🌙 Nightshift

The on-call data team that gets smarter every night.

Live demo · Break it yourself · JUDGING.md (60s) · Upstream PR datahub-skills#126 · Apache 2.0

try.* is the judge sandbox (break / wake / restore on a real graph). /app is the connected war room (your DataHub, Sentinel, history). Same agent loop; different entry.

At 2:47am an upstream team renames a column and tells nobody. By 9:07am the revenue dashboard reads zero, and Finance notices before the data team does. Nightshift takes that pager: a Claude agent wired into your DataHub graph works the incident like a senior engineer, and then does the thing humans never have time to do at 4am: it writes down what it learned, inside the graph itself.

The second time a pipeline breaks the same way, Nightshift does not investigate. It remembers. An investigation becomes a lookup.

The night, in numbers

Measured on DataHub's showcase-ecommerce datapack (1,049 entities), from the replayable shift reports in examples/shift-reports/ (wall-clock shift length + investigation tool calls counted in those reports):

	Night 1 (cold)	Night 3 (memory)
Investigation tool calls	14	5, including exactly the 2 reads memory prescribed
Shift wall-clock	2.2 min	1.1 min
Lineage re-walked	full path	none

By night 3 the agent recognized a recurrence, reframed the incident from a SQL problem to a process problem, escalated to the model's owner, and proposed a dbt source test so CI would block the regression. See the real draft PR: nightshift-dbt-demo#1.

Architecture

flowchart LR
    P([📟 the pager]) --> A

    subgraph agent [Nightshift agent (Claude)]
        A[1 RECALL] --> B[2 LINEAGE]
        B --> C[3 DIAGNOSE<br/>one root cause]
        C --> D[4 REMEMBER]
        D --> E[5 FIX PR<br/>draft]
    end

    subgraph dh [DataHub OSS]
        M[(the graph remembers<br/>postmortems · tags · guards)]
    end

    A -- "recall_incident_memory" --> N
    B -- "search · schemas · lineage + SQL" --> O[datahub MCP server<br/>official, reads]
    D -- "incidents · assertions · memory" --> N[nightshift MCP server<br/>ours, writes]
    O --> M
    N --> M
    E --> R([☀️ the morning report])

The official mcp-server-datahub covers the read surface completely. It has no write surface for the on-call loop: no incidents, no assertions, no memory. The Nightshift MCP server is that missing write surface, built on the OSS GraphQL API and metadata model, and upstreamed as a skill PR (plus a packaging bug report found along the way). Full detail in docs/architecture.md.

What the agents leave behind

Open the DataHub UI after a shift and every conclusion is already there, in the surfaces your team uses today:

• The incident, raised and resolved in DataHub (open_incident / resolve_incident).
• The postmortem, written into the dataset's documentation (prose for the human at 9am).
• A machine-readable memory, a structured property in JSON, for the next agent.
• A searchable failure-mode tag for everyone, via find_datasets_with_failure_mode.
• A column-presence assertion on the field that broke, visible in the Validations tab (guard_column). Honest scope: it marks that the column exists in the catalog; value-level checks (not-null / sum > 0) stay a follow-up for dbt tests / CI.
• A concrete fix: a dbt change built from the columns the catalog actually holds: a real draft PR.

Nightshift itself is stateless. The memory lives in DataHub aspects, so it survives restarts and model changes, and any MCP-capable agent pointed at the same graph inherits it.

One incident, whole-graph immunization

When a shift closes, Nightshift asks the graph one more question: where else does this exact exposure exist? immunize_graph finds every dataset carrying the same column, across platforms, and leaves an idempotent presence guard on each one. On the demo graph that is about 10 datasets over dbt, Looker, PowerBI, S3 and Postgres in a single call: datasets that never broke get the same Validations marker from an incident they never had.

The Sentinel

Nobody presses the button. The Sentinel fingerprints the schema of every watched dataset on an interval. When a column moves (renamed, dropped, or retyped), it names the drift and wakes the night shift itself (trigger: sentinel). The loop closes without a human pager at all.

Authority boundaries

• Agent may write (Nightshift MCP): incidents, documentation, structured memory, failure tags, column-presence EXTERNAL assertions, draft fix PR.
• Still human: merge the PR, value-level tests, production deploy.
• Memory lives in DataHub aspects, not in chat history. Nightshift is stateless; the graph remembers across nights and model changes.
• Presence guards assert the column exists in the catalog. They are not value-level quality tests (not-null / sum > 0 stay in dbt / CI).

Claims in this README are gated by python scripts/verify_judging_evidence.py.

Quick start

Prerequisites: Docker with ~8 GB of memory, and either an authenticated claude CLI or an ANTHROPIC_API_KEY.

make up datapack   # DataHub + a realistic 1,049-entity enterprise graph
make setup         # install Nightshift
make demo          # silently break the pipeline, hand the agent the pager

If datahub datapack fails on a fresh PyPI install of acryl-datahub 1.7.0 (missing resource file), that is datahub#19028. Workaround until the fix lands: install the CLI from the DataHub repo (pip install -e ./metadata-ingestion from a clone) or use a build that still ships the datapack assets, then re-run make datapack.

make demo renames an upstream column, tells nobody, and prints the morning report when the shift ends. nightshift war-room renders the night as a single dark page. Replayable shift reports and postmortems live in examples/.

Built on

The DataHub MCP Server and the DataHub Agent Context Kit. Nightshift adds the write surface OSS agents were missing and gives it back upstream.

License

Apache 2.0