Name resolution is not content retrieval
A URL includes a scheme, a hostname, and a path. DNS answers questions about names; it is not the mechanism that transfers the page body. A client can use resolution results to select an endpoint, then ask for a resource using HTTP. These operations cooperate, but their responsibilities and failure modes differ.
For a typical HTTPS exchange over TCP, it is useful to distinguish resolution, TCP establishment, TLS processing, and HTTP requests and responses. Caches and reusable connections mean a new request need not repeat every stage. HTTP/3 uses QUIC, so the TCP sequence is an example of a deployment, not an invariant of every web request.
Transport success is not business success
TCP has mechanisms for delivering an ordered byte stream. It cannot decide whether those bytes form valid JSON, whether a user is authorized, or whether an order committed. Inspect the HTTP response and the application contract instead of treating a connection result as the result of the complete business operation.
A client may connect successfully and then fail certificate validation during TLS. Another client may complete TLS and receive HTTP 403. The latter observation establishes that a component returned a refusal, but does not by itself identify the exact authorization rule or configuration responsible. Different stages require different evidence.
Divide delay into observable stages
If an initial request is slow and later requests are faster, resolution, setup, TLS, or cache behavior may be relevant hypotheses. Time until the response begins also differs from time until its entire body arrives. A large response can take substantial time even when connection establishment is quick.
Keep the target URL and execution conditions explicit when comparing measurements. Record connection reuse, cache state, network conditions, and available server observations. One sample cannot establish a stable performance property. These patterns guide which evidence to collect; they are not measured results proving a particular cause in an unspecified system.
A timeout can leave the outcome unknown
When a client times out waiting, the server may already have performed the operation. An order could commit and its response could be lost. Blindly retrying can then create duplicate effects. Check the method semantics and the application design before deciding that a repeat request is safe.
Successful DNS does not guarantee service health, and HTTP 200 alone does not guarantee the user intended the resulting state. Response contracts, operation identifiers, idempotency handling, and a way to query an operation result connect network behavior to reliable application design. Understand what each layer establishes and what it leaves uncertain.
Subtract cumulative timestamps to compare phases
curl timing fields such as time_namelookup, time_connect, time_appconnect, time_starttransfer, and time_total are cumulative timestamps. Consider a hypothetical fresh TCP/TLS transfer with no redirect or proxy: 0.020, 0.060, 0.100, 0.250, and 0.400 seconds. Subtraction gives 0.040 seconds from resolution completion to connection, another 0.040 to TLS completion, and 0.150 from first byte to completion. Adding all five values would count overlapping time repeatedly.
The 0.150-second difference between TLS completion and first byte is not a database execution measurement. Request transmission, round trips, queuing, and application work can contribute. Record reuse, redirects, and protocol conditions, then correlate server tracing for the same request identifier before assigning a phase to an internal component.