Expand description
DataFusion ExecutionPlan wrappers around native cuDF query plans.
Host-sink wrappers materialize native results back into Arrow streams for DataFusion. Composed wrappers can pass GPU-resident fragment results between native subplans, but the final DataFusion-facing boundary still streams host Arrow.
ExecutionPlan::execute on these wrappers performs no GPU work and claims
no grant; it returns a stream whose first poll admits the attempt, runs the
fragment’s whole schedule on the device’s GPU execution pool, and keeps the
result on the device. Each later poll moves one bounded chunk to host
memory under a DataFusion memory-pool reservation. Every failure crosses
into DataFusion once, as DataFusionError::External around an adapter
Error whose facts keep the native classification; a failure after native
execution started ends the attempt without a CPU replay. A GPU-resident
producer’s output belongs to the context of the composed root that runs it
and is released with that root’s stream. The lifetime contract is in
website/docs/design/admission.mdx#resident-handoff-and-release.
Structs§
- Aggregate
Strategy Execution Options - Runtime threshold policy for grouped aggregate strategy selection.
- Algeon
Native Exec - One native execution unit: an admitted native plan, the GPU-resident producers it reads, and its host or GPU-resident output.
- Algeon
Native Execution Config - Adapter-side native execution configuration.
- Dynamic
Filter Execution Options - Runtime and route-discovery thresholds for native dynamic filters.
- Keyed
Hash Join Strategy Options - Threshold policy used by capability admission and runtime keyed-hash join strategy selection.
- Lakehouse
Byte Cache - Local byte cache for
s3://Parquet objects. - Lakehouse
Cache Fill Config - Service-wide resources used only by complete-object lakehouse cache fills.
- Planning
Policy - Thresholds applied during plan optimization, admission, and execution.
- QueryId
- Query-wide diagnostic correlation id.
- Runtime
Observation Config - Query-local recorder configuration.
- Runtime
Observation Limits - Recorder limits copied into the manifest.
- S3Source
Storage Profile - S3-compatible storage settings for source object-store fetches.
- Sealed
Attempt Graph Claim Calibration Snapshot - Immutable, typed graph claim-versus-peak rows projected by one sealed attempt.
- Source
Chunk Planning Options - Expert policy for local Parquet source chunk planning heuristics.
- Source
Chunking - Source chunking of native reads. Device capacity always comes from the admitted grant, so a chunking carries no device budget.
- Source
Scheduler Policy - Native source scheduling controls.
- TraceId
- Optional parent trace id used to join multiple query ids under one request.
Enums§
- Graph
Claim Calibration Execution - Typed graph-execution classification sealed with an attempt’s calibration rows. A physical classification requires the engine to record that the graph executor started; peak qualification remains fail-closed downstream.
- Graph
Claim Calibration NonPhysical Reason - Closed vocabulary for graph attempts that intentionally did not execute a graph algorithm locally.
- Lakehouse
Byte Cache Fill - When a local byte-cache miss for an
s3://Parquet object is filled. - Source
Chunk Planning Mode - How source chunk planning sizes local Parquet chunks. The chosen mode and the reason for it are recorded in each chunk plan.
- Statistical
Aggregate Execution Mode - Execution contract for floating statistical aggregates such as
STDDEV_SAMP,STDDEV_POP,VAR_SAMP, andVAR_POP.
Constants§
- NATIVE_
ALGEON_ DATAFUSION_ METRIC_ NAMES - Stable public DataFusion metric-name contract exposed by native cuDF execution plans.
- SERVER_
CAPTURE_ LOSS_ METRIC_ NAME - DataFusion counter name for server runtime-artifact capture loss.
Functions§
- below_
native_ admission - The plan below the admission node the native optimizer places at a
query’s root, or
planitself when it has none.