The first look at every signal, in seconds.
When an alert fires, the on-call agent reads the pipeline before you do: heartbeat, connector and task state, the replication slot, the registry, the worker log. You get the cause, not a dashboard.
orders02:47 · heartbeat 94 s oldInvestigating heartbeat age, connector and task state, worker log.
Confirmed a task restarted after a short broker interruption.
Verified the slot caught up in 40 s. No events skipped.
Opens on real signals
Stale heartbeats, paused or failed connectors, a slot that stops moving, confirmed on two checks in a row. Noise does not open an investigation.
Reads everything that matters
Six read-only sources, sampled over a window long enough to see the slot move, not a single snapshot.
Checks before it speaks
Every finding is held against the evidence. Anything unsupported is flagged, not hidden.
Writes it down
Each investigation is sealed as an evidence bundle, write-once for one year.
Knows its limits
If the evidence does not show the cause, it says so and tells you what it could not see.
Measured, not claimed
30 of 30 blind fault trials diagnosed correctly in the Fortis lab.
From alert to answer
- Signal
The watch loop sees a stale heartbeat, a stopped connector or task, or a slot holding too much.
- Gather
It reads the six sources through read-only identities.
- Diagnose
Known causes are tested against the evidence and ruled in or out.
- Verify
The finding is checked, then written up with every source cited.
- Hand off
A fault your policy covers goes to the autopilot first. Anything else pages the person on call, with the cause and the fix.
Live in lab What it does today
- Watches change data capture pipelines on PostgreSQL, Debezium and Kafka
- Watches the PostgreSQL server itself: long and idle transactions, replication slots nobody reads, connections in use
- Notices when change data stops flowing at its usual volume, or the source schema changes, and names a schema change with the fault that follows it
- Names the platform fault underneath a data incident, such as a Kafka Connect worker with no node to run on
- Separates known causes: paused connector, failed task, registry down, CDC login locked, source idle
- Flags a downstream reader of your change data that stopped or keeps falling behind
- Writes a sealed evidence bundle for every investigation
- Pages the person on call only when it cannot settle the incident itself, never for a fix the autopilot verified
- Escalates through your rota when nobody acknowledges, and resolves the page when the system recovers
- Pages through PagerDuty as well as Teams: one incident per fault, resolved when the system recovers
Coming What is coming
- PagerDuty acknowledgements from Varqeni (built, live test pending) and back into Varqeni
- Managed change data capture services
Give your pipelines an on-call agent
Varqeni is in early access. Tell us about your pipelines and we will set up a trial around them.