DEV Community

Perch D
Perch D

Posted on Originally published at iotforall.hashnode.dev

Building Data Center Monitoring: Asset Models, Data Quality and Event Correlation

A cooling unit reports a fault. Rack inlet temperatures rise. Several servers subsequently report thermal warnings.

The monitoring backend now holds several events from different sources. Operators need to understand whether those observations are related, which assets may be affected and what requires investigation.

Displaying everything on one dashboard improves visibility. Connecting observations through verified dependencies makes the information more useful.

A data center monitoring architecture combines equipment acquisition, normalized observations, asset relationships and incident workflows.

This guide explains the implementation decisions behind that architecture.

1. Separate Protocol Acquisition from Application Meaning

Data center equipment can expose observations through Modbus, BACnet, SNMP, OPC UA or vendor APIs, depending on the installation.

The acquisition layer handles communication. The application model defines what each observation represents.

A temperature reading might originate from a register, a BACnet object or a vendor-specific SNMP object. Normalize those sources before using them in shared dashboards and rules.

A practical processing path is:

Acquire → Validate → Normalize → Store → Evaluate → Present
Enter fullscreen mode Exit fullscreen mode

At ingestion, define:

  • Asset identity.
  • Measurement type.
  • Engineering unit.
  • Scaling and conversion.
  • Timestamp origin.
  • Quality mapping.
  • Missing-value behavior.

Avoid implementing different interpretations of the same measurement in every dashboard.

For background on acquisition, supervisory functions and deployment boundaries, see Iotellect’s SCADA architecture guide.

2. Model Dependencies Separately from Location

Two assets sharing a room do not necessarily share a power or cooling dependency.

Represent relationships explicitly:

  • Located in.
  • Powered by.
  • Served by a cooling zone.
  • Connected through a network device.
  • Associated with an operational service.

An illustrative record might look like this:

{
  "assetId": "rack-a12",
  "assetType": "rack",
  "locationId": "room-a",
  "powerFeedIds": [
    "pdu-a-04",
    "pdu-b-04"
  ],
  "coolingZoneId": "zone-a-east",
  "networkDeviceIds": [
    "switch-a12-01"
  ]
}
Enter fullscreen mode Exit fullscreen mode

This is an application example, not a standard schema.

These relationships support different questions.

Location helps technicians find equipment. A power relationship supports investigation of electrical incidents. A cooling-zone relationship helps identify which measurements to review after a cooling exception.

Maintain relationship history. Equipment moves and configuration changes can invalidate assumptions used by correlation rules.

3. Preserve Event Time, Arrival Time and Quality

Observations can arrive late or out of order.

A gateway may upload buffered records after reconnecting. An upstream interface may return a cached value.

Store acquisition time separately from application receipt time:

{
  "assetId": "rack-a12",
  "measurement": "inlet_temperature",
  "value": 27.4,
  "unit": "degC",
  "sourceTimestamp": "2026-11-11T08:30:00Z",
  "receivedTimestamp": "2026-11-11T08:30:03Z",
  "quality": "good"
}
Enter fullscreen mode Exit fullscreen mode

Define where sourceTimestamp comes from. If the instrument provides no timestamp, document which component assigns it.

Keep historical storage separate from current-state selection. An older observation may fill a historical gap without replacing a newer current value.

For platform-specific background, see Iotellect’s documentation on Data Table Timestamps and Quality.

Make Freshness Explicit

Quality and freshness answer different questions.

A measurement may have been valid when acquired but now be too old for a live rule.

Distinguish:

  • Good and sufficiently fresh.
  • Uncertain.
  • Invalid.
  • Stale.
  • Missing.

Display the last known value with its age. Do not convert a missing measurement into zero.

4. Define Individual Alarms before Grouping Them

Correlation depends on understandable source events.

For each alarm rule, document:

  • The observation being evaluated.
  • Required quality and freshness.
  • Applicable operating mode.
  • Trigger condition.
  • Persistence period, where appropriate.
  • Recovery behavior.
  • Maintenance handling.
  • Responsible team.

A simplified temperature evaluation could look like this:

def evaluate_temperature(sample, policy, now):
    if sample.quality != "good":
        return "quality_issue"

    age = now - sample.source_timestamp

    if age < 0 or age > policy.maximum_age:
        return "timestamp_or_freshness_issue"

    if sample.value > policy.temperature_limit:
        return "temperature_exception"

    return "normal"
Enter fullscreen mode Exit fullscreen mode

This is illustrative pseudocode.

Production behavior may also require hysteresis, persistence, maintenance suppression and state tracking. Thresholds and timing need engineering review for the actual installation.

Distinguish equipment-reported faults from application-generated exceptions. Their evidence and investigation paths can differ.

5. Use Explainable Correlation Rules

Consider this sequence:

  1. Cooling unit CU-03 reports a fault.
  2. Several temperature exceptions appear in its assigned zone.
  3. Associated servers report thermal warnings.

A correlation service can identify this as a candidate incident using verified relationships and a defined time window.

A simplified matching function might be:

def related_to_cooling_event(event, cooling_event, asset_model, window):
    same_zone = asset_model.shares_cooling_zone(
        event.asset_id,
        cooling_event.asset_id
    )

    time_matches = (
        cooling_event.source_timestamp
        <= event.source_timestamp
        <= cooling_event.source_timestamp + window
    )

    return same_zone and time_matches
Enter fullscreen mode Exit fullscreen mode

This identifies a possible relationship. It does not establish root cause.

A real implementation needs policies for uncertain timestamps, delayed events, duplicate delivery and relationship changes.

Show Why Events Were Grouped

An incident should explain its evidence:

Cooling unit CU-03 reported a fault. Three inlet temperature exceptions were subsequently observed in its assigned zone. Two associated servers also reported thermal warnings.

Preserve each original event’s timestamp, severity, acknowledgement and history.

Grouping should not prevent independent escalation where required.

Test coincidental faults as well as shared failures. Nearby events can occur for unrelated reasons.

6. Separate Monitoring Access from Control Permissions

Reading a measurement and changing a setpoint require different privileges.

Define permissions for:

  • Viewing data.
  • Acknowledging alarms.
  • Editing rules.
  • Requesting equipment commands.
  • Approving configuration.
  • Managing users.

For remote actions, record the requester, target, requested change, relevant preconditions and result.

Distinguish a submitted request from a device-confirmed outcome. Where supported, use feedback to identify confirmed, failed and timed-out operations.

Keep supervisory workflows within their defined scope. Equipment protection and life-safety functions require their own engineering and validated implementation.

7. Design Site Behavior during Uplink Outages

Central visibility can disappear while local equipment continues operating.

Decide which functions must remain available at the site:

  • Acquisition.
  • Local dashboards.
  • Approved alarm evaluation.
  • Historical recording.
  • Operator access.

Inspect their dependencies. A local interface may still depend on a remote identity service or externally hosted resources.

Where buffering is supported, define capacity, acknowledgements, retries and overflow behavior.

After reconnection:

  • Preserve original observation times.
  • Recognize duplicate records.
  • Prevent older observations from replacing newer current values.
  • Apply explicit replay policies to alarms.
  • Control backlog traffic.

A lost site connection should appear as a visibility issue without implying that every monitored asset has failed.

8. Connect Incidents to Operator Actions

An incident should help operators move from evidence to resolution.

Include:

  • Affected assets and locations.
  • Original events.
  • Relevant measurement history.
  • Data quality and freshness.
  • Related dependencies.
  • Assigned responsibility.
  • Investigation notes.
  • Confirmed cause and recovery evidence.

Preserve the initial hypothesis alongside the final finding. That distinction helps improve future rules and incident analysis.

For teams implementing this operational layer, Iotellect’s data center operations and DCIM software provides a shared platform for power, cooling, IT, network and security monitoring, with dashboards, alarms and incident workflows.

The deployment still requires validated mappings, maintained dependencies and tested operating procedures.

Commissioning Tests Worth Running

Test the full information path under representative load.

Useful scenarios include:

  1. An instrument returning an invalid value.
  2. A gateway losing communication.
  3. Buffered observations arriving out of order.
  4. An outdated dependency relationship.
  5. A cooling fault without a temperature rise.
  6. An unrelated IT fault occurring in the same period.
  7. A command without confirmation.
  8. A site uplink interruption.
  9. Repeated events after recovery.

For each scenario, verify current state, historical records, alarm behavior and incident handling.

Measure acquisition and processing delays. Operators need to understand how recent their information is.

Frequently Asked Questions

Why combine facility and IT events?

Shared context helps operators investigate relationships between power, cooling, environmental and IT conditions.

Does correlation automatically identify root cause?

No. It groups evidence according to defined relationships and timing. Confirmation may require further investigation.

Can different equipment protocols feed one model?

Yes, where suitable interfaces are available and device-specific mappings are validated.

What should happen to delayed observations?

They can enter historical storage while following explicit rules for current-state updates and live alert evaluation.

What should a pilot demonstrate?

It should demonstrate reliable acquisition, meaningful relationships, explainable correlation and a complete incident workflow—including failure and recovery.

Top comments (0)