Skip to content

feat: generate an exact number of rows matching a declared criterion #790

Description

@nabinchha

Problem

DataDesigner users can generate candidate rows and filter them afterward, but they cannot declaratively request an exact number of rows that satisfy a quality or validation criterion. Doing this today requires user-managed loops across multiple create() calls, including candidate budgeting, artifact extension, trimming, and resume behavior.

Examples include generating 5,000 answers above a judge threshold, 10,000 policy-compliant conversations, or a fixed number of successful tool-use traces.

Proposed feature

Add engine-native record selection to a normal DataDesigner.create() run. A user declares a boolean predicate column and a hard candidate limit. When selection is configured, num_records means the desired number of accepted output rows:

builder.with_record_selection(
    dd.RecordSelectionConfig(
        predicate_column="meets_criteria",
        max_candidate_records=20_000,
        on_exhausted="raise",
    )
)

results = data_designer.create(builder, num_records=5_000)

The engine generates immutable candidate batches, evaluates the predicate, checkpoints accepted rows, and continues until the target is met or the candidate budget is exhausted.

V1 scope

  • Treat num_records as the accepted-row target when record selection is enabled.
  • Require an explicit boolean predicate column and hard max_candidate_records bound.
  • Support raise and return_partial exhaustion behavior.
  • Preserve deterministic candidate ordering and trim final overshoot exactly.
  • Track candidate attempts separately from accepted output.
  • Persist candidate-batch completion markers so zero-acceptance batches and resume are correct.
  • Preserve normal DAG generation, processors, profiling, plugins, and model-usage accounting.
  • Run one candidate batch at a time while retaining the scheduler's normal within-row-group concurrency.

Out of scope

  • Concurrent candidate batches.
  • Early cancellation of downstream row work after predicate rejection.
  • Exporting every rejected candidate.
  • Unbounded generation based on an expected acceptance rate.
  • Row-count-changing after-generation processors.

These are optional performance or expansion ideas, not requirements for a complete V1.

Acceptance criteria

  • A single create() call can reliably return exactly X matching rows.
  • Generation is bounded and exhaustion behavior is explicit.
  • Candidate progress is independent of accepted parquet row counts.
  • Runs resume after any committed candidate batch without repeating candidate offsets or accepted output.
  • Zero-acceptance candidate batches remain durably complete.
  • Output processing and profiling operate on accepted rows only, while usage metadata includes rejected work.
  • Default-buffer behavior works when the target is reached only after three or more candidate batches.
  • Documentation explains cost, bounds, partial exhaustion, and resume semantics.

Investigation and context

PR #773 explored workflow-level repetition. Real-run testing showed that extending a static row-group plan can stall when small increments remain inside the same buffer boundary, and that an interrupted append flow can request a target smaller than persisted output. Record selection therefore needs engine-owned candidate progress rather than repeated public create() orchestration.

Detailed plan

The source-of-truth design, architecture diagrams, implementation phases, test plan, and definition of done are in plans/790/engine-native-record-selection.md.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions