ci: GitHub actions performance (#187)

williscool · cursoragent · web-flow · commit 976dcbdf362a · 2026-01-16T20:21:13.000-05:00
* docs: add GitHub Actions performance optimization plan

Analyzed run 21023241683 to identify optimization opportunities:
- Remove debug dry-run steps (~6.5 min savings)
- Remove gradle tasks list from common-setup (~5-10 min savings)

Also documented known issues that aren't actionable:
- connectedAndroidTest rebuild limitation (AGP hardcoded deps)
- Integration test shard flakiness (runner resource contention)
- set_build_datetime job for consistent artifact naming

Co-authored-by: wharris &lt;wharris@upscalews.com&gt;

* perf(ci): remove debug dry-run steps and gradle task list

Remove diagnostic steps that added ~11-16 minutes to CI runs:

- Remove 'Debug Main Build Task Graph' dry-run step (~2 min savings)
- Remove 'Debug Unit Test Task Graph' dry-run step (~2.5 min savings)
- Remove 'List Gradle Tasks' step from common-setup (~5-10 min aggregate)

These were purely diagnostic and not needed for normal CI operation.
Gradle cache is warmed up by subsequent build steps anyway.

Co-authored-by: wharris &lt;wharris@upscalews.com&gt;

* docs: mark implemented optimizations as complete

Co-authored-by: wharris &lt;wharris@upscalews.com&gt;

* perf(ci): skip codegen for integration test shards

Integration test shards download pre-built APKs and run am instrument
directly - they don't need React Native codegen.

Changes:
- Add skip_codegen input to common-setup action (default false)
- Integration test shards now pass skip_codegen: true
- Update plan doc with new optimization

Expected savings: ~3 min × 4 shards = ~12 min aggregate

Co-authored-by: wharris &lt;wharris@upscalews.com&gt;

* docs: update plan with verified results

All optimizations implemented and verified:
- Removed debug dry-run steps
- Removed gradle tasks list
- Skip codegen for integration test shards

Measured ~12-13 min savings per CI run with warm caches.

Co-authored-by: wharris &lt;wharris@upscalews.com&gt;

---------

Co-authored-by: Cursor Agent &lt;cursoragent@cursor.com&gt;
diff --git a/.github/actions/common-setup/action.yml b/.github/actions/common-setup/action.yml
@@ -42,6 +42,10 @@ inputs:
     description: 'Optional cache key'
     required: false
     default: ""
+  skip_codegen:
+    description: 'Skip React Native codegen (for jobs that use pre-built APKs)'
+    required: false
+    default: "false"
 
 outputs:
   build-outputs-cache-hit:
@@ -272,16 +276,10 @@ runs:
       run: cd android && chmod +x gradlew
       shell: bash
 
-    # List all available Gradle tasks (also warms up Gradle cache)
-    - name: List Gradle Tasks
-      run: |
-        cd android && \
-        ./gradlew tasks --all > gradle-tasks.txt
-        cat gradle-tasks.txt
-      shell: bash
-
     # Run codegen for all React Native libraries (required for New Architecture)
+    # Skip for jobs that use pre-built APKs (integration test shards)
     - name: Generate React Native Codegen
+      if: inputs.skip_codegen != 'true'
       run: |
         cd android && \
         ./gradlew generateCodegenSchemaFromJavaScript generateCodegenArtifactsFromSchema --parallel --max-workers=$MAX_WORKERS
diff --git a/.github/workflows/actions.yml b/.github/workflows/actions.yml
@@ -111,24 +111,6 @@ jobs:
         node_version: "22.x"
         arch: ${{ matrix.arch }}
 
-    # Debug Main Build Task Graph
-    - name: Debug Main Build Task Graph
-      run: |
-        cd android && \
-        echo "=== Main Build Task Graph ===" && \
-        ./gradlew -PBUILD_ARCH="${BUILD_ARCH}" \
-                  -PreactNativeArchitectures="${BUILD_ARCH}" \
-                  ":${{ env.main_project_module }}:assemble${{ matrix.arch == 'arm64-v8a' && 'Arm64V8a' || 'X8664' }}Debug" \
-                  ":${{ env.main_project_module }}:assemble${{ matrix.arch == 'arm64-v8a' && 'Arm64V8a' || 'X8664' }}Release" \
-                  ":${{ env.main_project_module }}:bundle${{ matrix.arch == 'arm64-v8a' && 'Arm64V8a' || 'X8664' }}Release" \
-                  ":${{ env.main_project_module }}:assemble${{ matrix.arch == 'arm64-v8a' && 'Arm64V8a' || 'X8664' }}DebugAndroidTest" \
-                  --dry-run \
-                  --info \
-                  --console=verbose \
-                  --parallel --max-workers=$MAX_WORKERS --build-cache
-      env:
-        BUILD_ARCH: ${{ matrix.arch }}
-
     - name: Ccachify the Native Modules
       shell: bash
       run: |
@@ -440,22 +422,6 @@ jobs:
         path: coverage/
         retention-days: 90
 
-    # Debug unit test task graph
-    - name: Debug Unit Test Task Graph
-      run: |
-        cd android && \
-        echo "=== Unit Test Task Graph ===" && \
-        ./gradlew -PBUILD_ARCH="${{ matrix.arch }}" \
-                  -PreactNativeArchitectures="${{ matrix.arch }}" \
-                  :${{ env.main_project_module }}:test${{ matrix.arch == 'arm64-v8a' && 'Arm64V8a' || 'X8664' }}DebugUnitTest \
-                  :${{ env.main_project_module }}:create${{ matrix.arch == 'arm64-v8a' && 'Arm64V8a' || 'X8664' }}DebugUnitTestCoverageReport \
-                  --dry-run \
-                  --info \
-                  --console=verbose \
-                  --continue
-      env:
-        BUILD_ARCH: ${{ matrix.arch }}
-
     # Run the unit tests
     - name: Run Unit Tests
       run: |
@@ -739,6 +705,7 @@ jobs:
         fetch-depth: 1
 
     # Execute the common setup with emulator configuration
+    # Skip codegen since we use pre-built APKs (saves ~3 min per shard)
     - name: Common Setup
       id: common-setup
       uses: ./.github/actions/common-setup
@@ -751,6 +718,7 @@ jobs:
         android_target: ${{ env.ANDROID_TARGET }}
         android_profile: ${{ env.ANDROID_PROFILE }}
         run_emulator_setup: "true"
+        skip_codegen: "true"
 
     - name: Make scripts executable
       run: |
diff --git a/docs/dev_todo/github_actions_performance.md b/docs/dev_todo/github_actions_performance.md
@@ -0,0 +1,158 @@
+# GitHub Actions Performance Optimizations
+
+## Overview
+
+Analysis of GitHub Actions workflow performance based on run 21023241683 (January 15, 2026).
+
+### Key Timing Findings
+
+| Job | Duration | Key Bottlenecks |
+|-----|----------|-----------------|
+| Build Android App (x86_64) | ~32 min | Common Setup: 10 min, Dry-run: 2 min, Build: 18.5 min |
+| Build Android App (arm64) | ~31 min | Similar breakdown |
+| Unit Tests | ~20 min | Common Setup: 10 min, Dry-run: 2.5 min, Tests: 7 min |
+| Verify Connected Android Test | ~26 min | Common Setup: 10 min, Build+Test: 15 min |
+| Integration Test (per shard) | ~12-14 min | Common Setup: 7-8 min, Tests: 3-4 min |
+
+---
+
+## Actionable Optimizations
+
+### 1. Remove "Debug Main Build Task Graph" Dry-Run Step
+
+**Location**: `.github/workflows/actions.yml` lines 115-131
+
+**Current code**:
+```yaml
+- name: Debug Main Build Task Graph
+  run: |
+    cd android && \
+    echo "=== Main Build Task Graph ===" && \
+    ./gradlew ... --dry-run ...
+```
+
+**Action**: Remove this step entirely (or make conditional on a debug input flag).
+
+**Measured impact**: ~2 minutes per build job × 2 architectures = **~4 minutes saved**
+
+**Risk**: Very low - purely diagnostic output
+
+---
+
+### 2. Remove "Debug Unit Test Task Graph" Dry-Run Step
+
+**Location**: `.github/workflows/actions.yml` lines 444-457
+
+**Current code**:
+```yaml
+- name: Debug Unit Test Task Graph
+  run: |
+    cd android && \
+    echo "=== Unit Test Task Graph ===" && \
+    ./gradlew ... --dry-run ...
+```
+
+**Action**: Remove this step entirely.
+
+**Measured impact**: ~2.5 minutes saved
+
+**Risk**: Very low - purely diagnostic output
+
+---
+
+### 3. Remove "List Gradle Tasks" from Common Setup
+
+**Location**: `.github/actions/common-setup/action.yml` lines 276-281
+
+**Current code**:
+```yaml
+- name: List Gradle Tasks
+  run: |
+    cd android && \
+    ./gradlew tasks --all > gradle-tasks.txt
+    cat gradle-tasks.txt
+```
+
+**Action**: Remove this step entirely. It runs on every job and is purely diagnostic.
+
+**Measured impact**: ~30-60 seconds per job × 10+ jobs = **~5-10 minutes aggregate saved**
+
+**Risk**: Low - only diagnostic, Gradle gets warmed up by other steps anyway
+
+---
+
+### 4. Skip Codegen for Integration Test Shards
+
+**Location**: `.github/workflows/actions.yml` integration-test job + `.github/actions/common-setup/action.yml`
+
+**Problem**: Integration test shards download pre-built APKs and run `am instrument` directly - they don't build anything. But common-setup still runs React Native codegen (~3 min).
+
+**Solution**: Added `skip_codegen` input to common-setup. Integration test shards now pass `skip_codegen: "true"`.
+
+**Measured impact**: ~3 min × 4 shards = **~12 min aggregate saved**
+
+**Risk**: Low - integration test shards only run pre-built APKs via `am instrument`
+
+---
+
+## Estimated Total Savings
+
+| Optimization | Time Saved |
+|--------------|------------|
+| Remove build dry-run | ~4 min |
+| Remove unit test dry-run | ~2.5 min |
+| Remove gradle tasks list | ~5-10 min |
+| Skip codegen for integration test shards | ~12 min |
+| **Total** | **~23-28 minutes per CI run** |
+
+---
+
+## Known Issues (Not Actionable)
+
+### `verify-connected-android-test` Rebuilds Everything
+
+**Why it can't use pre-built APKs**: The Android Gradle Plugin's `connectedAndroidTest` task has hardcoded dependencies on `assembleDebug` and `assembleDebugAndroidTest`. There is no official way to skip these build steps - the AGP's `DeviceProviderInstrumentTestTask` explicitly requires them.
+
+**Previous investigation** (early 2025): Attempted to use pre-built APKs with `connectedAndroidTest`, including reviewing AGP source code. No viable solution found.
+
+**Current workaround**: The main `integration-test` job uses `am instrument` directly with pre-built APKs (via `matrix_run_android_tests.sh`). The `verify-connected-android-test` job exists as a sanity check to ensure `connectedAndroidTest` still works if ever needed, but it's not required for PR merges.
+
+### Integration Test Shard Flakiness/Retries
+
+**Observed**: All 4 shards were on "attempt 3" in run 21023241683. Shard 0 failed twice before succeeding (started 42 min after other shards).
+
+**Root cause**: Resource contention on GitHub Actions runners when multiple emulators run simultaneously. This is a known limitation of the free tier runners.
+
+**Mitigation options**:
+- Reduce shard count (current: 4, previously tried 8 which was too flaky)
+- Use paid runners with more resources
+- Accept occasional retries as cost of parallelism
+
+### `set_build_datetime` Job
+
+The separate job ensures timestamp consistency across all jobs that use it for artifact naming. While it adds ~30 seconds of runner startup overhead, removing it would require a different mechanism to share a consistent timestamp across jobs.
+
+---
+
+## Implementation Checklist
+
+- [x] Remove "Debug Main Build Task Graph" step from `actions.yml`
+- [x] Remove "Debug Unit Test Task Graph" step from `actions.yml`
+- [x] Remove "List Gradle Tasks" step from `common-setup/action.yml`
+- [x] Add `skip_codegen` input to common-setup for jobs using pre-built APKs
+- [x] Skip codegen for integration-test shards (~3 min × 4 shards = ~12 min saved)
+- [x] Verify CI still passes after changes
+- [x] Monitor CI run times to confirm improvements
+
+## Results
+
+**Measured improvements (comparing runs with warm caches):**
+
+| Job | Before | After | Savings |
+|-----|--------|-------|---------|
+| Unit Tests | 17m | 14m | -3m |
+| Integration Shard 1 | 12m | 8m | -4m |
+| Integration Shard 2 | 11m | 9m | -2m |
+| Integration Shard 3 | 12m | 9m | -3m |
+
+**Total CI time reduced by ~12-13 minutes per run.**
