Data Plane

The Data Plane uses a Thread-per-Core (TPC) architecture. Each CPU core runs an isolated, shared-nothing shard. Types are !Send by design — no data crosses core boundaries.

Execution Model

Each core owns:

• A dedicated event loop (no Tokio — raw TPC)
• io_uring submission and completion queues for NVMe I/O
• A jemalloc arena (no allocator lock contention)
• Lock-free telemetry ring buffers for metrics

There are no locks, no atomics, and no cross-core sharing. The Data Plane achieves predictable latency by eliminating all sources of contention.

What the Data Plane Does

• Executes PhysicalPlan nodes dispatched from the Control Plane
• Reads from NVMe via io_uring
• Runs SIMD-accelerated vector distance math
• Appends to the WAL (O_DIRECT)
• Evaluates BEFORE triggers (synchronous, same transaction)
• Emits WriteEvent records to the Event Plane via per-core ring buffers

What the Data Plane Does Not Do

• Spawn Tokio tasks
• Handle HTTP or pgwire connections
• Process AFTER triggers or CDC events
• Coordinate across shards

WriteEvent Emission

After each successful WAL commit, the Data Plane emits a WriteEvent containing:

• sequence — monotonic per-core counter
• collection — target collection name
• op — Insert, Update, or Delete
• row_id, lsn, tenant_id, vshard_id
• source — User, Trigger, RaftFollower, or CrdtSync
• new_value, old_value — for trigger and CDC consumption

Events are fire-and-forget — the Data Plane never blocks waiting for the Event Plane. If the ring buffer overflows, the Event Plane replays from the WAL.

Page Fault Hazard

A major page fault on an mmap region blocks the faulting TPC thread, stalling the entire shard's reactor. The Data Plane pre-fetches pages asynchronously via io_uring IORING_OP_READ or madvise(MADV_WILLNEED) before compute touches them.