2026-11-07 retro

· 314 words

is it worth the time?

  • if you do it 5-50x/day it always is
  • 1x/day: yeah if it takes 1-5min
  • 1x/week: yeah if it takes 30min+
  • 1x/month: yeah if it takes 1h+

on being unable to track ppl in the NYC marathon

another person's version of to-did list, perhaps more aptly named!

cool roundup newsletter format

Each edition collects 5-10 links with one-sentence framing. The format is intentionally minimal. If you can scan it in five minutes and walk away with one link worth reading, it did its job.

give up on trying to build your own cyberdeck and just use the daylight device?

thoughts

prod validation at scale

  • shadow traffic: duplicate read-only prod -> compare responses offline -> cutover after sufficient soak
    • challenges
      • noise: 1 candidate, 2 prod duplicates (which sometimes disagree -> discard datapoint) [Twitter Diffy]
      • writes: either use only for read paths or point write paths to stubbed/isolated dependencies
  • canary
    • challenges
      • to sticky assign or not sticky assign? yes -> consistent exp, no -> more user assignments, mixed-version sessions
      • blast radius (internal, single region/AZ?), stage duration and rollout %
      • automated baseline comparison: ideally against a freshly deploy old version, to avoid drift from non-code changes (e.g. cache warmup time, heap size, LB assignment) [Netflix/Google Kayenta], min 50 timeseries data points per metrics
  • fast rollback
    • blue/green/canary weight, old version fresh

TODO: rank things

++: Things I really enjoyed doing +: Things I enjoyed doing -: Things I felt indifferent about doing --: Things I hated doing

: Things I really liked the outcome of : Things I liked the outcome of <: Things I felt indifferent about the outcome of <<: Things I did not like the outcome of