Static implementation guide
Controlled Cache Experiments
Design cache-on and cache-off trials whose request counts, locations, inputs, and observation windows can be compared fairly.
Last updated 2026-08-27
Control the variables
Use a dedicated domain and projects without unrelated traffic. Both projects should belong to the same team, and both trials should use one client location, one guide URL, the same sequential request count, and equivalent delays. Avoid cookies, authorization, changing user agents, and query strings that could split the cache key.
Begin each trial from a documented state by redeploying the gateway or using a supported purge mechanism. Record the exact cache setting and deployment. For the enabled trial, separate the warming request from the measured sequence so the interpretation of warm hits is explicit.
Interpret differences conservatively
Compare how many client requests were served, how many reached the origin, and how the cache states changed. Edge request count and transferred bytes are different units and should never be presented as if one directly explains the other. Compression, headers, regional routing, and reporting attribution can prevent exact byte equality.
The experiment demonstrates lifecycle changes, not a permanent price quote. Plan allowances, product names, regional prices, and default caching behavior can change. Capture the dashboard’s current figures on the test date and leave monetary calculations outside the source code.