Wow! This is an incredible article. I do research and development for systems like this at GUN, and this article nails a lot of important pieces. Particularly t

Wow! This is an incredible article. I do research and development for systems like this at GUN, and this article nails a lot of important pieces. Particularly there ability to jump to an old message quickly. We built a prototype of a similar system that handled 100M+ messages a day for about $10, 2 minute screen cast here: https://www.youtube.com/watch?v=x_WqBuEA7s8 https://www.youtube.com/watch?v=x_WqBuEA7s8 . However, this was without FTS or Mentions tagging, so I want to explore some thoughts here: 1. The bucketing approach is what we did as well, it is quite effective. However, warning to outsiders, this only effective for append-only data (like chat apps, twitter, etc.) and not good for data that gets a lot of recurring updates. 2. The more indices you add, the more expensive it gets. If you are getting a 100M+ messages a day, and you then want to update the FTS index and mentions index (user messages' index, hashtag index, etc.) you'll be doing significantly more writes. And you'll notice that those writes are updates to an index - this is the gotcha and will increase your cost. 3. Our system by default backs up / replicates to S3, which is something they mention they want to perhaps do in the future. This has huge perks to it, including price reductions, fault tolerance, and less DevOps - which is something they (and you) should value! There backend team is amazingly small. These guys and gals seem exceptionally talented and making smart decisions. I'm looking forward to the future post on FTS!