Aruvi — Telegram Media Streaming Platform

Architecture

Browser/TV App → Cloudflare Tunnel → FastAPI (uvicorn) → PyroTGFork clients → Telegram MTProto
                                    ↕
                         Sliding Window Cache (500MB RAM global)
                                    ↕
                          NVMe Disk Cache (13GB, 3h TTL)

Key Components

Streaming Pipeline

1. All-bot warmup — all 14 bots fetch messages 1-20 at startup
2. Fast-start — first 13 chunks as 1-chunk batches across all 13 helpers
3. Batch fetch — remaining chunks in parallel (BATCH_SIZE=5 per bot)
4. Sliding window — 300MB ahead / 100MB behind stays in RAM; rest to NVMe
5. 100MB lookahead — maintains cushion against Telegram latency spikes
6. Global OOM guard — evicts farthest chunks across all streams at 500MB

Deployments

HidenCloud (current)

TELEGRAM_API_ID=
TELEGRAM_API_HASH=
TELEGRAM_BOT_TOKEN=
TELEGRAM_STORAGE_CHANNEL_ID=
TELEGRAM_HELPER_BOT_TOKENS=token2,token3,...,token14
DATABASE_URL=postgresql+asyncpg://...
TUNNEL_TOKEN=

Hugging Face Spaces (migrating)

Key Design Decisions

• BATCH_SIZE=5 — 5 × 1MB chunks = 5MB per bot batch; reduced from 10 for faster first-byte
• Global 500MB RAM — not per-stream; prevents OOM with multiple concurrent streams
• All chunks to NVMe — both ahead and behind chunks persist; enables instant rewind without Telegram refetch
• Per-bot fresh Message — each worker fetches its own get_messages() to avoid cross-bot FILE_REFERENCE_INVALID
• Sentinels for shutdown — concurrency None tuples on task queue signal workers to stop
• reconnect_client uses start() — gets new auth key on AuthKeyUnregistered, not just connect()

Tech Stack

• Backend: Python 3.11, FastAPI, SQLAlchemy async, Kurigram (Pyrogram fork)
• Database: PostgreSQL (Supabase) or SQLite
• Cache: In-memory + NVMe disk with 3h TTL
• Tunnel: Cloudflare Tunnel (cloudflared) + Cloudflare Workers proxy
• Frontend: React (pre-built, served as static files)
• Platform: HidenCloud ARM64 container / Hugging Face Spaces

License