Notes by @RyanCavanaugh

strictFunctionTypes for our codebase

#49929

• Most unsoundness is around visitor-like patterns
• Performance of additional needed checkDefined calls is ~zero
• Some possible inference issues identified, Jake will reduce offline
  • Update: Fixed at Improve logic that chooses co- vs. contra-variant inferences #52123
• Please code review, will merge for 5.0

Internal RWC Tests

• What's the deal?
  • These are internal (nonpublic) codebases
• Do we need this?
  • Probably not anymore
  • These codebases are ancient and are affected by 5.0 changes
• Would greatly prefer if external contributors could see all test collateral/results
• Sufficient coverage from top100, user, etc
  • Someone please write down the differences between all these
• Let's remove

CI coverage in general

• We need to be running ALL test suites on EVERY PR so we stop getting surprised
• Please kick off these runs manually for now
• TODO: Automate that
• Performance testing
• Current tests don't give an accurate statistical picture of what's happening
• New perf tools tell you whether results are statistically significant or not, and to what degree
• Getting apples-to-apples hardware is difficult
• But we've inarguably "drifted" slower one unmeasurable step at a time
  • Option 1: Measure the unmeasurable
    • Pros: Would be good
    • Cons: Not possible
  • Option 2: Measure some proxy measure of performance instead (allocations, comparisons, clock cycles, etc)
    • Pros: Discrete, no error bars
    • Cons: Proxy measures might only roughly correlate
  • Option 3: Measure release-to-release to keep tabs on how we're doing
    • Pros: Works
    • Cons: Misses things when they happen
• Consensus: Keep investigating perf and perf measures