BigQuery was one of the first decoupled storage and compute architectures. It is a unique piece of engineering and not a typical data warehouse in part because it started as an on-demand serverless query engine. It runs in multi-tenancy with shared resources, allocated as “slots” which represent a virtual CPU that executes SQL. BigQuery determines how many slots a query requires, without the ability of the user to control it. BigQuery can be priced on a $/TB scanned basis or through slot reservations. A slot in BigQuery is logically equivalent to 0.5 vCPU and 0.5GB of RAM. There are multiple models to allocate slots in BigQuery.
Redshift has the oldest architecture, being the first Cloud DW in the group. Its architecture wasn’t designed to separate storage & compute. While it now has RA3 nodes which allow you to scale compute and only cache the data you need locally, all compute still operates together. You cannot separate and isolate different workloads over the same data, which puts it behind other decoupled storage/compute architectures. Redshift runs as an isolated tenant per customer, and unlike other cloud data warehouses, it is deployed in your VPC. Redshift offers a serverless option which is based on an abstracted unit called Redshift Processing Unit (RPU) ranging from 8 to 512 in increments of 8. Each RPU provides 2 vCPU and 16GB RAM. Thus, 8 RPU is equivalent to 16 vCPU / 128GB RAM. The minimum RPU is 8.
BigQuery scales very well to large data volumes, and automatically assigns more compute resources when needed behind the scenes, in the form of “slots”. BigQuery works either in an “on-demand pricing model”, where slot assignment is completely in the hands of BigQuery and the state of the shared resource pool, or in “flat-rate pricing model” where slots are reserved in advanced. With reserved slots there is more control over compute resources, thus making scaling more predictable. Concurrency is limited to 100 users by default.
Redshift is limited in scale because even with RA3, it cannot distribute different workloads across clusters. While it can scale to up to 10 clusters automatically to support query concurrency, it can only handle a maximum of 50 queued queries across all clusters by default.
BigQuery lines up in benchmarks in the same ballpark as other cloud data warehouses but does come in consistently last in most queries. Beyond implementing according to best practices, there is little you can do to accelerate BigQuery performance, as it determines the amount of resources (slots) the query needs for you. BigQuery can be used together with “BigQuery BI Engine” for lower latency analytics. However, BI Engine is limited in terms of scale because it runs in memory. Its maximum capacity is 100GB.
Redshift does provide a result cache for accelerating repetitive query workloads and also has more tuning options than some others. But it does not deliver much faster compute performance than other cloud data warehouses in benchmarks. Sort keys can be used to optimize performance, but their contribution is limited. There is no support for indexes, and low-latency analytics at large data volumes is hard to achieve. Because Redshift decoupling of storage & compute is limited compared to other cloud data warehouses, it doesn’t support isolating workloads, which means performance can degrade under pressure and competition for resources.
BigQuery is a mature general-purpose data warehouse, which lends itself well to internal BI & reporting. The fact that it’s serverless in nature and tightly integrated with GCP, makes it very convenient for Ad-Hoc analytics and ML use cases on GCP. On the other hand, because BigQuery makes resource allocation decisions for you, it is not always the best fit for operational use cases and Data Apps where performance needs to be consistent and predictable.
Redshift was originally designed to support traditional internal BI reporting and dashboard use cases for analysts. As such, it is typically used as a general-purpose Enterprise data warehouse. With deep integrations into the AWS ecosystem, it can also leverage AWS ML service, making it also useful for ML projects. However, given the coupling of storage & compute, and the difficulty in delivering low-latency analytics at scale, it is less suited for operational use cases and customer-facing use cases like Data Apps. The coupling of storage and compute, together with the need to predefine sort & dist keys for optimal performance, make it challenging to use for Ad-Hoc analytics.