On a night shift in a warehouse, a supervisor watches a dashboard that says every robot completed its routes. Yet by morning the picking count is short and nobody can say where the missing hours went. The machines were fine. The trouble lived in the data trail behind them: logs that arrived late, a counter that reset, a report that quietly dropped a batch. Teams that run robot fleets generate a stream of telemetry, and the discipline that keeps such streams honest is DataOps. This checklist translates its ideas into the language of a robotics floor, using the Ops Atlas entry as the reference.
First, a short human picture
Think of the telemetry pipeline as another conveyor. Sensor readings and task events enter at one end; reports for managers come out at the other. In between are steps that clean, join and summarise. Like any conveyor, it can jam, run empty or deliver the wrong parcel, and a green status light says only that the belt is moving. The Atlas defines DataOps as running such pipelines with software discipline, meaning versioned and tested changes, audited loads and monitoring afterwards. The checklist below follows that order.
Checklist A: Changes to the pipeline
- Everything lives in version control. Transformation logic, schedules and alert thresholds can each be traced to a reviewed change.
- Tests run before a change goes live. They use a sample or a copy, not the real production feed.
- Sandboxes hold masked or synthetic records. Operator identifiers and shift details should not travel unmasked into practice environments.
Checklist B: Each night’s delivery
This is the part robotics teams most often overlook, because machinery does not change overnight and it is tempting to assume the data behaves the same way. The Atlas names it the run loop: data keeps changing even when no code does, so each load needs its own checks.
- Freshness. Is the latest event from the right shift? A fleet that stopped reporting still lets the job succeed on an empty file.
- Volume. Does the count of task events resemble a normal day of the same type? Compare a Monday with earlier Mondays, because a weekend shift pattern is a different process.
- Keys. Is every task identifier unique and non-blank? Duplicates inflate throughput figures.
- Schema. Did firmware or a vendor update rename a field or change a unit?
- Repeatability. If a night load is rerun after an error, does it replace that night’s partition and leave the totals unchanged? Appending would double the count.
- Late arrivals. Do you reconcile partitions with the source counts and re-read a short window of previous days, so that robots reconnecting after an outage are included?
Checklist C: Alerts that deserve attention
- Baselines respect the shift pattern. A limit set at three standard deviations either side of the mean presumes the process behaves the same way every day. In the Atlas lab, limits from all days let a partial load about 19% under a normal weekday through, while a weekday-only baseline raised the flag. A factory with weekday and weekend rosters has the same problem.
- False-alarm load is counted. At 0.0027 per check, hundreds of daily checks will fire by chance. If operators learn to ignore the alerts, the monitor has failed.
- Blocking versus warning is decided. Thousands of warn-only tests hide the important ones. Pick the few that must stop a report from publishing.
Checklist D: Ownership and measurement
- Every critical feed has an owner and a written agreement. Whoever produces robot telemetry states what the fields mean, what quality it promises and when it will arrive. A renamed status that keeps the same data type will pass a structural check and still corrupt your reports.
- Pipelines are tiered. Safety-relevant and billing-relevant feeds receive contracts, blocking audits and on-call cover; routine dashboards get key tests; experiments get Git only.
- Incidents are logged with detection times. Record who noticed first. The Atlas proposes the share of incidents found first by data users as a headline measure; in its worked example, most of the downtime was the time before detection.
A note on cost and rigour
Not every feed deserves every check. Holding data back until audits pass, the write-audit-publish pattern, protects readers but adds delay and sometimes duplicate writes, so the Atlas keeps it for the top tier. In a robotics plant that usually means the feeds behind safety reports, maintenance triggers or customer billing. A utilisation chart on a wall screen can tolerate a lighter regime. Scoping matters for money too: a full-history scan on a large table every day can cost far more than a partition-scoped test, and the entry works through an example where the gap is roughly a thousandfold under per-byte billing. Plan the checks, then plan what they cost to run.
Putting a score on it
Count the boxes you can tick with proof, such as a log, a commit or an alert export. Fifteen items are listed; below eight, start with Checklists A and B; between eight and twelve, work on C; above that, focus on shrinking the share of incidents your people spot before the monitors do. Redo the exercise each quarter, because fleets and firmware change.
When you are deciding which numbers a plant manager should see, the guide to robotic KPIs for factory managers is a good starting list, and the audit above helps you trust them. For the data feeding those numbers, how sensors connect into industrial robots explains where the readings originate. The full definitions, scope notes and cautions are in the Ops Atlas page on DataOps, which says plainly that the whole field is young and lacks a formal standard.