Expand description
Frontend-neutral GPU-resident data and reusable build handles. Frontend-neutral GPU-resident data, result, and reusable build handles.
Handles in this module are runtime/device-affine. Cloning shares the underlying native resources; host materialization is explicit.
Besides the cuDF frame, a handle carries the lifetimes that make the frame
safe to keep on the device: a ResidentOwnerLease for the bytes charged to
the producing attempt’s grant, transitive source owners for columns that alias
an input frame’s memory, a source-cache pin that keeps a cache entry from
being evicted while the frame aliases it, and graph provenance. A consumer
that reads a handle on its own stream wraps these owners with for_reader, so
the owner’s final release also waits for that stream’s work; dropping a handle
therefore never frees buffers a stream may still read. The release rules are
owned by website/docs/design/admission.mdx#resident-handoff-and-release.
Structs§
- GpuData
Frame - Owned cuDF DataFrame plus the placement observed when it entered the engine.
- GpuFiltered
Join Build Handle - Reusable cuDF filtered-join build artifact.
- GpuHash
Join Build Handle - Reusable cuDF hash-join build artifact.
- GpuHost
Batch Cursor - Bounded native GPU-to-host cursor retaining every owner required by the resident result until the final chunk is materialized or the cursor drops.
- GpuPrepared
Host Batch - Metadata-sized GPU output chunk awaiting a reserved device-to-host copy.
- GpuResult
Handle - GPU-resident query result bound to an Arrow schema and runtime.
Type Aliases§
- GpuResident
Input Map - Local lookup table for GPU-resident inputs consumed by a query.