The first time a customer sees the quarantine metric, the reaction is always the same: four percent of our records are failing? That's a problem with your pipeline.
It is not. It is the pipeline working, and the alternative is considerably worse.
What quarantine is
Records arrive that cannot be resolved with adequate confidence. Two customer records that might be the same person and might be a parent and child with the same name. An address that resolves to three different postcodes. A supplier whose legal entity changed and whose historical rows now straddle two identities.
Something has to happen to these. There are exactly three options.
Guess. Pick the most likely resolution and carry on. This produces a clean-looking pipeline and silently wrong decisions downstream, which surface months later as a pattern of individually-explicable errors.
Drop. Discard them. Now your decisions are made on a filtered population, with a filter nobody characterised, correlated with exactly the messy edge cases that matter most.
Quarantine. Hold them, flag them, route them for remediation, and exclude them from the decision path until they are fixed or expired.
Only the third is defensible, and it is the only one that produces a number people find alarming.
Why the number is reassuring
A reject rate of zero means one of two things: your data is perfect, or your quality bar is not doing anything.
We have never seen the first. Enterprise data comes from systems built over decades by teams that never agreed on a schema, migrated twice, and merged with an acquisition. A pipeline that finds nothing wrong with it is not looking.
Expect a non-trivial rate, higher in the first months and settling as remediation catches up with the backlog. What the steady-state number should be depends entirely on your source systems — the useful comparison is against your own trend, not against someone else's benchmark.
How quarantine goes wrong
It becomes a landfill. Records go in, nothing comes out, and eighteen months later it holds 400,000 records nobody has looked at and the metric has been muted in the dashboard.
Three things prevent this.
An owner. A named person or team responsible for the remediation queue, with it in their objectives. Not "the data team" — a person.
An expiry policy. Some records will never be fixable. The source system is gone, the customer is gone, the context is unrecoverable. Say so explicitly, expire them on a schedule, and record that they were expired rather than resolved.
Visible metrics. Inflow, outflow, age distribution, and the reason breakdown. The age distribution is the one that matters — a stable total can hide a queue where nothing is moving.
The reason breakdown is free diagnostics
Quarantine reasons, aggregated, are the highest-signal data quality report an organisation gets, because they are generated by actual use rather than by a profiling exercise.
One customer found that 40% of their quarantine came from a single upstream form that let users type a country name free-text. Fixing that form removed more bad data than the previous year's data quality programme.
That finding cost nothing. It fell out of counting rejects by reason, which is a thing you can only do if you have rejects.
The sentence to keep
A funnel with no rejects is not one anybody has run.
If your quality gate has never stopped anything, it is not a gate. It is a log line.
- Data
- Quality
- Governance