Per-core Worker Model#
Kernel modes (Tokio and Linux IOUring) run worker threads with local execution state and asynchronous I/O. worker_threads = 0 selects available parallelism; an explicit positive value selects the worker count. This worker count does not by itself promise CPU pinning.
Each worker owns session/transaction state and services data operations for its assigned partitions. Async tasks yield while awaiting I/O or internal RPCs. Local operations stay on the owning worker; partition routing can forward remote operations through the internal message bus.
With more than one worker, the server enables consecutive per-worker TCP ports starting at tcp_listen_port. Clients can discover worker identifiers, ports and partition placement through mcli server-topology, then open a session targeting the selected worker. routing_mode controls connection routing; it does not define a SQL hash-partitioning scheme.
See Partitioning for explicit range rules and placements, Suspendable Procedure Execution for guest scheduling, and Replication for node-level failover. Worker locality does not imply automatic live partition migration or distributed atomic commit.
Implementation: crates/db-kernel/server/src/server/serve/, crates/db-kernel/worker/src/, and the mudu_runtime backend startup code.