Fair catch, that line was carrying more weight than I backed up. The reason it's not linear is that failures are correlated, not independent. A big chunk of the web sits behind a handful of anti-bot and CDN vendors, so when one ships a change, every source behind it breaks in the same window instead of on its own schedule, and a dozen manageable sources become five fires in one week. Plus the shared-infra case you mentioned: one source moving to heavier JS or infinite scroll eats the headless pool's capacity and slows sources that never changed. So the load tracks sources times their shared dependencies, not source count alone. Updated the piece to say this properly, appreciate the push.
I've dealt with this exact issue on a project where we scraped a dozen sites and yeah splitting access from extraction so failures actually surface instead of hiding was the single biggest win we got out of a similar setup. Outsourcing makes sense too, doing it in house only holds up if you're dealing with a handful of sources that barely change. My one gripe is the bit about maintenance surface blowing up faster than expected, it's stated like a fact but there's nothing backing it up. Don't need a full cost breakdown for something this short, but even one real example, like shared rendering infra breaking or failures on different sources being correlated somehow would make that point land a lot harder