The short answer
Amazon Redshift is a distributed relational database designed for analytical queries on large datasets. It uses a massively parallel processing architecture where multiple servers execute queries at the same time.
+-------------------------------------------------+
| Leader Node |
| SQL endpoint, parser, optimizer, coordinator |
+-------------------------------------------------+
| high-speed private network
+----------------+----------------+
| |
+----------------+ +----------------+
| Compute Node | | Compute Node |
| Slice 0 | Slice 1 | | Slice 2 | Slice 3 |
+----------------+ +----------------+Leader node
Clients connect to a single leader node over standard PostgreSQL connections. The leader node does not execute user queries on table data. Instead, it parses incoming SQL, develops an optimized parallel execution plan, compiles the plan into C++ code, and distributes the compiled segments to compute nodes.
Compute nodes and slices
Compute nodes do the actual work. Each compute node is partitioned into multiple slices. A slice is an independent processing unit with its own dedicated memory and disk storage. If a cluster has 4 compute nodes and each node has 4 slices, 16 slices work on the query in parallel.
Storage formats: DC2 versus RA3
Older DC2 dense-compute nodes store data directly on local solid-state drives attached to each node. If you run out of disk space on DC2, you must add more compute nodes even if your CPU usage is low.
Modern RA3 nodes separate compute from storage. RA3 nodes keep hot data on local high-speed solid-state drives while offloading colder data blocks to Redshift Managed Storage, which is backed by Amazon S3. This allows your storage to grow automatically without forcing you to pay for idle compute nodes.
When explaining Redshift in an interview, emphasize the difference between the leader node and compute slices. A common junior mistake is assuming the leader node executes table queries directly.