StarRocks

Use it when

• The serving layer genuinely needs query-time joins against dimension tables; the cost-based optimizer makes denormalization optional rather than mandatory.
• One engine should serve seconds-fresh tables and accelerate Iceberg, Delta, Hudi, Hive, and Paimon lake data without copying it.
• Real-time UPDATE and DELETE on a primary key are required over streaming ingest.
• MySQL-compatible SQL keeps existing BI tools connected.

Think twice when

• You are also weighing Apache Doris; the projects share ancestry but have diverged substantially, so benchmark rather than inherit the choice.
• Denormalized slice-and-dice at extreme per-user query rates is the whole job; Pinot's index toolbox targets that directly.
• FE and BE (or compute node) groups are yours to operate; CelerData is the managed path.

How it runs

Front ends plan and coordinate; back ends store columnar data and execute. The 3.0 shared-data mode runs stateless compute nodes over object storage with local cache. Stream Load and Routine Load commit transactionally from Kafka; async materialized views with query rewrite carry the pre-aggregation. CelerData offers cloud and BYOC deployments.