bench-skill-rowbinary.yml

name: "skill: rowbinary benchmarks"

# Runs the RowBinary skill's benchmark suite (npm run bench) against a live
# ClickHouse server so the numbers cited externally (e.g. the @clickhouse/rowbinary
# blog post: RowBinary vs JSON, monomorphized vs API-combinator readers) have a
# timestamped, third-party-reproducible CI log + a machine-readable JSON artifact
# behind them — not just a maintainer's laptop.
#
# Decoupled from the correctness suite (tests-skill-rowbinary.yml): benchmarks are
# slower and their throughput numbers are noisy, so they shouldn't gate every PR.
# This runs on demand (workflow_dispatch) and on PRs that touch the skill or this
# workflow.

permissions: {}
on:
  workflow_dispatch:
  pull_request:
    paths:
      - .github/workflows/bench-skill-rowbinary.yml
      - skills/clickhouse-js-node-rowbinary/**
      - packages/datatype-parser/**

concurrency:
  group: "${{ github.workflow }}-${{ github.ref }}"
  cancel-in-progress: true

env:
  # Network resilience: npm's default of 2 fetch retries is not enough for
  # the transient registry errors (ECONNRESET) we regularly hit in CI.
  NPM_CONFIG_FETCH_RETRIES: "5"
  NPM_CONFIG_FETCH_RETRY_MINTIMEOUT: "10000"
  NPM_CONFIG_FETCH_RETRY_MAXTIMEOUT: "60000"
  NPM_CONFIG_FETCH_TIMEOUT: "600000"

jobs:
  benchmarks:
    timeout-minutes: 20
    runs-on: ubuntu-latest
    defaults:
      run:
        working-directory: skills/clickhouse-js-node-rowbinary
    steps:
      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3

      - name: Start ClickHouse (latest) in Docker
        uses: hoverkraft-tech/compose-action@11beaa1c2dae4e8ed7b1665aa074723b6cecb0e4 # v3.0.0
        env:
          CLICKHOUSE_VERSION: latest
        with:
          # The benchmarks only need the single-node HTTP server on :8123.
          compose-file: "docker-compose.yml"
          services: "clickhouse"
          down-flags: "--volumes"

      - name: Setup NodeJS
        uses: actions/setup-node@53b83947a5a98c8d113130e565377fae1a50d02f # v6.3.0
        with:
          node-version: 24

      - name: Record environment (CPU, Node, ClickHouse version)
        run: |
          echo "Node:       $(node -v)"
          echo "CPU cores:  $(nproc)"
          grep -m1 'model name' /proc/cpuinfo || true
          # The server may still be starting; poll briefly so the version is captured.
          for _ in $(seq 1 30); do
            chv=$(curl -s 'http://localhost:8123/?query=SELECT%20version()')
            [ -n "$chv" ] && break
            sleep 1
          done
          echo "ClickHouse: ${chv:-<unavailable>}"

      # Build + pack the in-repo @clickhouse/datatype-parser so the skill is
      # benchmarked against THIS checkout, not the version published to npm. The
      # committed package.json keeps a published range for end users; we only
      # override node_modules below (`npm install <tarball> --no-save`).
      - name: Pack local @clickhouse/datatype-parser
        working-directory: ${{ github.workspace }}
        run: |
          npm ci
          npm pack --workspace @clickhouse/datatype-parser --pack-destination "${{ runner.temp }}"

      - name: Install dependencies
        run: npm ci

      - name: Use the local datatype-parser build
        run: npm install "${{ runner.temp }}"/clickhouse-datatype-parser-*.tgz --no-save

      # Prints the human-readable tinybench summary (per-bench hz + "Nx faster"
      # lines, incl. the ledger RowBinary-vs-JSON comparison) to the job log, and
      # writes the same numbers as JSON for the uploaded artifact.
      - name: Run benchmarks
        run: npm run bench -- --outputJson "${{ runner.temp }}/rowbinary-bench-results.json"

      - name: Upload benchmark results (JSON)
        if: ${{ always() }}
        uses: actions/upload-artifact@bbbca2ddaa5d8feaa63e36b76fdaad77386f024f # v7.0.0
        with:
          name: rowbinary-bench-results
          path: ${{ runner.temp }}/rowbinary-bench-results.json
          if-no-files-found: error