A marketplace's technical dashboards are all green during a checkout failure that costs significant revenue. What kind of monitoring was missing?

What was missing

Monitoring of the business outcome rather than of the components. Every service can report success while the user journey fails, because the failure was in the composition: a valid response containing an empty result, a client-side error, a third party returning success with an error body, a feature flag misconfigured so a step is skipped.

Technical metrics measure whether the system did what it was asked. Business metrics measure whether the right thing happened.

What to monitor

• Funnel conversion rates at each step, with alerts on deviation from the expected pattern for this time of day and day of week. A checkout completion rate dropping by a fifth is unambiguous and is detected in minutes.
• Absolute counts of key events — orders placed, payments succeeded, sellers onboarded — because a rate can look normal while volume has collapsed.
• Comparisons against the seasonal baseline, not against a fixed threshold, since a marketplace's normal traffic varies enormously by hour and by day and a fixed threshold either never fires or fires constantly.
• Segmentation on the dimensions where failures concentrate: platform, app version, region, payment method, seller tier. An aggregate that looks normal frequently hides a total failure in one segment, and the segment is often the newest one.

Why this catches what technical monitoring cannot

The classes of failure it detects: a client release breaking a flow, a third party returning success with a failure body, a configuration change disabling a step, an experiment with an unintended effect, and any failure whose symptom is "nothing happened" rather than "something errored".

Absence of activity produces no errors, and technical monitoring is built around errors.

The implementation caution

Business metric alerts need care to avoid becoming the noisiest source in the system. They are inherently noisier — a marketing campaign, a holiday, a competitor's outage all move them legitimately. Alert on a large sustained deviation from a seasonally-adjusted baseline, not on any deviation, and route them to a channel where someone can apply judgement rather than to a pager.

The organisational value

These are the metrics an executive already watches, which means an alert on them needs no translation. The conversation about whether an incident matters is over before it starts, and that alone changes how quickly reliability work gets prioritised.