Add Spark-based data generation for K8s (replaces MapReduce) - #248
Add Spark-based data generation for K8s (replaces MapReduce)#248wjxiz1992 wants to merge 7 commits into
Conversation
Signed-off-by: Allen Xu <allxu@nvidia.com>
Greptile SummaryThis PR adds K8s-native TPC-DS data generation support via a new PySpark-based approach ( Key additions:
Review findings: Confidence Score: 4/5
Important Files Changed
Sequence DiagramsequenceDiagram
participant User
participant Driver as Spark Driver<br/>(K8s Pod)
participant Executor as Spark Executor<br/>(K8s Pod)
participant FS as Filesystem<br/>(HDFS/S3/etc)
User->>Driver: spark-submit with --archives dsdgen.tar.gz#dsdgen
Driver->>Driver: Parse args (scale, parallel, range)
Driver->>Driver: Create RDD with child indices
Driver->>Executor: Distribute archive + tasks
Executor->>Executor: Extract dsdgen.tar.gz to SparkFiles
Executor->>Executor: run_dsdgen_and_read(child_index)
Executor->>Executor: subprocess.run(dsdgen -child N)
Executor->>Executor: Stream .dat files line-by-line
Executor->>FS: Write partitioned data (table_name=xxx/)
Executor-->>Driver: Task complete
Driver->>Driver: All tasks finished
Driver->>FS: rename_partition_dirs()
Driver->>FS: Rename table_name=xxx to xxx/
FS-->>Driver: Rename complete
Driver->>User: Data generation complete
Last reviewed commit: 3160717 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Pull request overview
This PR introduces a new Spark-based approach for TPC-DS data generation that works with Kubernetes and other Spark cluster managers, complementing the existing Hadoop MapReduce-based approach. The implementation distributes the dsdgen binary via Spark's --archives mechanism and generates data in parallel across Spark executors.
Changes:
- Added
nds_gen_data_spark.py: PySpark application that replaces MapReduce approach for data generation, supporting K8s, YARN, Standalone, and local modes - Updated README with documentation and examples for the new Spark-based data generation workflow
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 6 comments.
| File | Description |
|---|---|
| nds/nds_gen_data_spark.py | New PySpark script implementing distributed TPC-DS data generation via Spark executors with support for incremental generation and multiple filesystem backends |
| nds/README.md | Added comprehensive documentation section explaining Spark-based data generation with usage examples for K8s and other cluster managers |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Signed-off-by: Allen Xu <allxu@nvidia.com>
Signed-off-by: Allen Xu <allxu@nvidia.com>
- Use dsdgen.tar.gz instead of dsdgen.jar for archive format consistency - Add return value checks in rename_partition_dirs to prevent silent failures - Use shutil.rmtree for robust temp directory cleanup - Add encoding="utf-8" to file open calls - Fix Partitions print alignment Made-with: Cursor
| if len(parts) != 2: | ||
| print("ERROR: --range must be 'start,end'", file=sys.stderr) | ||
| sys.exit(1) | ||
| range_start, range_end = int(parts[0]), int(parts[1]) |
There was a problem hiding this comment.
Missing exception handling for int() conversion - passing non-numeric values like --range abc,def will crash with unclear ValueError
| range_start, range_end = int(parts[0]), int(parts[1]) | |
| try: | |
| range_start, range_end = int(parts[0]), int(parts[1]) | |
| except ValueError as e: | |
| print(f"ERROR: --range values must be integers: {e}", file=sys.stderr) | |
| sys.exit(1) |
- Add tpcds-gen/Dockerfile.dsdgen for cross-compiling dsdgen on Linux - Add Dockerfile.k8s-test for building Spark K8s image with datagen - Remove k8s_datagen_smoketest.sh script - Update README with cross-compilation, image build, and minikube testing instructions Made-with: Cursor
| # Set up Spark directory structure | ||
| ENV SPARK_HOME=/opt/spark | ||
| ENV PATH="${SPARK_HOME}/bin:${PATH}" | ||
| ENV PYTHONPATH="${SPARK_HOME}/python:${SPARK_HOME}/python/lib/py4j-0.10.9.7-src.zip:${PYTHONPATH}" |
There was a problem hiding this comment.
Hardcoded py4j-0.10.9.7-src.zip version assumes Spark 3.5.x - will break with other Spark versions that bundle different py4j versions
- Add try/except for --range int() parsing with clear error message - Remove hardcoded py4j version in Dockerfile.spark-k8s using symlink - Add base image build instructions and ARG to Dockerfile.k8s-test - Add kubectl wait for PVC binding in README minikube guide Made-with: Cursor
- Add guard in Dockerfile.spark-k8s to fail build if no py4j zip found - Add usage instructions and build context note to Dockerfile.spark-k8s - Document both Dockerfile options (k8s-test vs spark-k8s) in README with comparison table and step-by-step instructions Made-with: Cursor
wjxiz1992
changed the title
wjxiz1992
changed the title
|
@copilot code review[agent] review |
gerashegalov
left a comment
gerashegalov
left a comment
There was a problem hiding this comment.
Four actionable findings: two P1 correctness/usability issues and two P2 efficiency/test-instruction issues.
| # delete_<update>.dat, inventory_delete_<update>.dat | ||
| for special in ("inventory_delete", "delete"): | ||
| if fname.startswith(special): | ||
| table_name = special |
There was a problem hiding this comment.
[P1] Emit the delete datasets only once for update runs. Every parallel dsdgen child produces identical delete_<update>.dat and inventory_delete_<update>.dat content; the existing generator explicitly keeps only the first copy. This loop instead yields those files from every child, so --update N with parallelism 200 writes every delete row 200 times and corrupts the maintenance workload. Please retain these special files only from child 1 (including ranged generation), or deduplicate them before writing.
| --deploy-mode cluster \ | ||
| --conf spark.kubernetes.container.image=<spark-image> \ | ||
| --conf spark.executor.instances=10 \ | ||
| --archives tpcds-gen/target/lib/dsdgen.tar.gz#dsdgen \ |
There was a problem hiding this comment.
[P1] This K8s cluster-mode command cannot use these client-relative dependencies as written. Spark-on-Kubernetes requires spark.kubernetes.file.upload.path for a local application/archive and otherwise fails at submission with Please specify spark.kubernetes.file.upload.path property. Since the proposed image already contains both files, use --archives local:///opt/spark/work-dir/dsdgen.tar.gz#dsdgen and local:///opt/spark/work-dir/nds_gen_data_spark.py, or document/configure a Hadoop-compatible upload path. The K8s example in the Python docstring needs the same correction.
| # Convert to DataFrame: [table_name: string, value: string] | ||
| # partitionBy("table_name") writes separate directories per table WITHOUT shuffle — | ||
| # each task independently splits its output into per-table files. | ||
| df = spark.createDataFrame(all_data_rdd, ["table_name", "value"]) |
There was a problem hiding this comment.
[P2] Pass an explicit schema here. A list of column names makes PySpark infer the schema by calling all_data_rdd.first(), which launches a complete dsdgen child; the RDD is then recomputed for the actual write. At large scales this regenerates one entire child partition unnecessarily and can leave temporary output when the inference job stops the iterator early. For example: spark.createDataFrame(all_data_rdd, "table_name string, value string").
| }' | ||
|
|
||
| # 6. Wait for completion and check logs | ||
| kubectl wait --for=condition=Ready=false pod/nds-test --timeout=300s |
There was a problem hiding this comment.
[P2] This does not wait for the data-generation pod to finish. A newly created Pending pod already has Ready=false, so this command can return immediately and the following log/data checks can run before generation completes. Wait for .status.phase=Succeeded instead, and explicitly detect/report Failed, before checking the completion message and output.
Summary
This PR adds Kubernetes-native support for TPC-DS data generation by introducing a new PySpark-based approach (
nds_gen_data_spark.py) that replaces the Hadoop MapReduce method. The dsdgen binary is distributed to Spark executors via--archives, enabling parallel data generation on any Spark cluster manager (K8s, YARN, Standalone, local) without requiring YARN or MapReduce.Changes
Core
nds/nds_gen_data_spark.py— New PySpark application that:dsdgenbinary across Spark executors via--archivestable_name=xxxdirectories to plainxxxfor NDS pipeline compatibility--rangefor incremental generation,--overwrite, and--updatefor maintenance datasetsDockerfiles
nds/tpcds-gen/Dockerfile.dsdgen— Multi-stage build for cross-compilingdsdgenon Linux (handles GCC 10+-fcommonissue)nds/Dockerfile.k8s-test— Layers datagen scripts onto the official Spark Python image (recommended, usesARG BASE_IMAGE)nds/Dockerfile.spark-k8s— Standalone full Spark + PySpark image from a Spark distribution (build context =$SPARK_HOME)Configuration
nds/datagen_submit.template— spark-submit template for use withspark-submit-templateutilityDocumentation
nds/README.md— Comprehensive new sections:Dockerfile.dsdgenTesting
Verified on local minikube (Docker driver, ARM64):
Dockerfile.dsdgenDockerfile.k8s-testnds_gen_data_spark.pywithlocal[2]mode inside a K8s pod (scale=1, parallel=2)Migration Notes
nds_gen_data.py(MapReduce-based) is unchanged — this PR adds a parallel path, not a replacementnds_gen_data_spark.py