The All File Converter service is designed as a scalable distributed SaaS system based on a modern asynchronous Python 3.12 stack. The architecture is divided into independent layers: network event reception, task orchestration, isolated execution of heavy binary processes, and an analytical pipeline.

1. General Technology Stack and System Architecture

The platform is built on the principles of high throughput, minimal memory consumption, and fault tolerance:

  • Telegram Bot Framework: Aiogram 3.13, operating in Webhook mode with secret token validation and custom message lifecycle filters.
  • Web Gateway and REST API: FastAPI based on the Uvicorn ASGI server with asynchronous streaming of binary files (FileResponse) and background cleanup via BackgroundTasks.
  • Queue Broker and Cache: Redis 7 (Celery task management, race condition locks, brute-force protection, session caching).
  • Background Executor (Task Queue): Celery 5.4 with a dedicated pool of workers in an isolated converter_worker container.
  • Database: PostgreSQL 16 with the SQLAlchemy 2.0 (asyncpg) ORM layer, a persistent connection pool (20+10 overflow), and automatic retry for failed transactions (@db_retry).
  • Network Contour: Tunneling via Cloudflare Zero Trust with direct IP access to the server blocked via Middleware.

2. Asynchronous Queues and Heavy Computation Isolation (Celery + Redis)

Media file and office suite conversion creates peak loads on CPU and RAM. To prevent incoming Telegram message processing from being blocked during heavy operations, strict isolation is implemented:

  • Task Delegation to Redis: Upon format selection, the Telegram handler registers a task in the DB with the PROCESSING status and places the job in the tasks.execute_conversion queue via Celery.
  • Isolated Worker Container: Conversion utilities are executed within a separate Linux container with its own CPU time and memory limits.
  • Freeze Control and Timeouts: External utility calls are wrapped in an asynchronous context with strict time control (conversion_timeout_sec = 180). If the limit is exceeded, the process is forcibly terminated via proc.kill(), freeing up resources.
  • Fault-Tolerant Fallback: In the event of temporary Redis broker unavailability, the task is automatically intercepted by the local asynchronous dispatcher and executed directly without disruption to the user.

3. Specialized Conversion Engine Pipeline

Highly specialized native utilities and libraries are used for each data type:

  • Documents and Spreadsheets (LibreOffice): A headless office suite (soffice --headless) for accurate rendering of DOCX, XLSX, PPTX, RTF, ODT into PDF or text files.
  • Streaming Audio and Video (FFmpeg): Multithreaded transcoding of video codecs (H.264), audio codecs (MP3, OGG Opus), audio track extraction, GIF generation (Lanczos filter), and square cropping (1:1) of Telegram video messages.
  • High-Speed PDF Processing (Poppler Utils): pdftotext utilities (instant extraction of formatted text in UTF-8) and pdftoppm (page-by-page rendering of PDFs into raster images without LibreOffice overhead).
  • Optical Character Recognition (Tesseract OCR): Neural network extraction of printed text from scans and photos across 40+ languages.
  • Raster and Vector Graphics: Pillow libraries (including HEIC and AVIF format support), CairoSVG for vector images, and lottie for animated Telegram stickers (.TGS).
  • Books, Subtitles, and Fonts: The Calibre engine (ebook-convert), the pysubs2 subtitle parser (SRT, VTT, ASS, SSA), and the fonttools font compiler (Brotli compression to WOFF2).

4. Preventive Resource Protection (System Guard)

To protect against server crashes due to insufficient memory (OOM Killer), the System Guard preventive diagnostics service has been introduced. Before accepting a file for processing, the system checks key host metrics:

  • Free RAM: Minimum 500 MB of free space (guard_min_free_ram_mb).
  • Disk Space: Minimum 2 GB of free space in the /tmp directory (guard_min_free_disk_mb).
  • Task Queue: Limitation on the Celery queue length (no more than 20 pending tasks).

If limits are exceeded, the service temporarily enables protection (HTTP 503 / chat message), preventing server overload and sending an instant alert to administrators in Telegram.

5. File Lifecycle and Security (GDPR)

The architecture is designed around the Zero-Data-Footprint model:

  • User files are uploaded to a secure tmp/conversions/ volume with unique prefixes based on task identifiers.
  • Files remain available strictly within the working session — no more than 15 minutes (900 seconds).
  • The automatic garbage collector erases source and completed files immediately after confirmation of successful chat delivery or upon session timeout expiration.
  • The PostgreSQL database does not store binary files or personal document texts — tables record only anonymized technical metadata (formats, sizes in bytes, runtime, statuses).

6. Universal REST API for External Web Services

The service was originally designed as a multi-platform backend. Alongside the bot, a fully functional secure API operates for websites (e.g., built on Django) and third-party bots:

  • GET /api/v1/formats — dynamic JSON matrix of available conversion directions, synchronised with Google Sheets settings.
  • POST /api/v1/convert — universal endpoint accepting multipart/form-data (file, target format, external client ID) and returning the resulting byte stream as a direct download.
  • Authorization via Bearer tokens (WEBHOOK_REFRESH_TOKEN) with timing attack protection via secrets.compare_digest.

7. Control Panel and Observability (NiceGUI + AG Grid)

Monitoring of business metrics and system status has been moved to a native Single-Page control panel based on NiceGUI 2.x:

  • Interactive AG Grid (v32+) tables with custom Excel-style checkbox filters (aggrid_filters.js).
  • Real-time monitoring of queues, API provider latencies, and PostgreSQL / Redis ping.
  • End-to-end date filtering with seamless integration of the Metabase BI dashboard via signed JWT tokens.