Data teams tend to treat their platform as a product and the service desk beneath it as somebody else’s problem. Then a nightly load slips past breakfast, three dashboards go stale, and the same four people spend the morning answering the same messages. That gap is where ITOps habits pay off. Below are eight of them, each translated into the language of pipelines, warehouses and datasets, with notes on how to adopt them without building a bureaucracy.
The habits come from our reading of the Atlas page that defines ITOps for the xOps School glossary, which describes the function as the work of keeping services running: taking requests, handling incidents, removing repeat causes, controlling change and maintaining a map of what exists. Everything numeric below is drawn from that entry; the data-platform translation is ours.
1. Separate the request lane from the incident lane
A request is something you expect to happen: a new analyst needs warehouse access, a dataset needs a new column exposed. An incident is something broken. When both land in one chat channel, urgent breakage waits behind routine asks. Give each lane its own queue and its own clock, and your on-call person stops context-switching between a password reset and a failed ingestion job.
2. Write priority as a formula, not a mood
The Atlas shows a sample convention from a popular service platform where you add impact to urgency and subtract a single point. Nobody has to copy that exact sum, yet the principle deserves adopting: priority follows from two stated inputs. For a data team, impact might be the number of downstream consumers or whether a finance close depends on the table; urgency might be how soon the data is needed. Once the formula is written, arguments about whose ticket jumps the queue become arguments about inputs, which are easier to settle.
3. Make service-level clocks honest
One of the Atlas labs follows an incident that stayed open for 92.5 clock hours although its target clock only advanced 4.5 of them, since a weekend and a holiday paused the timer. Teams that never read their own calendar rules misreport compliance in both directions. If your analytics platform promises a data-freshness target, state which hours count, what pauses the clock, and who may pause it. The same lab produced 6 of 8 targets met, or 75%, a number that looks poor until you notice how it was calculated.
The Atlas also shows how a customer-facing target can be decomposed. A customer promise of eight hours was carved into thirty minutes of desk triage, a two-hour diagnosis slice for the network team, a four-hour slot for the external carrier and ninety spare minutes. Apply that to a late dashboard: ingestion owner, transformation owner, BI refresh owner, and a buffer. The slice each team owns becomes explicit.
4. Keep a catalogue of what exists, and keep it honest
Configuration records are the ITOps name for the inventory of components and how they relate. In a data platform, that is your list of sources, pipelines, tables and dashboards. The Atlas warns against recording everything; the inventory needs what decisions require. Its health lab sampled ten records: nine were complete, two had gone untouched for more than two months, one belonged to nobody and one repeated another record’s serial. Even in a sample that small, hygiene issues appeared. Run an equivalent check on your dataset catalogue each quarter: owner present, last-reviewed date recent, no orphan pipelines feeding nobody.
If you want examples of what a well-run pipeline estate looks like in practice, our write-up on real-world implementations teams have shipped gives a broader view of the engineering side.
5. Classify changes before you review them
The Atlas distinguishes three change types. Standard changes are low risk and pre-authorised, so they need no committee. Normal changes get assessed. Emergency changes happen immediately, such as patching against active exploitation, and are reviewed after. For data work: adding a documented, tested column is standard; altering a table schema that finance reads is normal; rolling back a broken transformation at 02:00 is an emergency. Treating all three alike either slows every release or gives you no scrutiny where it matters.
6. Give change models a failure limit
One mechanic from the Atlas is worth copying outright: a change model with a limit on its own track record, for instance pulling a model out of service once it has failed more than once across its latest twenty runs. A standard change type earns its pre-authorisation by behaving. If your routine schema migration has failed twice in twenty runs, it has lost the right to skip review until someone fixes the template.
Alongside that, the entry recommends a written risk score with weights. Its example scored a change at 6 by adding three points for a tier-one system, two for a schema change and one for landing outside the maintenance window, which routed it to the advisory board. A data version: weight by downstream reach, by whether the change touches a shared model, and by timing relative to a reporting deadline.
7. Treat the alerting layer as production
The Atlas names a common anti-pattern: the event layer handled as if it were not a production dependency, so it can fail without anyone noticing. In data terms, the freshness monitor and the row-count check are production systems. If the checker itself stops running, you get silence, and silence looks like health. Alert on the absence of heartbeats, not only on bad values.
Grouping matters too. The Atlas lab turned thirteen alert rows into two incidents, then saw three late alerts open three more after the window closed. Choose a window length deliberately and review how many late arrivals it produces. If streaming is your worry, keeping real-time analytics from breaking at the pipeline covers that ground.
8. Kill the shadow spreadsheet
The last anti-pattern on the Atlas list is the shadow spreadsheet: a private tracker kept because the official tool is too awkward. Data teams are especially prone to it because they are good with spreadsheets. When you find one, treat it as feedback. Ask what the official queue lacks, then fix that gap instead of banning the sheet.
Where these habits tend to go wrong
Two failure modes show up repeatedly when data groups adopt service-operations vocabulary. The first is process added as a reflex after an outage: a bad load happens, so a new approval step appears, and six months later every release crawls. The Atlas names this pattern as a pitfall. Before adding a gate, ask whether a better check, a smaller change or a faster rollback would have prevented the same failure. The second is measuring only what the tool makes easy. A ticketing queue will happily report average close time while hiding that the same table has broken five times this quarter. Track repeats, because removing repeated causes is the part of the function that compounds.
A caution on borrowing too much
The Atlas is equally clear that heavyweight service management forced on one small team running one service is a mistake. If your platform has a handful of users, you need the habits above, not a certification programme. The frameworks it lists range from certifiable standards to a free European framework for small organisations, so scale your adoption to match.
If your platform spans several providers, the sequencing question gets harder, and our multi-cloud data operations guide for modern teams is a sensible next read.
A one-week starter plan
- Day one: split request and incident queues, even if both live in the same tool.
- Day two: write your priority formula on one page and agree it with two stakeholder teams.
- Day three: list change types and mark which already behave as standard.
- Day four: run the catalogue health check on twenty records.
- Day five: confirm that a stopped monitor would page someone.
Nothing on this list calls for a purchase order; what it calls for is that the people who keep the data flowing agree, in writing, on how work is sorted, timed and checked. That agreement is most of what the discipline amounts to.