How would you design a system to ingest, deduplicate, and process a daily web scrape of 10 billion documents for LLM training at xAI?