What it catches
Most data incidents are found by the person reading the dashboard. A loader stalls overnight, an upstream team renames a column, a backfill doubles a day of payments. Nullwake turns each of those into a promise the producer writes down and a check that runs on schedule.
freshness: {maxAgeMinutes: 90}Age of the newest row, and timestamps that sit in the futureStalled or failed loadsvolume: {windowHours: 24}Rows in the latest window against a robust baseline of earlier windowsPartial loads, duplicated backfillsschema: {columns: [...]}The live table definition against the contractRenamed, dropped or retyped columnsnullable: false, uniqueKeyNull counts and duplicate keysBroken joins, replayed eventschecks: [{sql: ...}]Any business rule you can write as a single SELECTNegative amounts, orphan rows, bad codesHow it works
Contracts live in your repository next to your dbt models and change through pull requests, like the rest of your code.
Write the contract
One YAML file per table: who owns it, how fresh it must be, its columns, its keys and any rules of your own.
nullwake initdrafts them from your existing tables or your dbt project, so you start from contracts that already run.Apply it from CI
nullwake apply contracts/loads the contracts.nullwake checkcan also run every contract straight afterdbt buildand fail the pipeline, with no server needed.Checks run on schedule
Workers run each check on its own interval through a read-only warehouse role, with a statement timeout on every query.
Incidents open and close themselves
A failing check opens one incident, not a stream of alerts. The owning team hears about it in Slack or through a signed webhook, can acknowledge it and add notes, and the incident closes when the check passes again.
apiVersion: nullwake/v1
kind: DataContract
metadata:
name: orders
owner: commerce-data@example.com
spec:
connection: warehouse
dataset: analytics.fct_orders
schedule: {intervalMinutes: 15}
schema:
columns:
- {name: order_id, type: bigint, nullable: false}
- {name: amount, type: numeric}
- {name: loaded_at, type: timestamptz}
uniqueKey: [order_id]
freshness: {column: loaded_at, maxAgeMinutes: 90}
volume: {column: loaded_at, windowHours: 24}
checks:
- name: no_negative_amounts
sql: select count(*) from analytics.fct_orders
where amount < 0

Runs inside your network
Nullwake is a set of containers you run next to your warehouse, on one VM or your own container platform. Your rows never leave, and neither do your credentials.
- Read-only by construction
- Checks run in read-only transactions with a statement timeout. Table and column names are validated, never pasted in from free text, and a custom rule must be a single
SELECT. - No rows leave the warehouse
- Built-in checks read back counts and aggregates only. Warehouse error messages are cut down to their error code before they are stored or logged.
- Credentials stay in your environment
- Contracts name an environment variable instead of holding a password, and a contract with an inline secret is rejected.
- Nothing phones home
- No telemetry and no licence server. The only outbound traffic is the alerts you configure and an optional heartbeat ping you point at your own monitor.
- Every line is readable
- Apache 2.0 source, an audit log of every sign-in and change, roles for viewers, editors and admins, and a web interface that ships no JavaScript.
Read the full security overview, including the exact read-only roles to create and what an alert contains.
Built never to show a false green
A monitor that says “passing” while the data is broken is worse than no monitor. Each of these came out of an external review and is covered by a regression test.
Empty answers are errors
A rule that returns NULL or a non-number is reported as an error. It never counts as a pass and never closes an incident.
Future timestamps fail freshness
One bad row dated next week can’t hide a load that stopped yesterday.
Late results are thrown away
If a worker loses its claim on a check, its result is discarded, so an old “pass” can’t close a newer incident.
Alerts survive crashes
Alerts are written in the same transaction as the incident and retried until delivered.
It watches itself
A health endpoint for your uptime monitor turns red when checks are overdue, workers are silent or alerts are stuck.
History is never deleted
Removing a check retires it. Its runs, incidents and audit trail stay, and re-adding it brings them back.
Already on dbt? Send me your manifest
Your dbt tests already describe what good data looks like. Nullwake turns them into checks that run on a schedule and alert when the data breaks, not only when someone runs a build.
Run
dbt parseIt writes
target/manifest.json. The file holds your model code, column names and test definitions. It contains no rows of data.Email it to me
Send it to hello@nullwake.si, or a trimmed copy your team has approved. If it has to stay internal, run
nullwake init --dbtyourself: it is open source and runs on your machine.Get contracts back within a day
uniqueandnot_nullbecome column rules,accepted_valuesandrelationshipsbecome SQL checks, source freshness becomes the freshness check andmeta.ownerbecomes the owner. I note anything that needs your decision.
$ nullwake init conn.yaml --dbt target/manifest.json
8 contract(s) drafted
$ nullwake check conn.yaml contracts/
PASS orders.unique.order_id
PASS orders.not_null.customer_id
PASS orders.schema
FAIL orders.check.relationships_customer_id
1 order points to a customer that doesn’t exist
FAIL orders.check.accepted_values_status
1 order has a status outside the allowed list
Start with a 30-day pilot
Free for design partners. I set it up with your team, inside your network, and we find out together whether it catches problems that matter to you.
- Deployment on one VM or your container platform, together on a call
- Contracts for your 5 to 10 most important tables, written with you
- Alerts routed to your Slack channel or incident tool
- A short weekly review of what fired and what it missed
After the pilot, a yearly subscription per deployment covers setup help, support and priority fixes. You get a quote before the pilot ends, with no obligation to continue. The software itself stays open source under Apache 2.0.
What you need
- Postgres 14 or later, Snowflake, or BigQuery
- A small Linux VM or a container platform inside your network
- A read-only warehouse role (we create it together)
- Slack or any HTTPS webhook endpoint
Who builds it
Nullwake is built by Akhil Pettem, a data engineer in Bengaluru, India. It is an early, independent product: you talk directly to the person who writes the code, and your feedback shapes what gets built next.
Where it stands, November 2026
- Version
- 0.6.2, running in production at app.nullwake.si
- Postgres
- Tested end to end against real databases
- Snowflake
- Verified against a live account: all six check types pass on healthy data and planted defects are caught
- BigQuery
- Verified against a live project with a read-only service account: all six check types pass and planted defects are caught
- Next
- Single sign-on (OIDC), PagerDuty and Opsgenie alerts, contracts generated from dbt, a Helm chart