Two weeks of hardcore performance work on Quartz, but now we can import 3GB containing 2.1 million contact list events in 9.4 minutes, *single-threaded* on a Pixel 9, using just a cool, calm, and collected 15MB (yes, MB) of RAM.
That's up to 10,000 events per second per CPU for lighter follow lists and 3,000 events per second on heavier ones.
By import, I mean receive, unzip, parse, string intern, verify, save, and index (SQLite) with individual transactions on write-ahead logging to guarantee DB consistency and match real use cases.
JSON parsing w/ string interning is now pretty much capped by IO and interning at 120,000 events per second per CPU or ~50MB/s per CPU.
JSON Parsing + interning + verifying runs at ~20,000 events per second per CPU or ~30MB/s per CPU.
None of these use SIMD instructions, which are still poorly supported on mobile CPUs.
