The pipeline migration was finished on a Thursday, the dashboards were green, and the data engineers went home tired and relieved. By the following Wednesday, three people had privately concluded that the whole effort had been a mess, and each blamed something different. That gap between a good outcome and a muddled memory is exactly what a start, stop and continue retrospective closes. Here is a step-by-step way for a DataOps team to run one after a release or sprint, told through the small, human moments that make it work.
Step one: pick the stretch of work and name it
Retrospectives drift when they cover “everything lately”. Choose a boundary the whole team recognizes: the migration of the ingestion layer, a sprint that reshaped the transformation jobs, or a month of covering the on-call rota for the warehouse. Write the boundary at the top of a blank page. A clear edge means people argue about what happened inside it rather than about where it started.
Then block the meeting while the work is still fresh. A week after is usually right for data teams, because one more batch cycle has run and the quiet failures, the ones that appear a few nights later as odd row counts, have had time to show up.
Step two: collect observations before opinions
Ask each person to jot down moments, not conclusions. “The nightly load slipped past the reporting deadline twice” is a moment. “The scheduler is bad” is a conclusion. Put the notes on a shared board without names, and read them aloud in silence if you can, simply giving everyone a few minutes to absorb them.
Data work is full of invisible effort. The analyst who quietly re-ran a failed model by hand, the engineer who patched a schema mismatch at midnight, the steward who answered the same question four times: these moments only count if somebody writes them down. This step is where they come into the light.
It also pays to bring a few artifacts to the table: the incident channel history, the list of tickets closed during the stretch, and any alert that fired more than once. Scrolling through them reminds people of episodes they had already forgotten, and it keeps the conversation tied to what actually occurred.
Step three: sort the notes into three columns
Draw three columns labeled Start, Stop and Continue, and move each note into the one it suggests.
- Start collects practices the team lacks. Example: attach a short expected-volume note to every new source, so a surprise drop in rows is recognized at once.
- Stop collects habits that waste effort. Example: stop copying extracts into personal folders for “safety”, since it creates versions no one can reconcile later.
- Continue collects what protected the team. Example: keep the habit of testing a transformation on a small sample before pointing it at the full table.
Some notes fit two columns. When that happens, split them. “We reviewed every change in a meeting” might yield a Continue for reviewing risky changes and a Stop for reviewing trivial ones.
Step four: shrink each item until it is doable
A vague wish dies in a ticket queue. Rewrite each entry so a single person could begin it next week. “Improve data quality” turns into “add a null-rate check to the two tables the finance team reads”. “Communicate better” turns into “post a three-line note in the team channel when a source schema changes”. If an item cannot be shrunk, it is a project, and it should go to planning instead of staying on the list.
Try asking which tiny slice of the idea would still alter how next week goes. Teams are often surprised that the answer is a single checklist line or a ten-minute conversation, and that the small version gets done while the grand version never would.
Step five: attach an owner and a date
Every item gets one name and one review date. The owner is not necessarily the person who does all the work; they are the person who will say, at the review, whether it happened. Teams that skip this step often discover a month later that eleven good ideas belong to nobody.
For Stop items, the owner’s job is a little different: they make the change visible. Switching off an old report or retiring a manual check should be announced to whoever relied on it, even if they relied on it only occasionally.
Step six: keep it inside a time box
Aim for about an hour. Spend ten minutes on observations, twenty on sorting, twenty on shrinking and assigning, and the remainder on a quick round where each person names one thing they are glad the team did. That last round is not decoration. People who hear their quiet contributions named are more willing to be honest about the Stop list next time.
Be careful about the temptation to treat the facilitator as a scribe. The person guiding the session should keep the time, protect quieter voices and ask what a note means when it is ambiguous, while someone else types. Splitting those jobs keeps the room moving and the record accurate.
A short story to make the three lists concrete
Picture a small team that had just changed how raw events landed in their lake. During the retrospective, a data engineer admitted spending several evenings hand-checking partition counts. That became a Start: an automated count comparison between source and landing zone. An analyst mentioned that two weekly extracts were being produced for a group that had stopped reading them months earlier. That became a Stop. And everyone agreed that the ten-minute morning sync, where the person on duty summarized overnight runs, had saved them more than once. That became a Continue. Three small sentences replaced a week of grumbling.
Where outside ideas can help
It can refresh a team’s thinking to look at how unrelated fields debrief. Aviation is famous for it, and the directory-style BestPilotsNow site, which splits its pages across pilots, schools, courses and resources and adds a forum plus a blog, illustrates how a community can arrange what it knows around a shared craft. You do not need to copy its subject matter; the point is that communities who care about craft tend to talk openly about what to repeat and what to drop.
Step seven: close the loop at the next planning session
At the next planning meeting, take two minutes to read the three lists and mark each item as done, dropped or carried over. An item carried over twice should be done or removed; keeping it forever teaches the team that the list is decorative.
If you want to connect the retrospective to technical habits, our piece on how live analytics stays steady while pipeline failures get headed off gives context for many Start items, while the overview of data operations spread over multiple cloud platforms offers further background for teams with a wider footprint.
Common stumbles and how to avoid them
- Turning the meeting into a status report. Remind everyone that the output is three lists, not a recap.
- Letting the loudest person set the agenda. Collect notes silently first.
- Writing Stop items that blame people. Describe the process instead.
- Forgetting that Continue is a decision. Say out loud what you are protecting.
- Never revisiting the lists. Put the review on the calendar before leaving the room.
The payoff once the habit sticks
Give it a handful of cycles and the pattern shifts. New joiners read the old lists and learn the team’s habits faster than any onboarding document could teach them. Stop items stop reappearing. The Continue list becomes a short, proud statement of how the team prefers to work. And the meeting gets shorter, because the observations are already half-written in people’s notebooks.
That is the real return: not any single item, but a team that treats its own routines as something it can edit, kindly and on purpose, one small change at a time.