Run comparison

The run comparison tool compares outputs of model runs side by side: pick a model, add historical runs (or raw workspace tables) to the comparison set, choose what to compare, and get a chart preview. The tool is a preliminary data selection with a quick preview — complex analysis belongs in the Datagrok workspace, where any comparison can be exported as a snapshot: the data table plus the chart with its current options.

Matching

Scalars and dataframe columns are matched across runs by name: exact, then normalized (case/underscore/dash-insensitive), then fuzzy. Values with mismatching units never match; partially missing units produce a warning icon.

When runs name the same quantity differently (for example, pressure and P), map the names manually: the exchange icon next to the value filter opens the Name mappings dialog with name pairs. Mapped names match as the same value across all runs, index columns included. The mapping is exact up to case and separators — no fuzzy matching — and pairs chain (mapping A to B and B to C makes all three equal).

Columns additionally require their tables to be comparable: the index columns must name-match. Beyond that, matching stays per column — the tables don't need to share their full column sets. Split columns don't affect matching: each run charts with its own split column even when the names differ, and runs without one chart as a single series in an unnamed split category.

Matched values whose data is the same in every run (equal scalar values, or equal value, index, and split column contents) are hidden from the compare list by default. Numbers count as equal when they differ by less than 0.1%, and datetime index values only when they match exactly. Turn off Hide equal values to show them.

Fine-tuning matches

Each compare row maps the same value across runs. Several items of one run can match the row, but only one of them is active at a time: the active item is the run's charted series. To chart several series within one run, use a split column instead — that is the native pattern for model results.

When an input and an output of the same step share a caption, the parameter name is shown after it, for example Data (result), so the two rows stay distinguishable.

Automatic matching picks the best candidate per run, which is not always the desired one — for example, when a run has several similar tables. Expand a compare row (the chevron on column targets) to see every compatible item grouped by run, with checkboxes:

• items picked by the automatic matching are checked and marked (auto);
• other compatible items are listed unchecked — checking one makes it the run's series and unchecks the run's previous pick (one item per run);
• unchecking an item excludes the run from the chart;
• raw workspace tables are checked in every target whose name matches exactly (up to case and separators), not just the best one; fuzzy name matches are listed unchecked;
• Reset to auto discards the manual toggles for that target.

Toggles survive index and split changes, and the coverage count next to each target reflects the checked items.

Index and split columns

Column comparison requires an index (x axis) column per table. An optional split column marks a categorical column whose values are separate series within one run.

• Numeric (int, bigint, float), datetime, and string columns are offered as indexes.
• Split columns must be string-typed and different from the index.
• With Merge same functions (on by default), tables produced by the same function collapse into one picker row; a selection applies to all runs.
• Charts are built by concatenating raw rows of all runs into one long dataframe; the line chart splits by run and, when set, by the split column. There is no row alignment or tolerance logic — differing grids simply chart as separate lines.

Row semantics

When a table's index column is numeric or datetime, a selector appears next to its pickers saying what the table's rows mean. The default is series: rows form an indexed sequence, and values chart as lines connected along the index. A value charts in a non-default mode only when every participating table has that mode set. Until then it charts as a plain series, and the runs whose tables lack the setting are flagged with an amber chip under Results.

Relative timeseries

The relative timeseries mode charts the table on an elapsed-time axis, which lets runs with different time representations — a datetime column in a workflow dataframe, plain float seconds in a CSV — share one line chart.

• Numeric indexes get a units selector (ms, s, min, h, days) saying what the numbers mean. It is prefilled from the column's units metadata when recognizable. Datetime indexes need no units.
• Runs are aligned automatically: each table's earliest time point is treated as 0, so runs overlay for shape comparison regardless of when they were recorded.
• The chart axis is a float column labeled with its unit, for example time (s). The unit comes from the first participating numeric index; a comparison of only datetime indexes picks a readable unit from the data range automatically.

Independent points

The independent points mode treats rows as unordered observations and charts them as unconnected (index, value) points instead of a line. Row order and grid alignment stop mattering, so this mode suits unordered data — optimization populations, sampled or stochastic results, per-item tables. And since any numeric column can serve as the index, it also gives value-vs-value views: pick one output quantity as the index and compare it against another across runs.

• Points are colored by the split column when one is set, otherwise by run. When the split takes the color, runs are distinguished by marker shapes.
• With multiple values selected, all values chart on one scatterplot sharing the y axis, colored by value name with runs as marker shapes. Whether the values are comparable on one scale (units, magnitude) is up to you.

Multiple values

When the selected value has compatible siblings, the Multiple values toggle (or Shift+click on a compare row) enables selecting several values at once. Unchecking the last selected value turns the mode off, keeping that value as the regular selection. Every scalar is compatible with every other scalar. Column values are compatible when they share at least one run from the same tables and their index is numeric or datetime everywhere. Scalars and columns never mix in one selection.

Several columns chart as stacked line-chart panels with runs (and split categories) as lines inside. In the independent points mode, the values share one scatterplot instead, colored by value name. A comparison of several values needs a numeric or datetime index — if an index moves to a string column, the chart is replaced with a hint until the index is changed back or a single value is selected.

Several scalars chart on a radar chart: one axis per value, one polygon per run. A switch above the chart flips it to a parallel-coordinates plot. With two values the tool always uses the parallel-coordinates plot, because a two-axis radar degenerates to a line. The radar viewer comes with the Charts package — without it, the parallel-coordinates plot is used for any number of values.

Editing matches never exits the mode: a run that is missing or re-sourced from another table in one of the selected values shows as a gap in that panel, and the value's row is marked partial. Reverting the edit restores the chart exactly.

Default index annotations

A model can declare how its dataframe output should be indexed in a comparison with the comparison option — a small JSON object on the output:

//output: dataframe result {comparison: {"index": "time", "split": "species", "mode": "timeseries", "units": "s"}}

Key	Value	Pre-fills
index	column name	the index (x axis) picker
split	column name	the split picker
mode	series, timeseries, or points	the row semantics selector
units	ms, s, min, h, or days	the time units selector

All keys are optional. When a run is added to the comparison:

• index pre-fills the index picker, and the column is offered even when its type is behind a toggle (an explicit annotation outranks the type gating);
• split pre-fills the split picker, provided the column is a valid split (string-typed, not the index);
• mode pre-fills the row semantics selector. units applies only together with mode: timeseries on a numeric index and outranks the column's units metadata;
• annotations never overwrite a selection the user has already made; a missing column name or an unknown mode or units value is silently ignored;
• merged rows work naturally: all merged tables come from the same function, so they carry the same annotation and agree on the default.

Raw workspace tables have no annotations.

Exporting

Results > Export... asks for a snapshot name and offers two ways out:

• Open in workspace adds the comparison to the workspace: the data table plus the chart with its current options. Tweak the chart before exporting — the tweaks carry over.
• Save & share opens the platform project dialog on the same snapshot, so the comparison can be saved as a project and shared without leaving the tool. On platforms without the project dialog, only Open in workspace is offered.