Skip to content

Server Sizing

Resource requirements for running the cast2md server and transcription workers.

Quick Reference

Deployment RAM Disk CPU Notes
Server only (external transcripts) 2 GB 10 GB 2 cores No local transcription
Server + local Whisper (base) 4 GB 20 GB 4 cores Good for small libraries
Server + local Whisper (large-v3-turbo) 4 GB 30 GB 4 cores Recommended
Server + RunPod workers 4 GB 20 GB 2 cores Offload transcription to GPU

Memory Usage

Server Process

Component Memory
FastAPI + workers ~200 MB
PostgreSQL (Docker) ~150 MB
Embedding model (if enabled) ~500 MB

Transcription Memory

With chunked processing (default):

  • Episodes are split into 30-minute chunks
  • Peak memory: ~2-3 GB regardless of episode length
  • Works with faster-whisper, mlx-whisper, and Parakeet
  • Configured via whisper_chunk_threshold_minutes and whisper_chunk_size_minutes
  • Enables 8GB M1 Macs to run large-v3-turbo on 3+ hour episodes

Without chunking (episodes < 30 min):

  • Memory scales with audio duration
  • ~1 GB for 30-minute episode
  • ~2 GB for 1-hour episode

Whisper Model Memory

Model VRAM/RAM Quality Speed
base ~500 MB Basic Fast
small ~1 GB Good Medium
medium ~2 GB Better Slower
large-v3-turbo ~3 GB Best Medium
large-v3 ~4 GB Best Slow

Disk Usage

Storage Breakdown

Component Typical Size Notes
Database 1-5 GB Grows with transcript count
Transcripts ~50 KB/episode Markdown with timestamps
Audio (if retained) 50-200 MB/episode Usually deleted after transcription
Whisper models 0.5-3 GB Downloaded on first use
Embedding model ~500 MB If semantic search enabled

Temp File Management

Temporary files are created during transcription:

  • preprocess_*.wav -- converted audio (mono 16kHz)
  • chunk_*.wav -- audio chunks for long episodes
  • .downloading_* -- incomplete downloads

Automatic cleanup:

  • Server cleans files >24 hours old on startup
  • Node workers also clean on startup
  • Files cleaned after each successful transcription

Manual cleanup:

# Check temp directory size
du -sh /opt/cast2md/data/temp

# Remove orphaned temp files
find /opt/cast2md/data/temp -name "preprocess_*.wav" -mmin +60 -delete
find /opt/cast2md/data/temp -name "chunk_*.wav" -mmin +60 -delete

Container Configuration

arch: amd64
cores: 4
memory: 4096
swap: 2048
rootfs: local-lvm:vm-xxx-disk-0,size=26G

Resource Limits

  • Memory: 4 GB minimum with local transcription
  • Swap: 2 GB recommended (handles occasional spikes)
  • Disk: 26 GB comfortable for ~500 episodes with audio deleted

Scaling Considerations

When to Add Remote Workers

Consider RunPod or distributed nodes when:

  • Queue consistently has >10 pending jobs
  • Local transcription can't keep up with new episodes
  • Processing a backlog of an existing podcast

RunPod GPU Sizing

GPU Cost/hr Speed Use Case
RTX A4000 ~$0.15 50x realtime Budget option
RTX A5000 ~$0.20 87x realtime Recommended
RTX A6000 ~$0.35 100x realtime Large backlogs

Warning

RTX 40-series GPUs have CUDA compatibility issues with Parakeet. See Performance for details.

Monitoring

# Memory usage
free -h

# Disk usage
df -h /opt/cast2md
du -sh /opt/cast2md/data/*

# Health check
curl http://localhost:8000/api/health

Troubleshooting

Out of Memory

Symptoms: Process killed, incomplete transcriptions

Solutions:

  1. Reduce whisper_chunk_size_minutes (default: 30)
  2. Use a smaller Whisper model
  3. Add swap space
  4. Offload to RunPod workers

Disk Full

Symptoms: Failed downloads, database errors

Solutions:

  1. Check for orphaned temp files
  2. Delete processed audio files
  3. Clean old trash entries
  4. Increase disk allocation

Slow Transcription

Symptoms: Jobs taking hours instead of minutes

Solutions:

  1. Enable GPU acceleration (CUDA)
  2. Use large-v3-turbo instead of large-v3
  3. Add distributed worker nodes
  4. Use RunPod for batch processing