Imported from terminalskills/skills (skills/gcp-networking-observability/SKILL.md). Install upstream with npx skills add terminalskills/skills --skill gcp-networking-observability. Copyright stays with the author (Apache-2.0).

GCP Networking Observability

Overview

Google Cloud emits multiple network telemetry streams: VPC Flow Logs (sampled IP traffic), firewall rule logs, Cloud NAT translation logs, Cloud Firewall Plus / Cloud IDS threat logs, and aggregated metrics in Cloud Monitoring. Connectivity Tests provide static path analysis for "why can't A reach B?" questions. The right source depends on the question — pick one and stop, do not chase verification across tools.

Instructions

Telemetry Streams

Enabling VPC Flow Logs

# Enable flow logs on a subnet (5% sampling, full metadata)
gcloud compute networks subnets update prod-subnet \
  --region=us-central1 \
  --enable-flow-logs \
  --logging-aggregation-interval=interval-5-sec \
  --logging-flow-sampling=0.5 \
  --logging-metadata=include-all

# Route flow logs to BigQuery via a Log Sink (recommended for analysis)
gcloud logging sinks create vpc-flow-sink \
  bigquery.googleapis.com/projects/my-project/datasets/network_logs \
  --log-filter='resource.type="gce_subnetwork" AND log_id("compute.googleapis.com/vpc_flows")' \
  --use-partitioned-tables

Top-Talker Analysis (BigQuery)

Always prefer BigQuery on _AllLogs datasets for volume-based queries. Cloud Logging's Logs Explorer is for spot-checking, not aggregation.

-- Top 20 source→destination IP pairs by bytes (last 24 hours)
SELECT
  jsonPayload.connection.src_ip AS src,
  jsonPayload.connection.dest_ip AS dst,
  jsonPayload.connection.dest_port AS dst_port,
  SUM(CAST(jsonPayload.bytes_sent AS INT64)) AS total_bytes,
  COUNT(*) AS flows
FROM `my-project.network_logs._AllLogs`
WHERE log_id = 'compute.googleapis.com/vpc_flows'
  AND timestamp >= TIMESTAMP_SUB(CURRENT_TIMESTAMP(), INTERVAL 24 HOUR)
GROUP BY src, dst, dst_port
ORDER BY total_bytes DESC
LIMIT 20;

-- Find external egress (to non-RFC1918) by VM
SELECT
  jsonPayload.src_instance.vm_name AS vm,
  jsonPayload.connection.dest_ip AS external_dst,
  SUM(CAST(jsonPayload.bytes_sent AS INT64)) / POW(1024, 3) AS gb_sent
FROM `my-project.network_logs._AllLogs`
WHERE log_id = 'compute.googleapis.com/vpc_flows'
  AND NOT NET.IP_IN_NET(jsonPayload.connection.dest_ip, '10.0.0.0/8')
  AND NOT NET.IP_IN_NET(jsonPayload.connection.dest_ip, '172.16.0.0/12')
  AND NOT NET.IP_IN_NET(jsonPayload.connection.dest_ip, '192.168.0.0/16')
  AND DATE(timestamp) = CURRENT_DATE()
GROUP BY vm, external_dst
ORDER BY gb_sent DESC
LIMIT 50;

If the query returns NULL VM names, the subnet has EXCLUDE_ALL_METADATA set. Retry using jsonPayload.connection.src_ip as the join key.

Firewall DENY Analysis

-- Top denied connection attempts (firewall rule + source IP)
SELECT
  jsonPayload.rule_details.reference AS rule,
  jsonPayload.connection.src_ip AS src,
  jsonPayload.connection.dest_port AS port,
  COUNT(*) AS hits
FROM `my-project.network_logs._AllLogs`
WHERE log_id = 'compute.googleapis.com/firewall'
  AND jsonPayload.disposition = 'DENIED'
  AND timestamp >= TIMESTAMP_SUB(CURRENT_TIMESTAMP(), INTERVAL 1 HOUR)
GROUP BY rule, src, port
ORDER BY hits DESC
LIMIT 25;

Connectivity Tests (Static Path Analysis)

# Test reachability between a VM and an external IP
gcloud network-management connectivity-tests create vm-to-api \
  --source-instance=projects/my-project/zones/us-central1-a/instances/web-1 \
  --destination-ip-address=203.0.113.10 \
  --destination-port=443 \
  --protocol=TCP

# Run and read the result
gcloud network-management connectivity-tests run vm-to-api
gcloud network-management connectivity-tests describe vm-to-api \
  --format='value(reachabilityDetails.result,reachabilityDetails.traces)'

Connectivity Tests answer reachability questions deterministically — no log digging. If the result is REACHABLE, the path works; UNREACHABLE includes the failing hop and rule.

Cloud NAT Port Exhaustion

-- Find which VMs are saturating NAT ports
SELECT
  jsonPayload.endpoint.vm_name AS vm,
  COUNT(*) AS allocations,
  COUNTIF(jsonPayload.allocation_status = 'DROPPED') AS dropped
FROM `my-project.network_logs._AllLogs`
WHERE log_id = 'compute.googleapis.com/nat_flows'
  AND timestamp >= TIMESTAMP_SUB(CURRENT_TIMESTAMP(), INTERVAL 30 MINUTE)
GROUP BY vm
ORDER BY dropped DESC
LIMIT 10;

If dropped > 0, increase min-ports-per-vm on the Cloud NAT config or add more NAT IPs.

Networking Metrics (Cloud Monitoring)

# Pull RTT/latency time series via Monitoring API
curl -H "Authorization: Bearer $(gcloud auth print-access-token)" \
  "https://monitoring.googleapis.com/v3/projects/my-project/timeSeries?\
filter=metric.type%3D%22networking.googleapis.com/vm_flow/rtt%22&\
interval.startTime=2026-04-15T00:00:00Z&\
interval.endTime=2026-04-15T01:00:00Z"

For dashboards, link the Flow Analyzer — it's the fastest way to view top-N traffic without writing SQL.

Investigation Workflow

1. Identify the right primary source for the question:
   • "Why is connection failing?" → Connectivity Tests + firewall logs
   • "Who is sending the most data?" → VPC Flow Logs (BigQuery)
   • "What changed in latency?" → Networking metrics (Monitoring)
2. Run one query against the primary source.
3. Present the answer — even if the answer is 0, null, or "no traffic." Stop there.
4. Never run a second source to "verify" — different tools have different sampling and aggregation, comparing them creates rabbit holes.

Examples

Example 1 — App can't reach a third-party API

User reports their service in us-central1 can't connect to a payments API. Run a Connectivity Test from the Cloud Run service to the API IP on port 443 — the result names the failing hop (e.g., firewall-rule deny-egress-public). Update the rule, rerun the test, confirm REACHABLE. No need to dig through firewall logs.

Example 2 — Surprise egress bill

User sees a $4k spike in egress charges for the month. Query VPC Flow Logs in BigQuery grouped by source VM and external destination IP, ordered by bytes. Top result: a VM exfiltrating 800 GB to a single IP. Cross-check the destination IP against threat intel, kill the VM, file an incident. One query — done.

Guidelines

• Pick one source per question — never query a second tool to "double-check" the first
• Always use BigQuery on _AllLogs for volume/aggregation queries; Logs Explorer is for spot-checks
• Print the SQL before running it — the user should see what you're querying
• Time-range upfront — calculate "last 1 hour" / "yesterday" once, don't re-derive
• Treat 0 / null / no traffic as a conclusive answer — don't search for "more interesting" data
• For routine reachability questions, Connectivity Tests beat log mining every time
• Enable VPC Flow Logs at 0.5–1.0 sampling for active troubleshooting; 0.05–0.1 for baseline
• Route to BigQuery via Log Sink — Cloud Logging retention costs add up at high volume
• Watch EXCLUDE_ALL_METADATA subnets — VM names will be NULL, fall back to IPs
• For NAT port exhaustion, increase min-ports-per-vm or scale NAT IPs before app changes