A network incident becomes difficult when observers cannot agree what failed. 'The internet is slow' could mean one client, one DNS resolver, a SaaS login, a regional transit path or a saturated access link. Convert it into a testable statement: source and destination, application, time window, expected response and actual observed difference. Preserve evidence before rebooting equipment. An isolated failure and a fleet-wide failure require different hypotheses even when the first user report sounds identical.
Troubleshooting starts with a falsifiable symptom
For example, change “the office VPN is broken” into “three employees on one access network cannot reach the internal finance application after authentication, beginning at 09:15, while employees on another network can.” That statement separates authentication from application reachability and identifies a comparison group. Confirm time synchronization across client, gateway and service logs before matching events. Record a failed request ID or five-tuple where policy allows it; do not collect credentials or personal payloads merely to prove a network symptom.
Follow the dependency chain instead of guessing a layer
A client must have a usable link, addressing, next-hop reachability and any required naming and transport services before an application exchange can succeed. Start where evidence first diverges from expectations. An ICMP ping result can be useful but may be filtered or travel differently from the application. A working DNS response does not prove TLS will succeed; a TLS handshake does not prove user authorization. Correlate these dependencies with packet captures and application logs as appropriate.
Use paired tests to separate name resolution and connectivity: query the expected DNS answer, then test a known endpoint IP and port with the same client network and policy. If DNS is healthy but a TCP handshake times out, check routes, security groups, stateful firewalls and return-path symmetry. If the handshake succeeds but an HTTPS request fails, inspect certificate name matching, authentication, reverse proxies and application logs. Avoid the simplistic conclusion that one successful ping proves the service is operating.
Packet capture is valuable when the failure has been narrowed enough to choose the right observation point. Collect only approved traffic, with timestamps and access controls appropriate to the organization. A missing TCP response may reflect routing, filtering, server behavior or capture location; it is not by itself proof that the remote host is down. Compare captures on both sides of a suspected boundary before assigning blame.
Differentiate packet loss, delay, queueing and name failures
High latency can arise from physical distance, queue buildup, retransmission, overloaded services or an inefficient route. Error counters describe device observations, while packet loss observed at the application may have multiple causes. A DNS NXDOMAIN response is not a network timeout. Do not infer that every retransmission note in a capture proves a switch dropped a packet; capture position, offloads and missed frames matter. Validate the symptom at multiple appropriate observation points before assigning blame.
When latency climbs, compare first-byte timing with transport retransmission and queue measurements. A stable RTT with a slow application may implicate server processing or downstream dependencies; an RTT that rises only at peak load can suggest queueing. Path probes may be filtered or rate-limited, so they cannot by themselves establish a faulty hop. Use a repeated application request, a capture at an approved boundary and server-side timing to determine which hypothesis survives measurement.
Topology and policy hide important boundaries
Modern networks contain VLANs, VRFs, overlays, cloud security groups, gateways and remote identity services. A packet may cross several independent administrative and failure domains. Make a hop-by-hop path map and label who owns each control. A successful local route lookup does not mean a remote firewall, NAT mapping, security policy or workload is correct. For intermittent problems, capture time-synchronized logs and identify the smallest common segment among affected clients.
Change management is part of diagnosis
Recent configuration changes are useful leads but not automatic proof of causation. Gather before-and-after state, tickets and monitoring events. If a reversible change is appropriate, alter one variable and compare the same measurement. A series of undocumented restarts can erase evidence and widen the outage. Keep stakeholders informed with a statement of what is observed, what is only suspected and the next controlled test. Good troubleshooting is partly a discipline of uncertainty.
Close the incident with reproducible evidence
A sound recovery test repeats the original failed operation from the same class of client and records whether latency, error rate and authorization behavior meet the pre-incident expectation. A green status dashboard may cover a different route or protocol, so compare application-level evidence as well as infrastructure counters. The conclusion should distinguish a demonstrated root cause from a plausible contributing factor, document the owner of any residual risk, and specify what will trigger a reopened investigation.
The final report should include the failed dependency, evidence for the root cause, corrective action, proof of service recovery, residual risks and any prevention plan. Distinguish symptoms that disappeared from causes that were demonstrated. Keep sanitized captures, diagrams and relevant command outputs so another engineer can reproduce the reasoning. Automation and AI-assisted anomaly detection can help triage events, but they do not replace a verified cause-and-effect chain or accountable change owner.