NOTES ON DEPENDABLE SYSTEMS
Clear explanations of the seams in connected software. Written for people who prefer a measured experiment to a confident guess.
CONNECTION LIFECYCLE
A client can reuse an established connection for several requests. That avoids repeating transport and TLS setup, so its next response may arrive much faster. The improvement does not necessarily mean the network path has changed.
Measure cold and warm behavior separately. Include a small request, a larger transfer and a restart of the test client. Keep the destination and the underlying network fixed while comparing the results.
NAMES & DESTINATIONS
A TLS service may host several names on one address. The server name in the handshake and the HTTP host identify the intended service. A request to a bare IP can fail certificate validation even when the named service works correctly.
Preserve those names when testing a different route to the same server. Do not disable certificate checks to turn an inconclusive test into an apparent success. Record which stage failed instead.
RECOVERY DESIGN
Retries can turn a small interruption into a burst of work. Use a bounded retry budget, distinguish temporary failures from rejected requests, and avoid letting background recovery occupy every worker.
Successful recovery should be observable: capture the reason, the action and the verification. Do not replace missing measurements with reassuring values. “Not measured” is a useful state when it is reported honestly.
WORKING PRACTICE
A repeatable test describes its vantage point, transport, destination and deadline. Compare like with like, and keep a control path. A small, well-labeled experiment is often more useful than an impressive volume of unstructured output.
Link Ledger is an independent collection of original engineering notes. It is a reading space, not a live monitoring console. All pages are static; this site uses no cookies, analytics, external fonts or third-party scripts.