feat: modal padding (#837)

* Add Modal backend for audio padding

- Create reflector_padding.py Modal deployment (CPU-based)
- Add PaddingWorkflow with conditional Modal/local backend
- Update deploy-all.sh to include padding deployment

---------

Co-authored-by: Igor Loskutov <igor.loskutoff@gmail.com>

server/reflector/hatchet/constants.py

@@ -37,5 +37,5 @@ LLM_RATE_LIMIT_PER_SECOND = 10
 TIMEOUT_SHORT = 60  # Quick operations: API calls, DB updates
 TIMEOUT_MEDIUM = 120  # Single LLM calls, waveform generation
 TIMEOUT_LONG = 180  # Action items (larger context LLM)
-TIMEOUT_AUDIO = 300  # Audio processing: padding, mixdown
+TIMEOUT_AUDIO = 720  # Audio processing: padding, mixdown
 TIMEOUT_HEAVY = 600  # Transcription, fan-out LLM tasks

server/reflector/hatchet/workflows/padding_workflow.py

@@ -0,0 +1,165 @@
+"""
+Hatchet child workflow: PaddingWorkflow
+Handles individual audio track padding via Modal.com backend.
+"""
+
+from datetime import timedelta
+
+import av
+from hatchet_sdk import Context
+from pydantic import BaseModel
+
+from reflector.hatchet.client import HatchetClientManager
+from reflector.hatchet.constants import TIMEOUT_AUDIO
+from reflector.hatchet.workflows.models import PadTrackResult
+from reflector.logger import logger
+from reflector.utils.audio_constants import PRESIGNED_URL_EXPIRATION_SECONDS
+from reflector.utils.audio_padding import extract_stream_start_time_from_container
+
+
+class PaddingInput(BaseModel):
+    """Input for individual track padding."""
+
+    track_index: int
+    s3_key: str
+    bucket_name: str
+    transcript_id: str
+
+
+hatchet = HatchetClientManager.get_client()
+
+padding_workflow = hatchet.workflow(
+    name="PaddingWorkflow", input_validator=PaddingInput
+)
+
+
+@padding_workflow.task(execution_timeout=timedelta(seconds=TIMEOUT_AUDIO), retries=3)
+async def pad_track(input: PaddingInput, ctx: Context) -> PadTrackResult:
+    """Pad audio track with silence based on WebM container start_time."""
+    ctx.log(f"pad_track: track {input.track_index}, s3_key={input.s3_key}")
+    logger.info(
+        "[Hatchet] pad_track",
+        track_index=input.track_index,
+        s3_key=input.s3_key,
+        transcript_id=input.transcript_id,
+    )
+
+    try:
+        # Create fresh storage instance to avoid aioboto3 fork issues
+        from reflector.settings import settings  # noqa: PLC0415
+        from reflector.storage.storage_aws import AwsStorage  # noqa: PLC0415
+
+        storage = AwsStorage(
+            aws_bucket_name=settings.TRANSCRIPT_STORAGE_AWS_BUCKET_NAME,
+            aws_region=settings.TRANSCRIPT_STORAGE_AWS_REGION,
+            aws_access_key_id=settings.TRANSCRIPT_STORAGE_AWS_ACCESS_KEY_ID,
+            aws_secret_access_key=settings.TRANSCRIPT_STORAGE_AWS_SECRET_ACCESS_KEY,
+        )
+
+        source_url = await storage.get_file_url(
+            input.s3_key,
+            operation="get_object",
+            expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
+            bucket=input.bucket_name,
+        )
+
+        # Extract start_time to determine if padding needed
+        with av.open(source_url) as in_container:
+            if in_container.duration:
+                try:
+                    duration = timedelta(seconds=in_container.duration // 1_000_000)
+                    ctx.log(
+                        f"pad_track: track {input.track_index}, duration={duration}"
+                    )
+                except (ValueError, TypeError, OverflowError) as e:
+                    ctx.log(
+                        f"pad_track: track {input.track_index}, duration error: {str(e)}"
+                    )
+
+            start_time_seconds = extract_stream_start_time_from_container(
+                in_container, input.track_index, logger=logger
+            )
+
+        if start_time_seconds <= 0:
+            logger.info(
+                f"Track {input.track_index} requires no padding",
+                track_index=input.track_index,
+            )
+            return PadTrackResult(
+                padded_key=input.s3_key,
+                bucket_name=input.bucket_name,
+                size=0,
+                track_index=input.track_index,
+            )
+
+        storage_path = f"file_pipeline_hatchet/{input.transcript_id}/tracks/padded_{input.track_index}.webm"
+
+        # Presign PUT URL for output (Modal will upload directly)
+        output_url = await storage.get_file_url(
+            storage_path,
+            operation="put_object",
+            expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
+        )
+
+        import httpx  # noqa: PLC0415
+
+        from reflector.processors.audio_padding_modal import (  # noqa: PLC0415
+            AudioPaddingModalProcessor,
+        )
+
+        try:
+            processor = AudioPaddingModalProcessor()
+            result = await processor.pad_track(
+                track_url=source_url,
+                output_url=output_url,
+                start_time_seconds=start_time_seconds,
+                track_index=input.track_index,
+            )
+            file_size = result.size
+
+            ctx.log(f"pad_track: Modal returned size={file_size}")
+        except httpx.HTTPStatusError as e:
+            error_detail = e.response.text if hasattr(e.response, "text") else str(e)
+            logger.error(
+                "[Hatchet] Modal padding HTTP error",
+                transcript_id=input.transcript_id,
+                track_index=input.track_index,
+                status_code=e.response.status_code if hasattr(e, "response") else None,
+                error=error_detail,
+                exc_info=True,
+            )
+            raise Exception(
+                f"Modal padding failed: HTTP {e.response.status_code}"
+            ) from e
+        except httpx.TimeoutException as e:
+            logger.error(
+                "[Hatchet] Modal padding timeout",
+                transcript_id=input.transcript_id,
+                track_index=input.track_index,
+                error=str(e),
+                exc_info=True,
+            )
+            raise Exception("Modal padding timeout") from e
+
+        logger.info(
+            "[Hatchet] pad_track complete",
+            track_index=input.track_index,
+            padded_key=storage_path,
+        )
+
+        return PadTrackResult(
+            padded_key=storage_path,
+            bucket_name=None,  # None = use default transcript storage bucket
+            size=file_size,
+            track_index=input.track_index,
+        )
+
+    except Exception as e:
+        logger.error(
+            "[Hatchet] pad_track failed",
+            transcript_id=input.transcript_id,
+            track_index=input.track_index,
+            error=str(e),
+            exc_info=True,
+        )
+        raise

server/reflector/hatchet/workflows/track_processing.py

@@ -14,9 +14,7 @@ Hatchet workers run in forked processes; fresh imports per task ensure
 storage/DB connections are not shared across forks.
 """
 
-import tempfile
 from datetime import timedelta
-from pathlib import Path
 
 import av
 from hatchet_sdk import Context
@@ -27,10 +25,7 @@ from reflector.hatchet.constants import TIMEOUT_AUDIO, TIMEOUT_HEAVY
 from reflector.hatchet.workflows.models import PadTrackResult, TranscribeTrackResult
 from reflector.logger import logger
 from reflector.utils.audio_constants import PRESIGNED_URL_EXPIRATION_SECONDS
-from reflector.utils.audio_padding import (
-    apply_audio_padding_to_file,
-    extract_stream_start_time_from_container,
-)
+from reflector.utils.audio_padding import extract_stream_start_time_from_container
 
 
 class TrackInput(BaseModel):
@@ -83,63 +78,44 @@ async def pad_track(input: TrackInput, ctx: Context) -> PadTrackResult:
         )
 
         with av.open(source_url) as in_container:
-            if in_container.duration:
-                try:
-                    duration = timedelta(seconds=in_container.duration // 1_000_000)
-                    ctx.log(
-                        f"pad_track: track {input.track_index}, duration={duration}"
-                    )
-                except Exception:
-                    ctx.log(f"pad_track: track {input.track_index}, duration=ERROR")
-
             start_time_seconds = extract_stream_start_time_from_container(
                 in_container, input.track_index, logger=logger
             )
 
-            # If no padding needed, return original S3 key
-            if start_time_seconds <= 0:
-                logger.info(
-                    f"Track {input.track_index} requires no padding",
-                    track_index=input.track_index,
-                )
-                return PadTrackResult(
-                    padded_key=input.s3_key,
-                    bucket_name=input.bucket_name,
-                    size=0,
-                    track_index=input.track_index,
-                )
+        # If no padding needed, return original S3 key
+        if start_time_seconds <= 0:
+            logger.info(
+                f"Track {input.track_index} requires no padding",
+                track_index=input.track_index,
+            )
+            return PadTrackResult(
+                padded_key=input.s3_key,
+                bucket_name=input.bucket_name,
+                size=0,
+                track_index=input.track_index,
+            )
 
-            with tempfile.NamedTemporaryFile(suffix=".webm", delete=False) as temp_file:
-                temp_path = temp_file.name
+        storage_path = f"file_pipeline_hatchet/{input.transcript_id}/tracks/padded_{input.track_index}.webm"
 
-            try:
-                apply_audio_padding_to_file(
-                    in_container,
-                    temp_path,
-                    start_time_seconds,
-                    input.track_index,
-                    logger=logger,
-                )
+        # Presign PUT URL for output (Modal uploads directly)
+        output_url = await storage.get_file_url(
+            storage_path,
+            operation="put_object",
+            expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
+        )
 
-                file_size = Path(temp_path).stat().st_size
-                storage_path = f"file_pipeline_hatchet/{input.transcript_id}/tracks/padded_{input.track_index}.webm"
+        from reflector.processors.audio_padding_modal import (  # noqa: PLC0415
+            AudioPaddingModalProcessor,
+        )
 
-                logger.info(
-                    f"About to upload padded track",
-                    key=storage_path,
-                    size=file_size,
-                )
-
-                with open(temp_path, "rb") as padded_file:
-                    await storage.put_file(storage_path, padded_file)
-
-                logger.info(
-                    f"Uploaded padded track to S3",
-                    key=storage_path,
-                    size=file_size,
-                )
-            finally:
-                Path(temp_path).unlink(missing_ok=True)
+        processor = AudioPaddingModalProcessor()
+        result = await processor.pad_track(
+            track_url=source_url,
+            output_url=output_url,
+            start_time_seconds=start_time_seconds,
+            track_index=input.track_index,
+        )
+        file_size = result.size
 
         ctx.log(f"pad_track complete: track {input.track_index} -> {storage_path}")
         logger.info(

server/reflector/processors/audio_padding_modal.py

@@ -0,0 +1,112 @@
+"""
+Modal.com backend for audio padding.
+"""
+
+import asyncio
+import os
+
+import httpx
+from pydantic import BaseModel
+
+from reflector.logger import logger
+
+
+class PaddingResponse(BaseModel):
+    size: int
+    cancelled: bool = False
+
+
+class AudioPaddingModalProcessor:
+    """Audio padding processor using Modal.com CPU backend via HTTP."""
+
+    def __init__(
+        self, padding_url: str | None = None, modal_api_key: str | None = None
+    ):
+        self.padding_url = padding_url or os.getenv("PADDING_URL")
+        if not self.padding_url:
+            raise ValueError(
+                "PADDING_URL required to use AudioPaddingModalProcessor. "
+                "Set PADDING_URL environment variable or pass padding_url parameter."
+            )
+
+        self.modal_api_key = modal_api_key or os.getenv("MODAL_API_KEY")
+
+    async def pad_track(
+        self,
+        track_url: str,
+        output_url: str,
+        start_time_seconds: float,
+        track_index: int,
+    ) -> PaddingResponse:
+        """Pad audio track with silence via Modal backend.
+
+        Args:
+            track_url: Presigned GET URL for source audio track
+            output_url: Presigned PUT URL for output WebM
+            start_time_seconds: Amount of silence to prepend
+            track_index: Track index for logging
+        """
+        if not track_url:
+            raise ValueError("track_url cannot be empty")
+        if start_time_seconds <= 0:
+            raise ValueError(
+                f"start_time_seconds must be positive, got {start_time_seconds}"
+            )
+
+        log = logger.bind(track_index=track_index, padding_seconds=start_time_seconds)
+        log.info("Sending Modal padding HTTP request")
+
+        url = f"{self.padding_url}/pad"
+
+        headers = {}
+        if self.modal_api_key:
+            headers["Authorization"] = f"Bearer {self.modal_api_key}"
+
+        try:
+            async with httpx.AsyncClient() as client:
+                response = await client.post(
+                    url,
+                    headers=headers,
+                    json={
+                        "track_url": track_url,
+                        "output_url": output_url,
+                        "start_time_seconds": start_time_seconds,
+                        "track_index": track_index,
+                    },
+                    follow_redirects=True,
+                )
+
+                if response.status_code != 200:
+                    error_body = response.text
+                    log.error(
+                        "Modal padding API error",
+                        status_code=response.status_code,
+                        error_body=error_body,
+                    )
+
+                response.raise_for_status()
+                result = response.json()
+
+            # Check if work was cancelled
+            if result.get("cancelled"):
+                log.warning("Modal padding was cancelled by disconnect detection")
+                raise asyncio.CancelledError(
+                    "Padding cancelled due to client disconnect"
+                )
+
+            log.info("Modal padding complete", size=result["size"])
+            return PaddingResponse(**result)
+        except asyncio.CancelledError:
+            log.warning(
+                "Modal padding cancelled (Hatchet timeout, disconnect detected on Modal side)"
+            )
+            raise
+        except httpx.TimeoutException as e:
+            log.error("Modal padding timeout", error=str(e), exc_info=True)
+            raise Exception(f"Modal padding timeout: {e}") from e
+        except httpx.HTTPStatusError as e:
+            log.error("Modal padding HTTP error", error=str(e), exc_info=True)
+            raise Exception(f"Modal padding HTTP error: {e}") from e
+        except Exception as e:
+            log.error("Modal padding unexpected error", error=str(e), exc_info=True)
+            raise

server/reflector/settings.py

@@ -98,6 +98,10 @@ class Settings(BaseSettings):
     # Diarization: local pyannote.audio
     DIARIZATION_PYANNOTE_AUTH_TOKEN: str | None = None
 
+    # Audio Padding (Modal.com backend)
+    PADDING_URL: str | None = None
+    PADDING_MODAL_API_KEY: str | None = None
+
     # Sentry
     SENTRY_DSN: str | None = None
 

server/reflector/utils/audio_constants.py

@@ -5,7 +5,9 @@ Used by both Hatchet workflows and Celery pipelines for consistent audio encodin
 """
 
 # Opus codec settings
+# ref B0F71CE8-FC59-4AA5-8414-DAFB836DB711
 OPUS_STANDARD_SAMPLE_RATE = 48000
+# ref B0F71CE8-FC59-4AA5-8414-DAFB836DB711
 OPUS_DEFAULT_BIT_RATE = 128000  # 128kbps for good speech quality
 
 # S3 presigned URL expiration