Audit stored flow timestamp coverage without exposing raw rows

Question

Do retained raw NetFlow/IPFIX records contain genuine flow start and end timestamps that can support duration-aware time-series reconstruction, or only collector receive-time fallbacks?

Inputs

A local Netdata Agent running netflow-plugin.
NODE_UUID, MACHINE_GUID, and AGENT_HOST for the token-safe direct-Agent wrapper.
The Agent cache directory. The system-package default is /var/cache/netdata; the default raw-flow root is /var/cache/netdata/flows/raw.

Raw flow rows reveal network endpoints and traffic volumes. This procedure prints only field names and aggregate counts.

Steps

Confirm the flow Function is available through the token-safe wrapper:

bash

source docs/netdata-ai/skills/query-netdata-agents/scripts/_lib.sh
agents_load_env

agents_call_function \
    --via agent \
    --node "$NODE_UUID" \
    --host "$AGENT_HOST" \
    --machine-guid "$MACHINE_GUID" \
    --function flows:netflow \
    --body '{"info":true}' \
  | jq '{status, type, has_history}'

The public Function intentionally does not expose FLOW_START_USEC, FLOW_END_USEC, or OBSERVATION_TIME_MILLIS as grouping or selection fields. Inspect their stored presence locally instead of printing Function rows.

Select one raw-journal instance directory and confirm its size and file count:

bash

CACHE_DIR="${NETDATA_CACHE_DIR:-/var/cache/netdata}"
RAW_ROOT="$CACHE_DIR/flows/raw"
RAW_DIR="$(find "$RAW_ROOT" -mindepth 1 -maxdepth 1 -type d -print -quit)"

test -n "$RAW_DIR"
du -sh "$RAW_DIR"
find "$RAW_DIR" -maxdepth 1 -type f -name '*.journal' | wc -l

If RAW_ROOT contains multiple instance directories, run the remaining steps once per directory. Do not merge them silently.

Check the exact retained field catalog and protocol values without reading row contents:
bash
```
journalctl --directory="$RAW_DIR" --no-pager --field=FLOW_VERSION

journalctl --directory="$RAW_DIR" --no-pager -N \
  | rg 'FLOW_START_USEC|FLOW_END_USEC|OBSERVATION_TIME_MILLIS|_SOURCE_REALTIME_TIMESTAMP' \
  | sort
```
Field-catalog absence means the field does not occur anywhere in the retained journal set. Presence proves only that at least one row has it; use the next step to estimate coverage.

Take a stratified sample from every retained journal file and emit only aggregate timestamp checks:

bash

find "$RAW_DIR" -maxdepth 1 -type f -name '*.journal' -print0 \
  | while IFS= read -r -d '' file; do
      journalctl --file="$file" --no-pager -n 500 -o json 2>/dev/null \
        | jq -s '
            def n($x): (($x // "0") | tonumber);
            reduce .[] as $r
              ({records:0, start:0, end:0, observation:0,
                end_eq_source:0, positive_duration:0};
               .records += 1
               | .start += (if n($r.FLOW_START_USEC) > 0 then 1 else 0 end)
               | .end += (if n($r.FLOW_END_USEC) > 0 then 1 else 0 end)
               | .observation +=
                   (if n($r.OBSERVATION_TIME_MILLIS) > 0 then 1 else 0 end)
               | .end_eq_source +=
                   (if n($r.FLOW_END_USEC) > 0 and
                       n($r.FLOW_END_USEC) == n($r._SOURCE_REALTIME_TIMESTAMP)
                    then 1 else 0 end)
               | .positive_duration +=
                   (if n($r.FLOW_START_USEC) > 0 and
                       n($r.FLOW_END_USEC) > n($r.FLOW_START_USEC)
                    then 1 else 0 end))'
    done \
  | jq -s '
      reduce .[] as $x
        ({files_sampled:length, records:0, start:0, end:0, observation:0,
          end_eq_source:0, positive_duration:0};
         .records += $x.records
         | .start += $x.start
         | .end += $x.end
         | .observation += $x.observation
         | .end_eq_source += $x.end_eq_source
         | .positive_duration += $x.positive_duration)'

Interpret the aggregate with Netdata's decoder semantics:
- start > 0 and positive_duration > 0: at least some records contain a usable interval. Validate clock ordering and duration distribution before using it for pro-rating.
- start == 0: no sampled record has a usable duration, regardless of whether end is populated.
- end == end_eq_source together with no start is evidence that the stored end is the collector input-time fallback. Netdata fills a missing flow end with input_realtime_usec before storage in src/crates/netflow-plugin/src/decoder.rs.
- observation > 0 is a separate timestamp and does not by itself provide a flow duration.

Output

The result is a compact aggregate such as:

json

{
  "files_sampled": 100,
  "records": 50000,
  "start": 0,
  "end": 50000,
  "observation": 0,
  "end_eq_source": 50000,
  "positive_duration": 0
}

This example means the retained records do not contain a usable flow interval. The apparent end timestamp is receive-time fallback data, not proof that the exporter supplied an end time.

Notes / gotchas

The field catalog checks the complete retained journal set; the numeric coverage command is deliberately a stratified sample to avoid scanning and materializing every sensitive row.
Equality with _SOURCE_REALTIME_TIMESTAMP must be interpreted together with the Agent's timestamp-source configuration. With the default input source, it is collector reception time.
Never save journalctl -o json output for raw flows in a durable or committed location. Stream directly into aggregate-only jq processing.
A start/end interval supports only uniform-rate estimation. It cannot recover bursts inside the interval.

Audit stored flow timestamp coverage without exposing raw rows

Audit stored flow timestamp coverage without exposing raw rows

Question

Inputs

Steps

Output

Notes / gotchas

Source guides