Load testing at scale and lessons learned
kevinlondon.com
kevinlondon.com
Also, the numbers doesn’t seem to reflect a horizontally scalable cluster deployment, the load numbers are low enough for a single server to handle.
> In the face of ambiguity, refuse the temptation to guess.
Great advice, but shouldn't it apply to the test traffic as well? Mimicking traffic patterns and making assumptions has a lot of downsides. You'll always be wrong, though the margin of error will vary. Traffic capture and replay is a great way to ensure load test traffic looks like the real thing.
The biggest blocker is that most traffic is highly sensitive to the environment. For example, UUID's and ID's likely don't exist in the development environments where you really want to test ahead of feature release--generally the datasets won't be the same. If they are the same, you'll complicate your life GDPR wise, increase your risk of breaches, and risk sharing PII.
Rather than capture traffic, you can get a similar accuracy, and more flexible setup, by looking at the traffic patterns and allocating your load based on transactions proportionally.