{
  "id": 20718221,
  "title": "Traffic mix matters",
  "url": "https://urgent.news/2026/11/19/traffic-mix-matters",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-19T07:33:44.000Z",
  "source": {
    "name": "Lobsters",
    "slug": "lobsters",
    "url": "https://brooker.co.za/blog/2026/11/18/workload-mix.html"
  },
  "original_language": "en",
  "account": "The recent M365 outage report highlights the significance of traffic mix in load testing, emphasizing its impact on the performance of computer systems. To illustrate this, a simple simulator was used, modeling a system with a table of locks. Requests were assigned to one of 16 worker threads, waiting in a FIFO queue for the lock corresponding to their key. After holding the lock for an exponentially distributed time (with a mean of 1 ms), the lock was released and the thread was freed. The keys were drawn from a Zipf distribution over N keys, with a coefficient of 0 indicating a uniform distribution. As expected, a skewed distribution led to less throughput due to a few highly contended keys, causing underutilization of the rest of the machine's resources. In contrast, when considering a relational database like Postgres, a fast in-memory cache is backed by a slower storage system. Keys are drawn from a Zipf distribution over N keys, with a cache size represented as a percentage of the total keys. The cache is significantly faster than storage, and a hit is served from the LRU cache, while a miss goes to storage, adding the key to the cache. The throughput curve shows that a skewed distribution results in better performance, but this is crucial for load tests and benchmarks to match real production traffic, as throughput can vary by multiple orders of magnitude, depending on factors such as key skew in the benchmark workload. Optimistic concurrency control and architectures with a single cost/latency mode, like Aurora DSQL, are designed to reduce workload sensitivity.",
  "summary": null,
  "key_points": [
    "Traffic mix significantly impacts load testing outcomes",
    "Skewed distribution reduces throughput due to contended keys",
    "Realistic traffic matching crucial for accurate performance evaluation"
  ],
  "editors_take": "Failing to replicate skewed production traffic during load testing risks massive performance discrepancies because benchmarks that assume uniform workloads ignore the resource underutilization caused by highly contended keys.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}