Fleet connectivity SLO queries (percent connected, boolean-dimension ratios, device ranking)

Question

For a fleet of devices (IoT gateways, robots, edge nodes) streaming to Netdata parents:

1. Single chart, 1 dimension: what percentage of the fleet is connected (streaming) right now / over time?
2. Single chart, 1 dimension: what percentage of devices have a boolean (0/1) collector dimension set to 1 (e.g. an "upstream reachable" check)?
3. The opposite: percentage of devices with that dimension at 0.
4. Rank/filter devices by the percentage of time the boolean dimension was 0 (find the devices that never connect).

Inputs

• TOKEN, SPACE, ROOM (see SKILL.md prerequisites)
• The boolean context and dimension name, e.g. context mycollector.upstream_connectivity, dimension Overall
• A time window (after in relative seconds)

Key insight: average of a 0/1 metric IS the ratio

All four questions reduce to one trick: the average of a 0/1 series is the fraction of 1s. No countif needed (see gotcha 3):

• Averaged across devices at each point in time → fraction of the fleet at 1 at that moment (questions 1, 2, 3).
• Averaged over time per device → fraction of time that device was at 1 (question 4).

Steps

1. Percent of fleet connected (from the parents' per-child streaming state). Netdata parents (v2.10.0-nightly 2026-06+ pulse) expose netdata.streaming.in.state — a per-child one-hot chart with 0/1 dimensions running, offline, archived, waiting, waiting replication, replicating. The fleet ratio is the average of running across all tracked children:

   bashCopy

   source docs/netdata-ai/skills/query-netdata-agents/scripts/_lib.sh
   agents_load_env
   agents_query_cloud POST /api/v3/spaces/$SPACE/rooms/$ROOM/data '{
    "scope":{"contexts":["netdata.streaming.in.state"],"dimensions":["running"]},
    "selectors":{"nodes":["*"],"contexts":["*"],"instances":["*"],"dimensions":["*"],"labels":["*"],"alerts":["*"]},
    "window":{"after":-21600,"before":0,"points":12},
    "aggregations":{"metrics":[{"group_by":["selected"],"aggregation":"avg"}],
                    "time":{"time_group":"average"}},
    "format":"json2","options":["jsonwrap","minify","unaligned"],"timeout":60000}' \
   | jq -r '.result.data[] | [(.[0]|todate), ((.[1][0]*10000|round)/100)] | @tsv'
   

   One column, values 0–1 (multiply by 100 for %). The denominator is every child the parents still track — including archived ones (see gotcha 1).

2. Percent of devices with a boolean dimension at 1. Same pattern on the collector context — average the 0/1 dimension across all instances:

   bashCopy

   agents_query_cloud POST /api/v3/spaces/$SPACE/rooms/$ROOM/data '{
    "scope":{"contexts":["CONTEXT"],"dimensions":["DIMENSION"]},
    "selectors":{"nodes":["*"],"contexts":["*"],"instances":["*"],"dimensions":["*"],"labels":["*"],"alerts":["*"]},
    "window":{"after":-21600,"before":0,"points":12},
    "aggregations":{"metrics":[{"group_by":["selected"],"aggregation":"avg"}],
                    "time":{"time_group":"average"}},
    "format":"json2","options":["jsonwrap","minify","unaligned"],"timeout":60000}' \
   | jq -r '.result.data[] | [(.[0]|todate), ((.[1][0]*10000|round)/100)] | @tsv'
   

3. The opposite ratio is 100 − (step 2), computed client-side:

   bashCopy

   ... | jq -r '.result.data[] | [(.[0]|todate), (100 - (.[1][0]*10000|round)/100)] | @tsv'
   

   Do NOT use time_group: countif "=0" through Netdata Cloud for this — see gotcha 3.

4. Rank devices by percent of time the dimension was 1 (or 0). Group by node, average over the whole window into a single point, then sort client-side:

   bashCopy

   agents_query_cloud POST /api/v3/spaces/$SPACE/rooms/$ROOM/data '{
    "scope":{"contexts":["CONTEXT"],"dimensions":["DIMENSION"]},
    "selectors":{"nodes":["*"],"contexts":["*"],"instances":["*"],"dimensions":["*"],"labels":["*"],"alerts":["*"]},
    "window":{"after":-86400,"before":0,"points":1},
    "aggregations":{"metrics":[{"group_by":["node"],"aggregation":"avg"}],
                    "time":{"time_group":"average"}},
    "format":"json2","options":["jsonwrap","minify","unaligned"],"timeout":120000}' > /tmp/rank.json
   
   # bottom 20 = devices with the LOWEST fraction of time at 1
   jq -r '.view.dimensions as $d
          | [range(0; ($d.ids|length)) | {n:$d.names[.], v:$d.sts.avg[.]}]
          | sort_by(.v) | .[:20][]
          | [((.v*10000|round)/100|tostring)+"%", .n] | @tsv' /tmp/rank.json
   

   sort_by(.v) ascending = most-disconnected first; sort_by(-.v) for the healthiest. The per-device value is in view.dimensions.sts.avg[] when points:1.

Output

• Steps 1–3: a timestamped series of fleet-wide percentages (one dimension — chartable as-is).
• Step 4: a ranked table percent<TAB>hostname of all devices.

Notes / gotchas

1. Denominator semantics. In step 1 the denominator is all children the parents track, including archived (long-gone or re-parented duplicates), until their charts obsolete. In steps 2–4 the denominator is only devices whose collector produced data in the window — fully offline devices drop out of the average entirely (they have gaps, and gap points are excluded from group-by aggregation). Pair step 1 (connectivity) with step 2 (health of the connected) for the complete picture.
2. Instances vs room nodes will not match. Parents can track more children than the room shows as nodes (archived duplicates after re-parenting), so do not expect totals.instances to equal the room node count.
3. countif through Netdata Cloud is unreliable (verified 2026-07-06). On a multi-parent space, time_group: countif with group_by: selected returned ~29–43% of the true value, and options: ["percentage"] returned >100% values. Plain time_group: average on the same data returned correct results. Until the Cloud aggregation of these is fixed, use the average-of-boolean trick above. (Direct agent queries are not affected.)
4. Tier-0 retention on busy parents is short. A parent with thousands of children may hold only minutes of per-second data (observed: ~18 minutes with ~5.6k children). Queries over longer windows silently use tier 1+ (per-minute averages) — fine for the average-of-boolean trick, another reason countif (which needs raw samples) is fragile here.
5. Boolean-ness matters. These patterns assume the dimension is strictly 0/1. If a collector emits other values, the average is no longer a ratio.

Source guides

• query-metrics.md — request body reference (scope, selectors, window, aggregations, response shape)
• SKILL.md — discovery endpoints (/spaces, /rooms, /nodes)