Bound ffmpeg video encoder threads

.github/workflows/test.yml

@@ -55,22 +55,10 @@ jobs:
         with:
           python-version: "3.12"
       - name: Install ffmpeg
-        run: sudo apt-get update && sudo apt-get install --yes ffmpeg strace
+        run: sudo apt-get update && sudo apt-get install --yes ffmpeg
       - name: Sync dev environment
         run: uv sync --frozen --extra dev
       - name: Run full-clip video test
-        run: |
-          set +e
-          timeout --kill-after=10s 90s \
-            strace -ff -tt -s 256 -o /tmp/video-e2e.strace \
-            uv run pytest -vv -o faulthandler_timeout=45 \
-            tests/test_video.py::TestVideoVisibleFullClip
-          status=$?
-          if [ "$status" -ne 0 ]; then
-            for trace in /tmp/video-e2e.strace*; do
-              echo "::group::$trace"
-              tail -n 100 "$trace"
-              echo "::endgroup::"
-            done
-          fi
-          exit "$status"
+        run: >-
+          uv run pytest -vv -o faulthandler_timeout=60
+          tests/test_video.py::TestVideoVisibleFullClip

docs/module-internals.md

@@ -149,11 +149,12 @@ regeneration. It centralizes container codecs, optional audio stream copying,
 metadata/chapter policy, encode-failure reporting, and atomic same-directory
 publication. Each mapped stream is allowed to reach its own end, so a copied
 audio tail is not shortened to the frame-input duration.
-Both the raw-BGR and timestamped-NUT stdin inputs disable ffmpeg probing before
-`pipe:0`: their format is already explicit, and ffmpeg 6 on Linux can otherwise
-wait for its normal analysis window while the producer blocks on a full pipe
-before the source-audio input opens. Command-order regressions cover both stdin
-modes; the real Linux full-clip CI job guards process completion.
+Both the raw-BGR and timestamped-NUT stdin modes cap the video encoder at two
+threads. A Linux full-clip trace showed ffmpeg creating dozens of worker stacks
+and severely delaying frame-pipe ingestion on a constrained hosted runner. The
+bounded codec pool avoids that scheduling collapse while leaving audio stream
+copy independent. Command regressions cover both supported video codecs; the
+real Linux full-clip CI job guards process completion.
 `probe_video_encode_profile` reads the first source video stream with ffprobe
 and preserves the supported properties that survive the 8-bit BGR boundary:
 `yuv420p`/`yuv422p`/`yuv444p` chroma sampling, recognized color tags, encoder

src/remove_ai_watermarks/video_encoding.py

@@ -21,6 +21,7 @@ if TYPE_CHECKING:
 log = logging.getLogger(__name__)
 
 _PIXEL_FORMATS = frozenset({"yuv420p", "yuv422p", "yuv444p"})
+_VIDEO_ENCODER_THREADS = 2
 _PIXEL_FORMAT_ALIASES = {
     "yuvj420p": "yuv420p",
     "yuvj422p": "yuv422p",
@@ -286,12 +287,8 @@ def raw_video_command(
     ffmpeg = shutil.which("ffmpeg")
     if ffmpeg is None:
         raise RuntimeError("Video processing requires ffmpeg on PATH")
-    # The pipe format and stream geometry are already explicit. FFmpeg 6 can
-    # otherwise wait for its normal analysis window while the producer blocks
-    # on a full pipe, before the second (audio) input has been opened.
-    pipe_input = ["-analyzeduration", "0", "-probesize", "32", "-i", "pipe:0"]
     frame_input = (
-        ["-f", "nut", *pipe_input]
+        ["-f", "nut", "-i", "pipe:0"]
         if timestamped_input
         else [
             "-f",
@@ -302,7 +299,8 @@ def raw_video_command(
             f"{width}x{height}",
             "-r",
             f"{fps:.12g}",
-            *pipe_input,
+            "-i",
+            "pipe:0",
         ]
     )
     command = [
@@ -319,6 +317,8 @@ def raw_video_command(
         "-map",
         "1:a?",
         *_video_codec_args(output.suffix.lower(), crf=crf, profile=profile),
+        "-threads:v",
+        str(_VIDEO_ENCODER_THREADS),
         *_profile_args(profile),
         "-c:a",
         "copy",

tests/test_video.py

@@ -1950,17 +1950,19 @@ class TestVideoVisibleScan:
 
 
 class TestVideoVisibleEncoding:
-    def test_raw_pipe_input_disables_redundant_ffmpeg_probing(
+    @pytest.mark.parametrize("suffix", [".mp4", ".webm"])
+    def test_encoder_bounds_codec_threads(
         self,
         tmp_path: Path,
         monkeypatch: pytest.MonkeyPatch,
+        suffix: str,
     ):
         from remove_ai_watermarks import video_encoding
 
         monkeypatch.setattr(video_encoding.shutil, "which", lambda _name: "/usr/bin/ffmpeg")
         command = video_encoding.raw_video_command(
             tmp_path / "source.mp4",
-            tmp_path / "clean.mp4",
+            tmp_path / f"clean{suffix}",
             width=12,
             height=8,
             fps=24.0,
@@ -1969,14 +1971,7 @@ class TestVideoVisibleEncoding:
             profile=video_encoding.VideoEncodeProfile(),
         )
 
-        pipe_position = command.index("pipe:0")
-        assert command[pipe_position - 5 : pipe_position] == [
-            "-analyzeduration",
-            "0",
-            "-probesize",
-            "32",
-            "-i",
-        ]
+        assert command[command.index("-threads:v") + 1] == "2"
 
     @staticmethod
     def _patch_single_frame_encode(

tests/test_video_invisible.py

@@ -102,13 +102,9 @@ def test_timestamped_encoder_reads_nut_and_passes_pts_through(
     )
 
     assert "-copyts" in command
-    assert command[command.index("-f") : command.index("-f") + 8] == [
+    assert command[command.index("-f") : command.index("-f") + 4] == [
         "-f",
         "nut",
-        "-analyzeduration",
-        "0",
-        "-probesize",
-        "32",
         "-i",
         "pipe:0",
     ]