Skip to content

test: test_frontend_screenshot_srgb i010_orange_is_srgb intermittently samples a black program frame #3960

Open

Description

@zackees

Refs #1390

The unchanged fixture intermittently samples a black program frame after producing the correct screenshot. This blocks migrating its registration in #3957; that PR restores its original tests.d registration and defers only this graph entry. The C++ fixture and public headers remain unchanged.

Observed failure on combined head ad963ae, Tier 2 image obs-rust-linux-validate:ed8bbd0fdf71e91a, GCC/g++, four CPUs, Xvfb 1920x1080x24 -noreset, C-reference core, CMake/CTest consumer:

i010_orange_is_srgb: test_frontend_screenshot_srgb.cpp:344
I010 orange: screenshot (255,96,0,255), program (0,0,0)
6 of 7 cases pass

The Cargo consumer passed all seven cases in that same combined-head replay. Batch 24 previously passed all seven through both consumers against both cores with GCC and Clang.

Focused follow-up used the same unchanged fixture, graph records and consumer setup, retaining only this fixture in a temporary diagnostic harness and preserving the resulting binaries. Results under Xvfb with no rebuilding between trials:

Core CMake/CTest Cargo/cc::Build
C-reference 10/10 pass 10/10 pass
Hybrid 5/5 pass 5/5 pass

A further bounded C-reference CTest probe passed 100/100. The original failure did not recur in these isolated trials; those passes do not erase it or prove the fixture reliable.

Setup audit: Cargo runs in runtime/, CTest in runtime/RelWithDebInfo/, as designed. Both directories have byte-identical copies of all 21 libobs effect/data files. Both consumer binaries resolve the exact same checkout-built core within each lane; GRAPHICS_MODULE_PATH points to the same checkout-built OpenGL module. They share the Xvfb display and inherited graphics environment. No missing data directory, renderer substitution, or incorrect cwd was found.

The fixture waits until a nonblack raw program frame is observed, then takes a screenshot, then reads program_pixel again (lines 271–317). The later read can observe a different frame; program_pixel also persists across cases. This is a plausible timing window, not an established root cause. Investigate synchronization with the exact program frame being asserted; preserve all screenshot/color assertions. This is distinct from the older initial cold-frame wait issue #2785 / #2787, whose wait is already present.

Evidence is retained in the assigned pool directories: land-b22-b27/runtime-c-reference-display_plugins.log (failure at line 3344; context from 3261), and fix-3957/{prepare-,repeat-,stress-,setup-,inventory-*}. Reproduce with the existing native_build::canonical_display_plugins_is_registered::runtime::canonical_display_plugins_runs_unchanged harness at the pre-deferral head, then repeat only test_frontend_screenshot_srgb through CTest under Xvfb. Logs and the temporary diagnostic harness diff are retained; no diagnostic source edits ship in #3957.

Activity

  1. added a commit that references this issue on Nov 8, 2026
  2. vpacher commented on Nov 13, 2026

    @vpacher
    Member

    Coordinator triage: the failure occurred once and then the test passed in all isolated replays since. So there is no reproduction and no acceptance criterion yet.

    Missing: a reproduction rate (for example under parallel CTest load, as in the combined-head replay) and an acceptance criterion (for example N of N passes in both cores, then the test moves into the native graph).

    Added as a sub-issue of #1390.

    /investigate

  3. moved this from Triage to Investigating in obs-rust Multiviewon Nov 14, 2026
  4. zackees commented on Nov 19, 2026

    @zackees
    MemberAuthor

    Hit again, this time in i210_orange_is_srgb rather than i010: merge_group run 44558387230, job 135686841242 (Dell runner dell-fsx3mn2-1, Tier 2 tests ON). The program sample was (0, 0, 0) while the screenshot was correct (255, 96, 0). It ejected #5249, a CI apt change that doesn't touch rendering; #5249 is re-queued unchanged.

  5. zackees commented on Nov 19, 2026

    @zackees
    MemberAuthor

    Claiming: investigation and the maintainer-assigned known-flaky mitigation in pool slot obs-studio5, branch fix/3960-known-flaky-screenshot. 1. Confirm the exact CTest registration and retry contract, then add the issue-linked registry entry. 2. Run the retry self-tests and full lint once; inspect the unchanged program-sample synchronization if cheap. 3. Publish a ready PR with Refs #3960 when backpressure permits. Expected first PR this session. The rendering root cause remains under investigation; this scoped mitigation does not complete the issue.

  6. zackees commented on Nov 19, 2026

    @zackees
    MemberAuthor

    #5277 (c24fc60) added test_frontend_screenshot_srgb to test/cmocka/known-flaky.txt, so a program-frame black sample now retries once instead of ejecting a merge group. This issue stays open for the root cause: the fixture waits for a nonblack raw frame newer than a baseline (about lines 271-290), takes the screenshot, and only then re-reads the shared program_pixel (about lines 307-317). That later read can see a different, black frame. Remove the known-flaky entry when the fixture is fixed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

needs-investigationUnderdefined: research before it can be Ready

Type

No type

Projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions