Operational, analytical, reporting: different rules

Three builds that get confused with each other

The word dashboard covers three different engineering objects. They diverge on latency requirements, on history depth, on the number of metrics and on what counts as a failure. Built by the rules of one type but used as another, a dashboard is guaranteed to disappoint: an operational screen refreshed once a day is useless, and a reporting screen rebuilt hourly argues with itself every month.

Operational

Answers the question "what is happening right now and do I need to step in". The viewers are people on shift, often on a wall screen or a phone. Data latency is minutes, refresh is automatic, history is one or two days, there are 4 to 8 metrics on screen, and each one carries an explicit threshold of normal. Here it is fine to trade accuracy for speed: an approximate number in a minute beats an exact number in an hour. A failure means the feed stopped, not a mismatch in the third decimal place.

Analytical

Answers the question "why did it turn out this way". The viewers are analysts and product owners, looking occasionally and with the intention to dig. Latency is a day, history runs from a year, there are 10 to 20 metrics, and breakdowns are mandatory: segment, channel, cohort, geography. Filters and drill-down earn their place here, and accuracy outranks speed: a mismatch with financial reporting is not acceptable.

Reporting

Records the result of a period for the people who report it outward or allocate resources against it. Viewed once a week or once a month. The defining property is immutability: numbers for a closed period must not change after the fact. That requires snapshots. Values as of the close date are stored separately, and later corrections to the data do not touch them.

                  operational    analytical      reporting
latency           up to 15 min   up to 24 h      until close
history           1-3 days       12-36 months    36 months
metrics           4-8            10-20           6-12
breakdowns        none or one    many            fixed
recompute past    forbidden      allowed         forbidden after
                                                 close
what is a failure feed stops     mismatch with   a closed number
                                 the source      changes

How to identify the type up front

Ask what happens if the data arrives two hours late. For an operational screen that is an incident. For an analytical one it goes unnoticed. For a reporting one it does not matter at all as long as the period is still open. The second question: is it acceptable for yesterday's number to change tomorrow. For operational it is acceptable and normal, for analytical it usually is, and for reporting after the close it is not acceptable under any circumstances.

Both answers come from the business, not from the data team, and they take about two minutes to collect. Ask them in the first conversation and write them into the passport. Every later argument about refresh schedules, storage cost and query cost resolves against those two lines.

Cheat sheet

• Operational: minutes of latency, 4 to 8 metrics, thresholds of normal.
• Analytical: a day of latency, breakdowns, accuracy first.
• Reporting: snapshots as of the close date, numbers never move afterwards.
• Type is decided by two questions: the cost of late data and whether the past may be recomputed.