Cudf Query Engine Builder

NVIDIAOfficialCreated by Joe SackDistributed by NVIDIA

Skill
Workflow
Software development
Apache-2.0
Medium risk
Has scripts

From the author

Build the first native, single-GPU cuDF path in a query engine, or complete its initial adapter. Use for libcudf C++ or cuDF Java implementation, buffer ownership, worker readiness and CPU/reference comparison, including identifying missing POC inputs. Excludes standalone dataframe analysis, cuDF mainline changes, production integration and distributed or multi-GPU execution. Bundled maintainer evaluations use CPU simulations and source review; they do not run CUDA.

— the description by Joe Sack

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill provides a structured workflow for query-engine engineers to integrate cuDF into their existing systems. It focuses on building a functional, single-GPU proof of concept (POC) that demonstrates native execution, correct buffer management, and proper asynchronous cleanup. The process emphasizes maintaining clear boundaries between the engine's orchestration responsibilities and cuDF's device-wide algorithms, ensuring that memory ownership, lifetime management, and producer-consumer dependencies are handled safely.

What it does

Maps query engine semantics to version-matched cuDF APIs.
Defines buffer ownership, lifetime, and release strategies for GPU data.
Implements adapter glue code to bridge the engine and libcudf or cuDF Java.
Establishes correct asynchronous execution patterns and completion synchronization.
Validates POC correctness against CPU-based reference implementations.
Provides a framework for reporting native execution evidence and identifying blockers.

When to use it

When you need to integrate cuDF into a query engine for the first time.
When developing a new native adapter for libcudf C++ or cuDF Java.
When you need to resolve buffer ownership or asynchronous cleanup issues in an existing integration.
When you need to verify that a query engine is correctly offloading work to a single GPU.

Advantages

Ensures safe memory management and exception handling in asynchronous GPU workflows.
Maintains separation of concerns between engine orchestration and cuDF algorithms.
Provides a clear path for validating native execution versus CPU simulations.
Reduces integration risks by focusing on incremental, verifiable POCs.

Usage examples

“Build a POC for a single-GPU filter operator in our C++ query engine using libcudf.”: The agent will guide you through mapping the filter semantics to cuDF, implementing the adapter glue, managing the buffer ownership, and verifying the result against your CPU reference implementation.
“How do I safely handle buffer cleanup in my cuDF adapter when an asynchronous call might throw?”: The agent will provide a pattern for retaining owners outside of throwing scopes and establishing completion before releasing resources, as detailed in the asynchronous cleanup guidelines.

Install

Download the skill as a ZIP:
https://mp-staging.autogineer.com/agent-skills/nvidia/cudf-query-engine-builder/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip cudf-query-engine-builder.zip -d .agents/skills

Documentation by Joe Sack

cuDF query-engine POC builder

Outcome and scope

Build the first native cuDF proof of concept in the target engine for a query-engine engineer. Produce working target-engine code, a runnable single-GPU program, and a correctness result, not only advice.

The POC may cover one operator, a sequence of operators, or a query selected for the POC. Do not reduce it to one operator by default. If a smaller increment is needed for feasibility or safe implementation, explain why and preserve the path to the requested POC.

For a request limited to an initial adapter correction or source review, deliver that requested work and state which build, native execution and comparison steps remain unperformed. Source review and CPU simulation cannot establish acceptance of the native POC.

Gather inputs and choose sources

Before editing, identify:

the target engine or adapter code and build entry point;
the data source or fixture, its schema, and how the engine reads it today (its own scan layer, or none);
the query, operator sequence, or CPU implementation;
the expected output or CPU/reference result and target-engine semantics; and
the target-engine language, any existing libcudf C++ or cuDF Java integration code, and the cuDF version when identifiable from the code or build environment.

Determine the cuDF entry and exit interfaces from the supplied engine code. Define schema conversion, buffer ownership and release, lifetime, and producer/consumer readiness as part of the POC. Ask the engineer only when the code and tests cannot establish a required choice.

Ask only when a missing input changes the code, correctness, safety, or ability to build and run. Gather only the environment facts needed for this POC.

If the supplied query, CPU behavior, target-engine semantics, and expected result disagree, ask which behavior controls. Do not choose silently.

Use target-engine code and tests, installed public headers, and version-matched official cuDF documentation and source. Treat partner discussions as problem signals, not technical authority.

Build the POC

Establish the named CPU or reference result. Run the reference implementation when it is executable.
Map the requested semantics to public, version-matched cuDF APIs. Handle relevant null, type, order, error, ownership, lifetime, and stream behavior. When native calls can leave GPU work pending, read cleanup after asynchronous work before implementing ownership and exception paths.
Implement the requested operators and minimal engine or adapter glue without changing cuDF mainline. If the required public cuDF API is missing, preserve any buildable engine-side work, document the smallest cuDF gap, and report BLOCKED when the POC cannot proceed safely.
Build a runnable program and execute the workload on one GPU. Record exact commands and working directories.
State the applicable row-order handling, null and type equality, and floating-point tolerance before comparing the output with the named CPU or reference result.
Fix evidenced defects, then rebuild, rerun, and recompare. Continue until the checks pass or a named blocker remains. Report unsafe boundaries, missing prerequisites, unresolved semantics, and larger engine changes.

Before accepting an asynchronous path, inspect every operation that can throw after submission, including native calls, output allocation and result-wrapper construction. C++ destroys objects declared inside a try block before entering its catch handler: a wait in that handler cannot protect those destroyed owners. Keep input, intermediate and output owners outside the throwing scope, and establish completion before releasing them on failure. On success, wait after the last queued operation, including a device-to-host copy, before the caller reads the result. A wait before the copy does not make its destination ready.

Honor the supplied caller's completion promise even when filtering produces no rows. An empty result can still follow queued work or depend on input readiness. For a result handed between workers, retain its buffer, stream and resource owners and establish the producer-to-consumer dependency. Follow the ownership patterns and validation steps in cleanup after asynchronous work.

Exercise a recoverable allocation failure while work is pending, assert that no owner is released early, and check a subsequent valid call when the engine promises recovery. Treat a failed ownership or readiness check as a failed POC even when normal output values match.

If a required codebase, toolchain, compatible cuDF runtime, GPU, or reference result is unavailable and prevents a required step, report BLOCKED, name the missing prerequisite, preserve completed work, and give the exact next command or check. Do not report the blocked step as run or passed.

POC acceptance criteria and handoff

POC passes only when the source change builds, the program exits successfully, its output matches the named CPU or reference result under the stated comparison policy, and a reliable engine signal or runtime observation confirms that the requested result-producing work ran through cuDF.

Report the native-execution evidence separately from numerical agreement. If reliable confirmation is unavailable, mark native execution unresolved and do not report an overall POC pass. Do not infer native execution from GPU availability or elapsed time. A profiler is not required when another reliable engine signal or runtime observation establishes native cuDF execution.

Return:

the patch and runnable program path;
exact build and run commands with their working directories;
the inputs, selected binding and cuDF version, named CPU or reference result, comparison policy, and observed result;
any available cuDF-path observation; and
unresolved questions and larger engine work, with facts separated from inference.

Keep cuDF and engine responsibilities separate

cuDF provides device-wide algorithms for one GPU. The engine owns query-plan and pipeline orchestration, spill, production fallback, multi-GPU coordination, and distributed execution.

I/O sits on the boundary, and the POC must pick one of two paths and say which. cuDF ships data sources that read local files, remote files, and remote objects directly. The engine can either pass cuDF a path or data source and let cuDF perform the read, or keep its own I/O stack and hand cuDF the bytes in host buffers, usually pinned. Choose the path that matches the supplied engine code: if the engine already has a scan or storage layer that produces host buffers, feed those buffers to cuDF; if it does not, use a cuDF data source. Record the choice, the buffer type (pinned or pageable), and who frees each buffer in the handoff. Caching, prefetch, retries, credentials, and format negotiation stay with the engine and are out of scope.

Write the engine or adapter code needed for the requested POC, including who owns and releases each buffer and how consumers wait for producer work to complete. Report larger engine work separately; do not implement it in the POC.

Performance and profiling

Profiling is optional for an initial POC and is never a blocker: if Nsight Systems cannot be installed or run, report that and continue. When it is available, capture one nsys profile of the run and check that the GPU is busy in the timed region. Report idle gaps, host-side stalls, and synchronous copies; do not fix them in the POC unless asked. Preserve the exact run command and workload inputs so the run can be profiled later.

When performance work is requested, profiling belongs in the same tuning loop. Benchmark the same logical work under stated conditions. Obvious code changes may be validated with benchmark results, but broader diagnosis and performance claims require profile evidence. If profiling cannot run, report the blocker rather than inferring bottlenecks from code alone.

A GPU result that remains slower than the CPU result is not automatically a failed POC. Preserve the implementation and report the measurements, profile evidence collected, observed bottlenecks, and remaining engine work.

Evaluation files

The evals/ directory is for maintainers evaluating this skill. It contains incomplete adapter starters and CPU models used to test host ownership and readiness; these are not cuDF implementations or native-execution evidence. Do not run these fixtures as part of an engineer's POC unless the user requested skill evaluation. Read evals/EVAL.md before evaluating them and use its isolated execution procedure.

Files

ATTRIBUTION.mdAdded by AutoGineer977 B
BENCHMARK.md10.7 KiB
LICENSE29.8 KiB
LICENSE-APACHE11.2 KiB
LICENSE-CC-BY-4.018.3 KiB
SKILL.md9.1 KiB
THIRD_PARTY_NOTICES.md3.2 KiB
evals/.reference/Dockerfile.cudf-cpp272 B
evals/EVAL.md7.4 KiB
evals/config.yml505 B
evals/evals.json19.7 KiB
evals/files/async-copy-cleanup/TASK.md3.1 KiB
evals/files/async-copy-cleanup/adapter.hpp683 B
evals/files/async-copy-cleanup/run_checks.pyScript4.9 KiB
evals/files/async-copy-cleanup/sim_backend.hpp6.2 KiB
evals/files/async-copy-cleanup/test_adapter.cpp3.6 KiB
evals/files/worker-result-handoff/TASK.md3.3 KiB
evals/files/worker-result-handoff/adapter.hpp1.4 KiB
evals/files/worker-result-handoff/run_checks.pyScript4.9 KiB
evals/files/worker-result-handoff/sim_backend.hpp6.2 KiB
evals/files/worker-result-handoff/test_adapter.cpp4.1 KiB
evals/grader.pyScript11.4 KiB
references/asynchronous-cleanup.md7.7 KiB
skill-card.md5.1 KiB
skill.oms.sig9.1 KiB

references/asynchronous-cleanup.md

Complete native work before releasing its owners

Read this before implementing a path that uses asynchronous copies, borrowed events, multiple streams, engine-owned device buffers, or an owning result. Establish the caller's rules for input recycling, host-readable output, result destruction and failed CUDA contexts from its code and tests.

A stream view, table view or resource reference does not keep its owner alive. Retain the objects that own input storage, native intermediates, output storage, the stream and the allocator for every dependent use. Check the installed public headers for the types and APIs being used.

Choose what return means

A GPU-to-GPU handoff may return pending work when the caller receives both its owners and a readiness dependency. The consuming stream must wait on that dependency before using the buffers.

A function that promises host-readable output must finish its device-to-host copies before returning. cudaStreamWaitEvent(consumer, ready, 0) orders future work on the consumer stream; it does not wait for host completion. Check cudaStreamSynchronize(consumer) or a completion event that covers the copies before exposing the host buffers. Keep borrowed events unchanged and preserve their owner's lifetime. CUDA stream management.

Apply the caller's completion rule to empty output too. If the caller requires its producer event to be complete at return, a zero-row branch must establish that completion even though it has no bytes to copy. Avoid an early return before the common completion step. Do not impose this synchronous rule on an interface that explicitly permits pending empty results.

Put retained owners outside the throwing scope

A native call can enqueue GPU work and then throw because a host allocation fails. Later output allocations, container growth and result-wrapper construction can also throw. Keeping owners only in a successful result does not protect these paths.

In C++, an object declared inside a try block is destroyed before that block's catch begins. A wait in the catch is too late if the needed owner was already destroyed. Declare the owner bundle and any partially constructed output owners before the protected try, or use a correctly ordered scope guard that runs before those owners are destroyed. Prepare owner-container capacity and failure-retention storage before submission when practical. If work is already pending on entry, first retain its existing owners through a nonthrowing transfer or an existing caller lease; an allocation while constructing the protection must not open another release path.

This structural C++ example uses engine-specific placeholder types and operations. It shows the required scopes, not a cuDF API or a reusable implementation:

RetainedOwners retained = take_existing_owners_noexcept();
// retained outlives the try block and catch. It also holds partial outputs.
try {
    // This call may enqueue work before throwing. Do not arm cleanup afterward.
    submit_native_work(retained);
    construct_result_in(retained);  // This can allocate and throw too.
    establish_required_completion_or_result_dependency(retained);
    return transfer_complete_result_ownership(retained);
} catch (...) {
    // All required owners are still alive here.
    auto status = cudaStreamSynchronize(retained.stream());
    if (status != cudaSuccess) {
        // Engine policy must retain these owners without a new allocation,
        // preserve both errors, and stop normal unwinding until safe.
        enter_engine_failed_context_policy(status, retained);
    }
    throw;  // Propagate the original error after completion is established.
}

The failed-context operation above is a required engine policy, not a supplied helper. It must not return to ordinary unwinding while owners can still be used. If the engine has no such policy, identify the missing recovery boundary instead of inventing one. A checked wait may report an earlier asynchronous error; a no-throw wait does not by itself prove safe cleanup. For multiple streams, complete every dependent use or first join those uses onto the stream being waited on.

An event recorded after a native call returns cannot protect an exception thrown before the record. Stream destruction can return while queued work remains, so destroying a stream is not a completion mechanism. A valid synchronous POC may wait on success and exception paths. An asynchronous result must retain owners and a usable completion dependency through its final consumer.

Complete downloads before returning host buffers

When the host-output interface permits it, allocate all output buffers before submitting the first download. This removes the avoidable failure window in which the first copy is pending and allocation of a later output fails. It does not remove the need to protect other exceptions.

The following body belongs inside the protected scope above, with output and the native-result owners retained outside that scope:

check_cuda(cudaStreamWaitEvent(download_stream, borrowed_ready, 0));
if (row_count != 0) {
    allocate_all_host_outputs(output);
    enqueue_all_downloads(native_result, output, download_stream);
}
check_cuda(cudaStreamSynchronize(download_stream));  // Also for empty output.
return output;

The unconditional completion point applies when this interface promises ready output and completed producer work. On an allocation or copy error, the outer catch must keep both host and device owners alive while already submitted work completes. Resetting output owners before waiting reverses the required order.

Check the actual allocator requirement

Use the engine's stream and requested resource for the allocations covered by the API contract. Read the installed parameter documentation: an mr parameter documented for returned memory is not a guarantee that every internal temporary uses that resource. Treat counts of temporary or default-resource allocations as observations unless the engine explicitly restricts them. Do not reject a valid result solely because an internal temporary used another supported resource.

Retain the allocator owner until every allocation made from it has been destroyed. Retain the stream owner needed by stream-ordered deallocation. In a result object, declaration order matters because members are destroyed in reverse order: declare the context owner before the dependent buffers. cuDF's memory-resource interfaces distinguish returned and temporary allocations; check the version in use rather than changing a process-wide default as a workaround. cuDF resource guidance.

Test the required handoffs and failures

Exercise a delayed producer and a delayed final download, so a missing wait cannot pass because work happened to finish early. Test zero-row output under the same caller readiness rule as nonempty output. For successful calls, compare actual returned values with the named independent reference.

Inject a recoverable allocation failure after native submission while work is demonstrably pending. Cover later output allocations and result-owner construction, including library allocations when the test mechanism supports them. Observe whether an owner is released before its dependent use finishes. Check a subsequent valid call when the interface promises recovery. Also run a failure before submission to distinguish cleanup defects from a broken setup.

Accept synchronous or asynchronous implementations when the caller allows them. Keep host allocation failure, device out-of-memory and fatal CUDA errors separate. A passing controlled host-allocation test does not establish recovery from the other failures. Report what was exercised and what remains untested.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs Python

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs
Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

Joe Sack

Publisher

NVIDIA (Official)

Distributed by

NVIDIA

Source

NVIDIA/skills · skills/cudf-query-engine-builder · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-08

AutoGineer is not affiliated with or endorsed by Joe Sack or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download cudf-query-engine-builder.zip