Model Operations
Machine learning on System9 is not a separate console bolted onto the data
platform. It is expressed in SQL against the catalog, a model is a catalog
object, a prediction is a SQL call, and an evaluation is a query, which is what
makes it reachable to the analysts who already own the data. On the customer
console this surface is Model Operations, reached at /model-operations.
In one paragraph
A model is a catalog object created with DDL; you predict and evaluate in SQL; AutoML searches the candidate space under a budget; a drift monitor and a retrain controller manage the lifecycle; and federated training aggregates cell-signed parameter updates so a model trains across data that never moves.
The ML control plane
| Capability | What it does |
|---|---|
| Model DDL | A model is a catalog object created and altered like a table, with a DDL handler. |
| Predict / evaluate in SQL | PREDICT over a table, EVALUATE in SQL, a scoring service, and turnkey forecasting: all as SQL surfaces. |
| AI & vector SQL functions | Each behind a governor, so an invocation is policy-checked rather than open. |
| Training assets | A training-set builder, feature signatures, feature groups and materializations, experiments, evaluations, and an evaluation harness. |
| Drift & retrain | A drift monitor and a retrain controller, so a degrading model is detected and re-trained under policy rather than silently decaying. |
| Promotion & guards | Promotion with approval and a lease, artifact-trust and reachability checks, a column guard and a relocation guard so a model cannot quietly reach data or move to a cell policy forbids. |
| AutoML | Hyper-parameter search with k-fold cross-validation and a leaderboard, under hard budget_trials / cv_folds caps and a fixed seed for reproducibility, plus pluggable training backends behind one interface, an export gate, and an artifact packer. |
| Federated training | Each cell trains on its local shard and ships a cell-signed parameter update to a control-owner cell that aggregates a sample-weighted quorum: training across data that never moves. |
ML as a SQL surface, governed
A predict call is a SQL call, so it enters the same governed path as any other query, the model is a catalog object with a classification, the invocation passes a governor, and a guard refuses a model that would reach data or a cell it should not.
*Figure: ML expressed as governed SQL. CREATE MODEL / PREDICT / EVALUATE
pass a function governor; the model is a catalog object carrying
provenance, lineage, and classification; AutoML, drift → retrain, and
promotion run under policy, with a column guard and relocation guard
refusing a model that would over-reach; and federated training aggregates
cell-signed updates without moving data.
Table equivalent
| Stage | Mechanism | Governance |
|---|---|---|
| Author | Model DDL | Model is a catalog object |
| Invoke | Predict / evaluate in SQL | Function governor policy-checks the call |
| Search | AutoML (HPO + k-fold CV, budgeted) | Hard trial/fold caps; fixed seed |
| Maintain | Drift monitor → retrain controller | Retrain under policy |
| Promote | Approval + lease + column/relocation guards | A model cannot over-reach data or cells |
| Train distributed | Federated training (signed updates, quorum) | Data never moves |
Why it is different
- The people who own the data can do the ML. Because a model is a catalog object and prediction is SQL, an analyst does not leave the query surface to score a table.
- A model is governed like any asset. Provenance, lineage, classification, and the column and relocation guards mean a model is subject to the same policy plane as the tables it reads.
- Federated training moves the math, not the data. Cell-signed parameter updates aggregated by quorum let a model learn across an estate whose data stays put, the same data-gravity discipline as the scale-out plane.
Across the deployment matrix
- Single-cell, model DDL, predict/evaluate, and AutoML run against the cell's own catalog and engines.
- Multi-cell: federated training aggregates across cells; each cell trains locally and ships only a signed update.
- Sovereign / air-gapped, the whole surface is local; there is no hosted ML service in the path. Note the AI group is off by default and must be opted up.
- Cloud-enabled: identical surface; third-party model reach is a separate tier decision (see the AI system).
Related
- The AI system (core + edge), the tier model and the governance every AI call passes.
- Model Serving & Provisioning, the pipeline that signs and serves an artifact.
- The catalog and lineage, the catalog a model is an object in.
- Scale-out & distributed execution, the data-gravity model federated training follows.