Skip to main content

Data centers — power & cooling

The problem. A data center's two biggest risks are heat and the power bill. A cooling unit that quietly fails lets a row of racks climb toward thermal shutdown; meanwhile cooling and power overhead — everything not spent on actual compute — is the largest controllable cost. Facilities teams need to catch a thermal problem before racks throttle, and to drive overhead down by seeing exactly where the watts go.

What we solve here is preventing heat-related outages and cutting the energy that never reaches a server.

Set up demo data​

New workspace? Run this once (Python) to create a rack's inlet-temperature series, create the subscription before we listen (with a hot spell), and the rack→cooling graph — so step 3 finds the shared unit. Safe to re-run.

import datahub_sdk, numpy as np, pandas as pd

client = datahub_sdk.DataHubClient.from_env()

client.timeseries.create([datahub_sdk.TimeSeries(external_id="rack_r14_inlet_temp_c", name="Rack R14 inlet temp", unit="deg_c", value_type="float")])
client.subscriptions.create([datahub_sdk.Subscription(
    external_id="hall_2_thermal", name="Hall 2 thermal", timeseries=["rack_r14_inlet_temp_c"])])
idx = pd.date_range(end=pd.Timestamp.now(tz="UTC"), periods=60, freq="1min")
t = np.full(60, 24.0); t[-10:] = 29.4                # above the 27 °C band
client.timeseries.insert_from_lists(timestamps=idx, values=t, ts="rack_r14_inlet_temp_c")

client.resources.create(
    [datahub_sdk.Resource(external_id=x, name=x, labels=[lbl]) for x, lbl in
     [("rack_r14", "Rack"), ("rack_r15", "Rack"), ("crac_unit_2", "Crac")]],
    [datahub_sdk.RelForm.by_external_ids("rack_r14", "crac_unit_2", "cooled_by"),
     datahub_sdk.RelForm.by_external_ids("rack_r15", "crac_unit_2", "cooled_by")])

1. Catch a thermal problem early​

Rack inlet temperatures and cooling-unit status stream live. When inlet temperature climbs past the safe band, raise a thermal_warning so the team intervenes before hardware protects itself by shutting down. See Consume live data.

import ai.intellistream.datahub.sdk.subscriptions.SubscriptionMessage;

try (var stream = client.subscriptions().listen(List.of("hall_2_thermal"))
        .stream((SubscriptionMessage msg) -> {   // auto-acks after each message
            if (inletAbove(msg.payload(), 27.0)) {        // °C — above ASHRAE band
                EventModel warning = new EventModel();
                warning.setExternalId("thermal_warning_r14_" + System.currentTimeMillis());
                warning.setType("thermal_warning");
                warning.setStatus("open");
                warning.setMetadata(Map.of("rack", "rack_r14", "inlet_c", "29.4"));
                warning.setEventTime(ZonedDateTime.now());
                client.events().create(List.of(warning));
            }
        })) {
    awaitShutdown();   // your app lifecycle; closing the stream ends delivery
}

2. Drive down overhead (PUE)​

PUE is total facility power over IT power. Roll both meters up to hourly sums, divide, and the trend shows whether a cooling tweak actually helped. See Query & aggregate.

3. When racks overheat together, find the cooling unit​

Several racks climbing at once usually means one shared cooling unit, not a coincidence. Model racks → CRAC units as a graph and walk from each hot rack to the unit they share — the exact two-alarms-one-cooling-system pattern.

ResourceNetwork a = client.resources().fetchRelated("rack_r14", 3);
ResourceNetwork b = client.resources().fetchRelated("rack_r15", 3);

Set<String> aNodes = a.nodes().stream().map(Resource::getExternalId).collect(Collectors.toSet());
Set<String> shared = b.nodes().stream().map(Resource::getExternalId)
    .filter(aNodes::contains).collect(Collectors.toSet());
// shared contains "crac_unit_2" → one failing unit, not two hot racks

See the result​

The hot rack trips the loop, and the two warming racks share one cooling unit:

thermal_warning raised for rack_r14
shared unit: crac_unit_2        ← one failing CRAC, not two hot racks

See also​