Notesfoundations🌱 seedling

Load Parameters

Also: Load Parameter · Workload Characterization

Load parameters are the few numbers that determine how much work a system does, such as requests per second, read/write ratio or followers per account. Which ones matter depends on the architecture, and often the one that matters is the shape of a distribution, not a total.

Why it exists

You can't reason about growth until you can say what grows. "More traffic" is too vague to design for; "300k timeline reads/s, 5k writes/s, follower counts up to 30M" is not.

How it works

  • Common parameters: requests per second, read/write ratio, concurrent users or connections, cache hit rate, dataset size, fan-out per request.
  • Peak, not average. Traffic follows the day, spikes at cron boundaries and during events. Size for the peak or for how fast you can reach it.
  • Distribution, not total. Skew dominates: one customer with 100× the data, one key everyone reads (Hot Spot), one account with millions of followers.
  • Pick the parameters that drive the expensive path. For a home timeline that was fan-out: followers per account weighted by how often each posts.

Trade-offs

Gains Costs
a concrete target to design and test against parameters drift as the product changes; revisit them
makes skew visible before it hurts collecting distributions needs real telemetry, not guesses

In practice

  • Twitter (2012): ~5k tweets/s average (12k+ at peaks) versus ~300k timeline reads/s. Tweet volume was easy; the follower distribution was the problem, which led to Fan-Out on Write and plans for a special case for huge accounts.
  • Bluesky (2025): the skew that hurt was accounts that follow hundreds of thousands of others, not accounts with many followers.

Pitfalls

  • Using averages for a skewed workload: the outlier is where the system breaks.
  • Load-testing with uniform random keys when production traffic is heavily skewed.

Recall

These cards are also in your review deck, scheduled with spaced repetition.