AKS Network Capture

Capture packet-level evidence on AKS when read-only diagnostics are inconclusive. Use this to prove where a packet is dropped — inside the pod, on the node, at an NSG, on a route, or at the Azure load balancer.

This is an escalation tool. For most networking symptoms (DNS, connectivity, ingress 502s), start with aks-troubleshooting; come here when you need a pcap or wire-level proof.

Safety model

Packet capture requires elevated node access, so these scripts are built to be safe by construction:

• No shell injection. User-supplied filters and targets are validated against strict allowlists, passed to tcpdump as a single trailing argument (never a shell string), and compile-checked in-pod with tcpdump -d. There is no eval. A negative regression test (evals/tests/aks-network-capture/injection.test.sh) proves malicious inputs are rejected.
• Scoped access, not privileged. Capture pods use only NET_ADMIN + NET_RAW with hostNetwork. They mount only /var/log/aks-network-captures from the node, never the node root, and never enable hostPID.
• Pinned images. Capture Jobs use Microsoft Retina's network-tool image from Microsoft Container Registry, pinned by digest; no Docker Hub, no :latest.

The live-cluster smoke test is manual and is not run in CI. Before relying on distributed capture in production, run evals/tests/aks-network-capture/smoke-live-cluster.sh against an AKS cluster you control; it uses an isolated namespace, generates bounded same-node DNS traffic, retrieves the exact run, decodes packet records, and verifies Kubernetes plus host-artifact cleanup.

Capture workflow

# Install the fixed capture runner in the namespace
./scripts/setup-capture-configmap.sh default

# Capture DNS traffic across all Linux nodes for 2 minutes
./scripts/create-capture.sh --name dns-debug --tcpdump-filter "udp port 53" --duration 120s

# Capture for specific pods (resolves each pod to its host node and narrows to the pod IPs)
./scripts/setup-capture-configmap.sh production
./scripts/create-capture.sh --name frontend --pod-selector "app=frontend" --namespace production

# Retrieve and clean up
./scripts/retrieve-captures.sh --name dns-debug

create-capture.sh targeting flags: --node-selector, --node-names, --pod-selector, --pod-names (with --namespace). The same namespace contains the ConfigMap and capture Jobs; pass it to retrieve-captures.sh --namespace as well. Creation prints the run ID; pass --run-id during retrieval when a capture name has multiple runs. Capture flags: --duration (max 30m), --packet-size, --tcpdump-filter. The filter must be a plain BPF expression — no flags, no shell metacharacters.

Azure-side analysis (AKS)

A dropped packet often dies in the Azure network layer, not in Kubernetes. After (or instead of) a capture, run scripts/collect-azure-network-info.sh and check, in order:

• NSG — rules on the node subnet and NICs. Effective rules: az network nic list-effective-nsg --ids <node-nic-id>.
• Routes / UDR — user-defined routes that redirect or blackhole traffic. Effective: az network nic show-effective-route-table --ids <node-nic-id> -o table.
• Azure Firewall — if egress is forced through a firewall via UDR.
• VNET peering — for cross-VNET flows.
• Load balancer — backend pools and health-probe path/port for inbound Services.
• Service / private endpoints — for reaching Azure PaaS (SQL, Storage, Key Vault).
• Private DNS — the script enumerates zones subscription-wide, because a zone linked to the cluster VNET frequently lives in a different resource group than the cluster.

Egress to Azure PaaS fails (Azure SQL, Storage, Key Vault)

When pod-to-pod (east-west) traffic works but egress to external Azure services fails, the fault is almost always in the Azure network layer, not the cluster CNI. Investigate in order, before concluding it is DNS:

1. NSG rules on the node subnet (and NIC) — outbound Deny rules blocking the destination port/service tag (Sql, Storage, AzureCloud). Most common cause; inspect first.
2. Route tables / UDR — a 0.0.0.0/0 route to a firewall/NVA can black-hole or redirect PaaS-bound traffic. Confirm the effective routes on the subnet.
3. Service / private endpoints — verify the service endpoint is enabled on the subnet, or that the private endpoint's DNS resolves to the private IP and the NSG/route path to it is open.
4. Azure Firewall / NVA — if egress is forced through the firewall, confirm an application/network rule allows the destination FQDN or service tag.
5. DNS — only after the above, confirm the FQDN resolves to the intended (public vs. privatelink) endpoint.

scripts/collect-azure-network-info.sh gathers the NSG rules, route tables, firewall config, and endpoint state to identify which layer blocks the flow.

Successful vs. broken flow

When you have the pcap and the Azure config, map the path and mark where it breaks:

graph LR
    A[Source Pod<br/>10.244.1.5] -->|1. veth| B[Node netns]
    B -->|2. iptables/CNI| C[Routing]
    C -->|3. NSG / UDR| D[Azure fabric]
    D -->|4. FORWARD allow| E[Dest Pod<br/>10.244.2.8:8443]

Highlight the first hop where the packet is absent in the capture or denied by an NSG/route rule — that is the failure domain.

Analysis checklist

1. Extract the retrieved tarballs and open the pcap (tcpdump -r <file>); confirm whether the traffic was seen at all.
2. Cross-reference with the Azure network config (collect-azure-network-info.sh output).
3. If the pcap is empty, drive traffic with generate-test-traffic.sh during a fresh capture.
4. Report the failure domain and the evidence (the exact rule, route, or missing hop).