Large Data and Performance

XY keeps exact canonical columns in Python and chooses a rendered representation for each trace. The goal is to keep transport and draw work bounded by what the viewport can distinguish while retaining exact rows for readout and refinement.

The current defaults begin M4 line decimation above 10,000 rows and automatic scatter density above 200,000 points. Density grids default to 512×384 cells, and very large density traces can build a lazy multiresolution pyramid for viewport queries. These are pre-1.0 policy thresholds, not API guarantees; write code against the behavior and recorded tier, not a hard-coded count.

• density=None chooses automatically and is the normal large-data setting.
• density=True requests a density overview explicitly.
• density=False forces direct points. Above the soft ceiling XY warns about fill-rate and allocation risk, but honors the explicit opt-out.

Density aggregation is computed by the native core and rendered as a compact WebGL texture. Calling it “GPU density” would blur those two stages: the GPU draws the result, while the source-row binning happens before that draw.

Reduction changes visible geometry, not the canonical source store. Hover, selection, pick(), and drilldown resolve source rows when the active tier has an exact mapping. When a dense view cannot support exact per-marker semantics, XY exposes its aggregate/sample representation rather than pretending that a bin is an individual row.

Line decimation preserves extrema for the visible buckets, but it does not claim that every intermediate vertex is drawn. A narrow view is re-decimated against its own window.

There are two different scaling regimes:

1. Ingest, validation, range scans, binning, and decimation inspect source data and therefore retain row-dependent work.
2. The resulting wire payload, WebGL geometry, and static SVG/native-PNG scene are bounded by the chosen viewport representation.

This is why a “cost scales with pixels, not points” slogan needs qualification: it describes the output side after reduction, not the complete data-to-pixels pipeline.

Tier decisions are also recorded in the built payload rather than being silent. Use the memory report for many-chart dashboards and long-running append workloads, and use the committed benchmark harness before publishing a performance comparison.

See the benchmark snapshot for a measured example, or Interactions and selections for exact readout behavior.