A repeatable method, Linux/Windows tools, Azure Network Watcher, on-prem path
Practise these as flashcards. One or two lines each; the ones you miss come back first.
Start freeRefused: the host answered with a RST, so the network path works but nothing is listening on that port (or a firewall actively rejects). Timed out: no answer at all, packets are being dropped by a firewall/NSG, a routing problem, or the host is down.
Bandwidth is the capacity of the link. Throughput is what you actually achieve. Latency is the time for a round trip. A high-bandwidth link can still give low throughput if latency or packet loss is high.
Linux: nc -zv host 443 (or curl -v telnet://host:443). Windows: Test-NetConnection host -Port 443. Run it from the actual source machine.
ICMP is blocked by a firewall or NSG, or the host doesn't answer ping, while TCP 443 is allowed. Ping is not a valid reachability test in many networks, especially cloud. Test the real port instead.
The routers along the path and the latency to each. Limits: many devices drop or de-prioritize ICMP/UDP probes, so you see stars or fake loss; the return path may differ; firewalls may block it. Use TCP traceroute (traceroute -T -p 443) or mtr for better data.
Clarify who, since when, all or some users, what changed. Time the request phases with curl -w (DNS, connect, TLS, TTFB, transfer). Check LB metrics per backend, host USE metrics, APM for slow transactions and DB calls, and the network path with mtr. Narrow by elimination and back every conclusion with data.
Something specific to that site: its DNS servers, proxy, WAN/VPN link, a firewall rule for its subnet, or routing. Test from a machine at that site (Test-NetConnection, tracert, nslookup) and compare to a working site.
curl -o /dev/null -s -w with time_namelookup, time_connect, time_appconnect, time_starttransfer and time_total. Big DNS = resolver; big connect = network/loss; big TLS = handshake/cert/LB CPU; big TTFB = backend; big after TTFB = transfer/bandwidth.
Test with iperf3: if parallel streams are much faster than one, a single TCP stream is limited by the window vs bandwidth-delay product, made worse by any loss. If iperf3 is fast, the bottleneck is disk or the copy tool (scp is single-threaded and encrypts). Also check NIC errors, MTU, shaping and VM network caps.
Correlate timing. Candidates: one unhealthy backend in rotation, LB or firewall idle timeouts killing long-lived connections, SNAT port exhaustion under load, a failing DNS server in the resolver list, GC pauses or connection pool exhaustion in the app. LB logs per backend and packet captures narrow it quickly.
DNS resolves correctly? Effective NSG on the NIC and subnet (IP flow verify). Next hop: does the UDR send it to the hub firewall, and does the firewall allow and log it? Gateway/ExpressRoute up with the route advertised? On-prem firewall allows the Azure subnet? Return route from on-prem back to Azure? Test with test-connectivity and a capture on both sides.
Whether the handshake completes (SYN without SYN-ACK = dropped), who sends RSTs, retransmissions (loss), zero-window (receiver overloaded), TLS alerts, and timing gaps between request and response that show where time is spent. Capture both ends to see where packets vanish.