[KanchanShu]

This PR add Linux ODBC Performance Benchmark Pipeline

Summary

Adds a dedicated Linux performance benchmark pipeline for the BigQuery ODBC driver.

Changes

• Builds and runs performance_test using CMake.
• Benchmarks the Existing, Current, and Main Google ODBC drivers.
• Runs each benchmark multiple times and uses the median execution time.
• Adds benchmark_results.py to parse results and generate a comparison table with performance percentage changes.
• Downloads driver artifacts and uploads benchmark results to GCS.
• Adds a Cloud Build configuration to run the benchmark pipeline on main.

Result

Provides automated Linux performance comparison between the existing driver, current branch, and main to identify performance improvements or regressions.

Benchmark Results: Linux Performance Benchmark Results

[shivamd-gpartner]

new_timestamp_table test case seems failing on our driver @KanchanShu can you see that

[shivamd-gpartner]

nit: please make it 2026

[KanchanShu]

Done

[shivamd-gpartner]

seems duplicate declaration

[KanchanShu]

removed it

[shivamd-gpartner]

@sachinpro this is pending your review, please take a look

[sachinpro]

/gcbrun

[sachinpro]

/gcbrun

[sachinpro]

/gcbrun

[sachinpro]

/gcbrun

[sachinpro]

/gcbrun

[sachinpro]

Few changes to the original PR

1. Split Linux Benchmarks into 3 Separate Cloud Build Triggers (Matching Windows Setup):

   • Google Driver Benchmark (linux-bq-driver-benchmark-ci / ci/cloudbuild/builds/linux-bq-driver-benchmark.sh):
     • Downloads gs://odbc-integration-builds/<SANITIZED_BRANCH>/libgoogle_cloud_odbc_bq_driver.so, runs the benchmark suite (BQ_DRIVER_INTEGRATION_TESTS=ON), and uploads performance_benchmark_results_BqDriver.txt to gs://bq-dev-tools-testing-drivers/odbc-perf/<SANITIZED_BRANCH>/linux_results/.
     • On main, it is triggered automatically by ci/cloudbuild/builds/integration-production-bq-driver-dm.sh right after libgoogle_cloud_odbc_bq_driver.so is uploaded to GCS (preventing the race condition where the benchmark previously started before the .so build finished). On other branches, it is run manually.
   • Existing Driver Benchmark (linux-existing-driver-benchmark-ci / ci/cloudbuild/builds/linux-existing-driver-benchmark.sh):
     • Manual trigger only (BUILD_SHARD="Core"). Runs the benchmark suite against the Existing (Simba) driver (BQ_DRIVER_INTEGRATION_TESTS=OFF) and uploads performance_benchmark_results_Core.txt to gs://bq-dev-tools-testing-drivers/odbc-perf/<SANITIZED_BRANCH>/linux_results/.
   • Benchmark Results Table Generator (linux-benchmark-results-ci / ci/cloudbuild/builds/linux-benchmark-results.sh):
     • Automatically triggered at the end of either linux-bq-driver-benchmark-ci or linux-existing-driver-benchmark-ci (and can also be run manually).
     • Avoids re-running main benchmarks on every run; instead downloads performance_benchmark_results_BqDriver.txt (current branch), performance_benchmark_results_Core.txt (current branch), and main/linux_results/performance_benchmark_results_BqDriver.txt from GCS, generates benchmark_summary_table.txt, and uploads it to gs://bq-dev-tools-testing-drivers/odbc-perf/<SANITIZED_BRANCH>/linux_results/benchmark_summary_table.txt.

2. Aligned GCS Artifact & Result Paths with Windows:

   • Driver builds (gs://odbc-integration-builds): Updated ci/cloudbuild/builds/integration-production-bq-driver-dm.sh and ci/cloudbuild/builds/linux-bq-driver-benchmark.sh to upload and fetch the Linux driver .so at gs://odbc-integration-builds/<SANITIZED_BRANCH>/libgoogle_cloud_odbc_bq_driver.so using the same branch sanitization (sed -E 's/[^a-zA-Z0-9-]/_/g') as Windows MSI builds.
   • Benchmark results (gs://bq-dev-tools-testing-drivers): Standardized Linux result filenames (performance_benchmark_results_BqDriver.txt, performance_benchmark_results_Core.txt, benchmark_summary_table.txt) under gs://bq-dev-tools-testing-drivers/odbc-perf/<SANITIZED_BRANCH>/linux_results/.

3. Shared benchmark_results.py Between Linux and Windows & Added Commit SHA:

   • Updated ci/cloudbuild/builds/lib/benchmark_results.py to:
     • Parse [ METRIC ] lines (Time to first byte, Iteration time) and normalize [ OK ] lines for DataFetchPerformanceParamTest to <base_name> - Total execution time.
     • Group and sort data-fetch tests by (table_name, 10k -> 100k -> 1M, metric_order) ahead of catalog tests.
     • Include Commit ID: \<commit_sha>`at the top ofbenchmark_summary_table.txt (--commit-sha`).
     • Gracefully handle missing or partial input files by warning and displaying N/A instead of raising FileNotFoundError, ensuring failures in any test or shard do not block summary generation.
   • Replaced the inline Python parser in .github/workflows/test-runner.yml (windows-benchmark-results) with a direct call to ci/cloudbuild/builds/lib/benchmark_results.py.

[sachinpro]

Thanks!

I made changes to make the pipeline more consistent with the existing windows benchmarks. We are utilizing the same benchmark_results.py now for windows and linux.

I also split the benchmark for Google and existing driver, so they can be run independently(and parallelly).