The schema holds. The meaning moves.
A column stays NOT NULL VARCHAR while its top value quietly flips from card to wallet. Every test passes. Every dashboard updates. Every number is wrong.
61%of incidents pass all existing tests
Baseline · last 48 runs · payments.checkout
4.2B
rows profiled every day
900
pipelines under watch
11 min
median time to detection
3%
of alerts dismissed as noise
Why you find out last
Nothing throws an exception.
Pipelines stopped failing loudly years ago. Now they succeed, on time, with the wrong numbers in them.
The model degrades on a curve.
No page, no stack trace — just precision sliding half a point a week until somebody questions the quarter and the retro starts three months late.
5.2weeks to notice a silent regression
A human tells you first.
The first signal is a message from someone in finance asking why a number looks odd. By the time you open the query, it's archaeology.
1 in 3incidents first reported by a person
How it works
Connect, baseline, watch.
01
Connect
Point Driftline at a warehouse, a stream, or a feature store. It reads metadata and column statistics through credentials you issue and can revoke. Row values never leave your account.
02
Baseline
Driftline profiles fourteen days of history and learns the normal shape of every column: distribution, cardinality, null rate, numeric range, arrival cadence. That shape becomes the baseline everything is scored against.
03
Watch
Every run gets scored. When a column moves past your threshold you get the column, the segment it moved in, the run that introduced it, and the diff against baseline — in Slack, before the dashboard refreshes.
What it watches
Twelve ways a column lies to you.
Each one is scored on every run, per column and per segment, against your own history.
distribution_shift
Population stability index against the fourteen-day baseline.
cardinality_break
A dimension gains or loses distinct values faster than its own history allows.
null_rate_spike
Nulls arrive where they never used to, or stop arriving where they always did.
arrival_lag
The run landed, but later than its own p95. Freshness drift before staleness.
segment_collapse
A segment that was 8% of rows is suddenly 0.2%, while the total holds steady.
range_drift
Numeric bounds creep past the minimum and maximum the column has ever held.
category_churn
The top categories reshuffle: yesterday's leader is today's fourth.
volume_anomaly
Row count leaves the day-of-week seasonality it has followed for months.
join_fanout
A join that returned 1.0 rows per key now returns 3.4. Duplicates with a clean exit code.
freshness_gap
Upstream stopped; downstream kept computing on yesterday.
type_coercion
Strings that used to parse as dates quietly stopped parsing.
duplicate_surge
The natural key is no longer natural.
“A payments column changed meaning on a Tuesday. Driftline flagged it at 11:04 and we shipped the fix before the weekly report ran. That used to be a three-week archaeology project with four people in a room.”
Staff Data Engineer, Halcyon Freight
Pricing
Priced per pipeline, not per seat.
Everyone on the team can read every alert. You pay for what Driftline watches, not for who looks at it.
Starter
$0 / month
3 pipelines, 7-day baseline, Slack alerts.
Team
$480 / month
40 pipelines, 14-day baseline, segments, Replay.
Scale
Custom
Unlimited pipelines, private deployment, SSO.
Find out what moved last week.
Connect one pipeline. Driftline backfills a baseline from your history and shows you the drift you already shipped.