Skip to main content

Use AI/ML

The AI/ML area currently supports training through AutoML, generating augmented data, and inspecting selected MLflow experiments and registered models. It is not yet a complete train-to-serving workflow.

Choose the workflow​

GoalStart here
Train from a fileMachine Learning → Auto ML → Upload CSV
Train from a tableMachine Learning → Auto ML → Lake House
Create augmented dataMachine Learning → Auto ML → Synthetic Data
Reopen a terminal AutoML runMachine Learning → Auto ML → Run History
Inspect MLflow runsMachine Learning → Experiments
Inspect registered model versionsMachine Learning → Models

Oqullus currently labels this navigation group Machine Learning. The paths above preserve the deployed UI wording until the product adopts the broader AI/ML label described in UX-ML-004.

Train from a CSV​

  1. Open Machine Learning → Auto ML and select the CSV input.
  2. Upload the dataset.
  3. Choose the task intentionally. Select a target column to predict that value, or leave the target empty to discover clusters. An empty target does not ask AutoML to infer one.
  4. Optionally request task suggestions. Suggestions profile the dataset; they do not start training or reserve compute.
  5. Review the displayed training options and start the workflow.
  6. Keep the page open while Oqullus polls the run. Record the run ID if you need to investigate it later.

AutoML stores the upload, creates a workflow, and runs ingestion, exploratory analysis, feature engineering, model selection, and training. The current no-target stage plan also shows Labeling. On the default request path, its generated fallback rule creates no useful label, but the stage can still appear successful; clustering creates its separate cluster output later during model validation. AI Gateway assists with LLM reasoning; it does not train the model.

Train from a Trino table​

  1. Open Machine Learning → Auto ML → Lake House.

  2. Select the catalog, schema, and table.

  3. Choose the task intentionally:

    • select a target to predict it through classification or regression;
    • leave the target empty to cluster rows from their numeric features.

    AutoML does not infer a target when this field is empty. See UX-ML-003.

  4. Set a row limit. A limit is optional in the UI, but using one is the current safe workaround because extraction is not bounded automatically before data is materialized.

  5. Optionally list columns to exclude.

  6. Set the train split between 0.5 and 0.95.

  7. Select Run and monitor the returned run ID.

The current deployment queries Trino with the AutoML service identity, not the signed-in user's SQL Editor/Ranger identity. Access behavior may therefore differ from running the same table query in SQL Editor. See SEC-ML-003.

Understand compute selection​

Users do not select pandas, local Spark, or a distributed cluster. AutoML assigns an internal small, large, or massive algorithm tier from estimated data size and row count.

In the assessed deployment all Spark tiers still use local[*] in the AutoML API pod. The names large and massive do not mean that distributed compute was created. Keep datasets within a measured local limit until ARCH-ML-001 is resolved.

Read progress and results​

Oqullus polls the AutoML backend and presents the current pipeline stage. A completed result may include:

  • ingestion and data-quality summaries;
  • selected and engineered features;
  • selected model and evaluation metrics;
  • warnings and recommendations;
  • prediction inputs when the trained artifact supports them;
  • a generated report or notebook where available.

Run History currently shows terminal runs. Active-run recovery depends partly on browser storage, so another browser or device may not rediscover running work. There is no visible Cancel control. If a run remains active after a pod failure, an operator must inspect it; do not assume it is still executing.

Generate augmented data​

  1. Open Machine Learning → Auto ML → Synthetic Data → Generate.
  2. Upload a CSV.
  3. Optionally choose a target column. Leaving it empty uses unsupervised generation.
  4. Choose a whole-number row multiplier. The backend accepts 0.1–5, but the current UI accepts whole numbers starting at 1 and does not display the maximum.
  5. Select Generate Synthetic Data and wait for polling to reach Completed or Failed.
  6. Review quality, fidelity, coverage, model, generated-row, deduplication, and warning information.
  7. Use History to reopen a terminal generation run.

The generated result is currently an augmented dataset, not synthetic-only data: it contains the original rows shuffled together with generated rows. It must retain the original dataset's security classification. Low quality or a failed quality evaluation produces a warning rather than blocking completion.

The backend limits input to 200 MB and original plus requested rows to 500,000. Oqullus currently displays the internal result path but does not provide an authorized download action. See DATA-ML-002.

Inspect experiments​

  1. Open Machine Learning → Experiments.
  2. Find an experiment and expand its fetched runs.
  3. Open a run to inspect parameters, metrics, artifacts, traces, or system metrics when those records exist.

Experiments is read-only. There is no direct View Experiment transition from AutoML, and AutoML ownership tags do not consistently match the page's filter. An AutoML run may therefore be present in MLflow but absent for its author.

Inspect registered models​

  1. Open Machine Learning → Models.
  2. Select an existing registered model.
  3. Inspect its versions and follow available lineage to the source MLflow run.

Models is read-only. Supervised AutoML does not register its candidate, and the current UI does not provide Register, Alias, Promote, Deploy, Roll back, or Delete actions.

What is not currently supported​

  • durable queued AutoML execution and automatic recovery after pod loss;
  • immediate end-to-end cancellation of training or synthetic generation;
  • reliable navigation from AutoML to its MLflow experiment and model version;
  • registration of a supervised AutoML candidate through Oqullus;
  • trained-model serving, rollout, monitoring, or rollback;
  • distributed large or massive AutoML compute;
  • using AI Gateway as the serving destination for an AutoML model.

For assessment evidence rather than user instructions, see evidence and reproduction. For system boundaries, see Machine Learning architecture.