[nickvanw]

Great question! We have a technical blog post about how PlanetScale Boost is implemented: https://planetscale.com/blog/how-planetscale-boost-serves-yo...

In short, it can be compared in consistency to an up-to-date read replica; PlanetScale Boost uses Vitess' VStream to process events as they happen and keep itself up to date. The blog has much more information if you're curious.

[valbu]

I think the SQL example should be bit more extreme, the count() group by is quite common and has just linear scaling and is plenty fast for majority of use cases. Tested with 1 process thread, 1k stars = 0.285ms; 10k = 2.85 ms; (and 0.1k = 45us that is same as overhead or just selecting 1 row without join and group by). So with 1k stars you need the system to average 3500 calls/s to saturate 1 thread or have meaningful latency impact. Sure, for bigger IO or row counts this does not scale and materialized view is indeed >100x faster.

[dang]

(We've merged the threads so that blog post is now the URL at the top)

[nickvanw]

Thank you as always for everything that you do @dang!