Data sources and freshness
What Flux reads, how current each source is, and how to tell whether a number is complete before you rely on it.
Sources #
| Source | Provides | Cadence |
|---|---|---|
| Resource inventory | Every resource, tags, configuration | Several times daily |
| Cost management | Actual and amortized cost, daily history | Daily, checkpointed |
| Billing exports (FOCUS) | Charge-level detail, commitments, pricing | Per export delivery |
| Vendor recommendations | Advisor-style findings | Daily |
| Platform metrics | CPU, memory, disk, network | Rolling |
| Third-party monitoring | Additional utilization coverage | Rolling, incremental |
| Policy | Compliance posture, exemptions | Daily |
| Reference data | Resource types, regions, pricing units, commitment eligibility | Weekly |
Reading freshness #
Integrations shows per-source freshness: when it last completed, whether it is healthy or stale, and which scopes are complete, partial, or stalled.
Staleness is schedule-aware. A daily source is not stale an hour after its last run; it becomes stale when it misses an expected run plus a grace window. A source that has never been provisioned is reported as such, not as failing.
The two-day billing latency #
Billing data is not final for roughly two days. Flux excludes unfinalized days from totals, forecasts, and anomaly baselines.
This means month-to-date figures end a couple of days ago, and comparisons use equal spans. A “missing” recent day is nearly always this, not a fault.
Coverage states #
Telemetry carries an explicit coverage state per resource, because partial observation is common and matters:
| State | Meaning |
|---|---|
| Covered | Sufficient observation across the window |
| Partial | Some data, below the threshold for a confident finding |
| Not covered | No telemetry for this resource |
Not covered is never treated as idle. A resource with no data cannot be called unused, and Flux will not do so.
When sources disagree #
Where two telemetry sources report materially different utilization for the same resource, the finding is routed to review rather than resolved by preferring one. Both values, with their sources and windows, are shown.
Snapshot version #
Every serving instance reports which snapshot version it holds and when that version was published. The data-currency indicator in the header shows this, and the same version appears in Excel export metadata.
If two people see different numbers, comparing snapshot versions is the first diagnostic — one may simply be on an older version for a few more seconds.
Pipeline status #
Administrators have an operational endpoint reporting the whole pipeline in one call: queued and running collections with ownership and claim ages, publication version and age, apply backlog, throttle state, and the snapshot version being served.
It answers “is the pipeline moving?” without reading logs.