| Execution model | Vectorized (2k values per column) | Row-at-a-time | Row-at-a-time (some batching) | Vectorized, columnar |
| SIMD usage | Yes (AVX/NEON intrinsics, auto-vectorization) | No | Limited | Yes (broad SIMD usage in scans & aggregates) |
| Storage layout | Columnar (.duckdb, Parquet, Arrow) | Row store | Row store | Columnar (native format + Parquet/ORC integration) |
| Best for | Embedded analytics (OLAP on local data/Parquet) | Small apps, embedded OLTP | Full-featured OLTP/OLAP hybrid | Distributed, large-scale OLAP |
| Parallelism | Yes: in-process morsel-driven scheduler | No (single-threaded) | Yes (process-based, higher overhead) | Yes: massively parallel, multi-node clusters |
| Deployment | Embedded, in-process | Embedded, in-process | Client/server daemon | Distributed client/server (scale-out cluster) |
| I/O pushdown | Yes: projection + predicate pushdown for Parquet/Arrow | N/A | Limited (index-based) | Yes: column/partition pruning, compression indexes |
| Overhead | Very low (same-process threads, no IPC) | Very low (single-threaded) | Higher (IPC, process-per-query) | Moderate (network RPC, cluster coordination) |
| OLTP performance | Weak | Strong | Strong | Weak (not designed for transactional workloads) |
| OLAP performance | Excellent (SIMD + columnar + vectorized, embedded) | Poor | Good (but slower vs columnar) | Excellent at scale (terabytes → petabytes, distributed) |
| Scalability | Single-machine (fits well into laptops, servers) | Single-machine | Scales vertically with replication/partitioning | Horizontal scale-out (clusters across 100s of nodes) |
| Common use cases | Data science, local analytics, Python/R notebooks | Mobile apps, small embedded DBs | Production transactional DBs, hybrid workloads | Big data analytics, event logs, monitoring, real-time dashboards |
