/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 * @file Merges one source's overlay parquet files into shuffled output files.
 *
 *   The training sampler can read a whole per-source draw from one row group,
 *   so a source split into per-country files would reach only one country per draw.
 *   A shuffle across the inputs mixes each source's countries within each row group.
 */
import type { PathBuilderLike } from "path-ts";
/**
 * Options for {@link mergeSourceFiles}.
 */
export interface MergeSourceOptions {
    /**
     * The parquet files to merge, read in the order given.
     */
    inputs: readonly PathBuilderLike[];
    /**
     * The output path.
     * Each written file receives a five-digit index in its stem.
     */
    output: PathBuilderLike;
    /**
     * The number of rows held in memory while shuffling.
     * It defaults to the corpus writer's window.
     */
    windowSize?: number;
    /**
     * The shuffle seed.
     *
     * The default matches the corpus writer's seed.
     * This makes rebuilds reproducible.
     */
    seed?: number;
    /**
     * The maximum rows per output file.
     *
     * `writeParquetFile` builds one Arrow table per file.
     * Arrow's list builder overflows on very large tables.
     */
    rowsPerFile?: number;
}
/**
 * The files that {@link mergeSourceFiles} wrote and the row counts they hold.
 */
export interface MergeSourceResult {
    inputs: readonly string[];
    /**
     * The files written, in order.
     */
    outputs: readonly string[];
    rows: number;
    /**
     * Row counts per `country` value.
     */
    byCountry: Record<string, number>;
    /**
     * Row counts per `source` value.
     * A successful merge has one entry.
     */
    bySource: Record<string, number>;
}
/**
 * Merges the inputs through a windowed shuffle into one or more parquet files and reports their contents.
 *
 * @throws When the inputs contain more than one `source`, because a manifest
 * entry records one source per file.
 * The check runs after the stream ends, so files flushed before it remain on disk.
 */
export declare function mergeSourceFiles(options: MergeSourceOptions): Promise<MergeSourceResult>;
//# sourceMappingURL=merge-source.d.ts.map