks, especially when external storage is managed by experts. While it supports partitions, there is no support for indexing, and together with the fact that resources are pooled from a shared multi-tenant service, low-latency and consistent performance are not Athena’s sweet spot. A cloud data warehouse is more performant than Athena in most cases.
Druid is designed as an OLAP engine to provide fast access to aggregations that are run against large volumes of data. Druid is typically used for customer facing analytics and streaming data processing. Druid is used as an add-on with other data warehousing products that are efficient at scaling, joining, and filtering large volumes of data. It is not a suitable option for data warehouse replacement.
Athena is a great choice for Ad-Hoc analytics. You can keep the data where it is, and start querying without worrying about hardware or pretty much anything else, given that Athena is serverless and takes care of everything behind the scenes. However, it is not a great fit when you need consistent and fast query performance, and/or high concurrency. This is why it is typically not the best choice for operational and customer-facing applications. It can be also easily and flexibly used for batch processing, which is often leveraged for ML use cases.