When Tim reviewed the HTTPS support detection survey in 2013, he identified several issues with the instrumentation that made us doubt the quality of the results:
Since Erik would like us to re-run this test, I have revised the instrumentation code to address these issues.
- The requests are now for //performance.wikimedia.org, which is not likely to have an open connection.
- The order of the two tests is randomized.
- Both tests are set to time out after 5s, but we never log before 6 seconds have elapsed. This ensures that if the user navigates away from the page before both tests have had 5 full seconds to succeed, we don't log anything at all. This should control for systemic error resulting from different performance characteristics of HTTP and HTTPS.
The original instrumentation code is available here:
The diff containing my proposed changes is available here:
At Erik's request, I am keeping this private. I am obviously aware that the code from the previous round of data collection is publicly available, and that the changes will be public too if and when we start serving them to clients.