fixing-flaky-e2e-tests

作成者: streamlit

不安定なPlaywright e2eテストを診断して修正します。テストが断続的に失敗する場合、タイムアウトエラーが表示される場合、スナップショットの不一致がある場合、またはブラウザ固有の…

npx skills add https://github.com/streamlit/streamlit --skill fixing-flaky-e2e-tests

Fixing flaky E2E tests

Diagnose and fix flaky Playwright E2E tests in e2e_playwright/.

When to use

  • Tests fail intermittently (pass sometimes, fail others)
  • Timeout errors (TimeoutError: wait_until timed out)
  • Snapshot mismatches with pixel differences
  • Browser-specific failures (firefox, webkit, chromium)
  • User asks to fix top flaky tests from CI

Finding top flaky tests

Run the script to identify the most flaky tests from recent CI runs:

uv run scripts/fetch_flaky_tests.py

Options:

  • --days N: Look back N days (default: 4)
  • --top N: Return top N flaky tests (default: 10)
  • --min-reruns N: Minimum total reruns to include (default: 2)
  • --json: Output as JSON for programmatic use

The script downloads playwright_test_stats artifacts from successful playwright.yml runs and aggregates tests that required reruns.

Filtering tests to fix

Skip tests already marked with @pytest.mark.flaky---these are known flaky tests being tracked separately.

# Check if a test file has the flaky marker
grep -l "pytest.mark.flaky" e2e_playwright/<test_file>.py

Investigation workflow

1. Reproduce the flakiness locally (REQUIRED)

IMPORTANT: Only attempt to fix tests that fail locally. If you cannot reproduce the flakiness after 25 runs, do NOT attempt a fix—the test may be flaky due to CI environment factors that cannot be addressed locally.

Run the test up to 25 times with the affected browser(s). The loop breaks on first failure and captures full output:

for i in {1..25}; do
  result=$(make run-e2e-test e2e_playwright/test_file.py::test_name -- --browser firefox 2>&1)
  if echo "$result" | grep -q "FAILED"; then
    echo "=== FAILURE ON RUN $i ==="
    echo "$result"
    break
  fi
  echo "Run $i: PASSED"
done

If all 25 runs pass, skip this test and move to the next one.

2. Check test artifacts

After failure, examine:

  • e2e_playwright/test-results/ - traces, screenshots, videos
  • e2e_playwright/test-results/snapshot-updates/ - actual vs expected snapshots

For persistent snapshot flakiness: If a test keeps failing due to snapshot mismatches, compare the actual vs expected images in e2e_playwright/test-results/snapshot-updates/. Look for:

  • Pixel-level differences (use an image diff tool or overlay)
  • Subtle layout shifts, font rendering variations, or timing artifacts
  • Browser-specific rendering quirks (especially Firefox subpixel issues)

This helps identify whether the flakiness is due to timing (content not loaded), animation state, or browser rendering differences.

Common causes and fixes

Timing issues (most common)

Symptom: Screenshots taken before element fully renders, animations not complete.

Fix: Add explicit waits before interactions or screenshots:

# Before
element.click()
assert_snapshot(element, name="snapshot")

# After
element.click()
expect(element).to_be_visible()  # Wait for visibility
assert_snapshot(element, name="snapshot")

For popups/modals/calendars that animate:

calendar = page.locator('[data-baseweb="calendar"]').first
expect(calendar).to_be_visible()  # Wait for animation to complete
assert_snapshot(calendar, name="calendar-snapshot")

Browser retry causing extra events

Symptom: Assertion expects exact count but gets more (e.g., assert 44 == 41).

Fix: Use >= instead of == when browsers may retry failed operations:

# Before
assert error_count == expected_count

# After - browsers may retry failed image loads
assert error_count >= expected_count

Timeout too short

Symptom: TimeoutError on slower browsers.

Fix: Increase timeout for operations that can be slow:

# Before
wait_until(app, lambda: check_condition(), timeout=10000)

# After
wait_until(app, lambda: check_condition(), timeout=20000)

Snapshot mismatch due to timing

Symptom: Snapshot mismatch for ... (X pixels difference).

Causes:

  • Element still animating when screenshot taken
  • Font rendering not complete
  • Async content not loaded
  • Images not fully loaded/decoded (especially in webkit)

Fix: Ensure element is stable before screenshot:

element = page.locator(".my-element")
expect(element).to_be_visible()
# For elements with animations, wait for specific CSS state:
expect(element).to_have_css("opacity", "1")
assert_snapshot(element, name="snapshot")

For elements containing images, wait for images to be fully loaded and decoded:

from e2e_playwright.shared.app_utils import wait_for_images_loaded

element = page.locator(".my-element")
wait_for_images_loaded(element)  # Waits for load + decode
assert_snapshot(element, name="snapshot")

Browser-specific considerations

BrowserCommon Issues
FirefoxSlower console logging, may retry failed requests, subpixel rendering differences
WebkitMay have timing differences with layout
ChromiumGenerally most reliable, use as baseline

Firefox subpixel rendering flakiness

Symptom: Firefox screenshots flake with 1-pixel differences due to subpixel rendering variations.

Fix: Add a one-liner markdown element above the element being tested. This shifts the subpixel position to a more stable value:

# In the test app (.py file)
st.markdown("---")  # Stabilizes subpixel rendering for elements below
st.date_input("Pick a date")

This is a workaround for Firefox's subpixel rendering behavior and can reduce snapshot flakiness when other timing fixes don't help.

If you've exhausted timing fixes and the flakiness persists only on a specific browser due to known browser limitations (not test bugs), skip_browser may be appropriate as a last resort:

# Only use after confirming this is a browser-level limitation, not a fixable timing issue
@pytest.mark.skip_browser("webkit", reason="Webkit has known layout timing issues with this element")
def test_problematic_on_webkit(app: Page):
    ...

Important: Using skip_browser requires justification. Prefer fixing the underlying timing issue first. See "Rules" section for guidance on when skipping is acceptable.

Verification

After applying fix, verify with multiple runs:

# Run 10+ times to ensure stability
for i in {1..10}; do
  make run-e2e-test e2e_playwright/test_file.py::test_name -- --browser firefox 2>&1 | grep -E "(PASSED|FAILED)"
done

Target: 10/10 passes before considering fix complete.

Key utilities

From e2e_playwright.conftest:

  • wait_for_app_run(page) - Wait for Streamlit script execution
  • wait_for_app_loaded(page) - Wait for initial app load
  • wait_until(page, fn, timeout) - Poll until condition is true

From e2e_playwright.shared.app_utils:

  • expect_no_skeletons(element) - Wait for loading skeletons to disappear
  • reset_focus(page) - Click outside to trigger blur events
  • reset_hovering(locator) - Move mouse away from element
  • wait_for_images_loaded(element) - Wait for images to be loaded and decoded (important for webkit)

Complete workflow

  1. Fetch flaky tests: uv run scripts/fetch_flaky_tests.py --top 10

  2. Filter out marked tests: Skip tests with @pytest.mark.flaky

  3. For each remaining test:

    • Read the test code to understand what it's testing
    • Reproduce the flakiness locally (up to 25 runs)
    • Skip if not reproducible: If 25 runs all pass, move to the next test
    • Identify the root cause (timing, browser-specific, etc.)
    • Apply the minimal fix
    • Verify with 10+ runs on affected browser(s)
  4. Run checks: make check before committing

Rules

  • Reproduce locally first: Only fix tests you can reproduce locally (up to 25 runs)
  • Minimal fixes: Smallest change that fixes the issue
  • Don't disable tests without justification: Never skip tests just to "fix" flakiness. skip_browser is acceptable only when:
    1. You've exhausted all timing/wait fixes
    2. The flakiness is due to a documented browser limitation (not a test bug)
    3. You include a clear reason explaining why
  • Verify thoroughly: Run 10+ times on affected browser after fix
  • Preserve test intent: Understand what the test is validating
  • Document cause: Add comments explaining why waits/timeouts are needed

streamlitのその他のスキル

developing-with-streamlit
streamlit
**[必須]** すべてのStreamlitタスク(Streamlitアプリケーションの作成、編集、デバッグ、美化、スタイリング、テーマ設定、最適化)に使用します。また必須…
official
addressing-pr-review-comments
streamlit
現在のブランチのstreamlit/streamlitリポジトリにおけるPR上のすべての有効なレビューコメントに対応します。インラインのレビューコメントと一般的なPR(イシュー)の両方を対象とします…
official
checking-changes
streamlit
コミット前にフォーマット、リント、型チェック、ユニットテストを実行して、すべてのコード変更を検証します。バックエンド(Python)またはフロントエンド(TypeScript)の変更後にお使いください…
official
discovering-make-commands
streamlit
Streamlit開発で利用可能なmakeコマンドを一覧表示します。ビルド、テスト、リント、フォーマットタスクに使用します。
official
finalizing-pr
streamlit
ブランチの変更をマージするために、コードの簡略化、チェックの実行、変更のレビュー、必要に応じてPRの作成を行います。変更をマージする準備ができたときに使用します。
official
fixing-flaky-e2e-tests
streamlit
不安定なPlaywright e2eテストを診断して修正します。テストが断続的に失敗する、タイムアウトエラーが発生する、スナップショットの不一致がある、ブラウザ固有の挙動を示す場合に使用します。
official
fixing-streamlit-ci
streamlit
現在のブランチ/PRにおける失敗したGitHub Actions CIジョブを分析し修正します。CIチェックが失敗した場合、PRチェックに失敗が表示された場合、または診断が必要な場合に使用します…
official
generating-changelog
streamlit
2つのgitタグ間のdocs.streamlit.io向けの洗練されたウェブサイトリリースノートを生成します。新しいStreamlitリリースの準備や変更の確認の際に使用します。
official