spread

Partitions the data and lays the groups out along an axis, with a gap between them. The workhorse operator for bar charts and small multiples.

from gofish import chart, spread, rect

chart(seafood).flow(spread(by="lake", dir="x")).mark(rect(h="count")).render(
    w=500, h=300, axes=True
)

Signature

spread(*, by=None, dir, **options) -> Operator

Parameters

Parameter	Type	Description
by	str | field(...) | Callable	Field, dotted path, field(...) accessor, or callable to partition by (see path-aware by). Omit to spread per row.
dir	"x" | "y"	Required. Axis to lay groups out along.
spacing	int	Gap between groups in pixels. Ignored when glue=True.
alignment	str	Cross-axis alignment of the groups.
glue	bool	Glue children together: collapse data-driven sizes into a single positional axis at this level. stack sets this.
w, h	int | str	Fixed pixel size, or a field name sizing this operator's own box from data (data-driven operator extent — e.g. a mosaic's column width).
size	int | str | field(...) | list	Per-entry stack-axis extent — a field name, a field(...) accessor, or an explicit list with one value per split entry. size=field("count").normalize() makes the stack axis a space-filling spine. See Space-filling spines.

Returns an Operator for use inside .flow().

Examples

# One bar per lake
chart(seafood).flow(spread(by="lake", dir="x")).mark(rect(h="count"))

# Wider gaps
chart(seafood).flow(spread(by="lake", dir="x", spacing=64)).mark(rect(h="count"))

# Nest spreads for grouped layouts
chart(seafood).flow(
    spread(by="lake", dir="x"),
    spread(by="species", dir="x", spacing=2),
).mark(rect(h="count", fill="species"))

Path-aware by

by accepts a field name, a dotted path string, a callable, or a field(...) accessor:

spread(by="species", dir="x")            # field on a raw record
spread(by="datum.species", dir="x")      # path (e.g. after a selection)
spread(by=lambda r: r.datum.species, dir="x")  # callable escape hatch
spread(by=field("species").sort(), dir="x")    # field(...) accessor

Path strings matter after a ref / selectAll selection: the stream items are then refs, not raw records, so you re-encode by the datum path — by="datum.species".

How a datum.field path resolves (homogeneity collapse)

A ref's .datum is the raw bag of rows that flowed into the node (a list; a fully-split leaf is a 1-row list). A by="datum.field" path does not just read the field off the first row — it projects with homogeneity collapse:

datum.field resolves to a scalar iff every row in the node's bag agrees on that field; otherwise it is None — the "this field is multi-valued here, grouping by it is ill-posed" signal.

This is exactly SQL's ONLY_FULL_GROUP_BY / functional-dependency rule: you may only group by a column that is constant within each row-bag.

Example. After selectAll("bars") where each ref is a lake aggregate of 5 species rows:

group(by="datum.lake")     # resolves — all 5 rows share one lake
group(by="datum.species")  # None — 5 distinct species; ill-posed

To group by a field that is multi-valued in the current bag, disaggregate first (split the bag so each child is homogeneous in that field). A fully-split cell (1 row) trivially collapses, so by="datum.species" works once each node holds a single record.

by vs. channel: an intentional asymmetry

by operates on the selection stream, but a mark's channels operate on the raw record — and they are addressed differently:

Place	Reads	How to write species
by on an operator after a selection	the ref stream	"datum.species" (path)
a channel on a mark, e.g. rect(...)	the raw record	"species" (bare field)

So a ribbon chart reads:

chart(selectAll("bars")) \
    .flow(group(by="datum.species")) \
    .mark(ribbon(opacity=0.8))         # channel → raw record, no prefix

Do not "consistency-refactor" channels into datum.count — a channel like rect(h="count") reads the bound record directly and is never path-prefixed. Only by (on group/spread/stack/scatter) is path-aware, because only by sees the selection stream.

Field-expression pipeline

field(name) returns a chainable accessor — a builder where each method appends one op to an ordered pipeline. It works in two disjoint places:

• As by (a domain slot): .sort(by=None, order=None), .sort(values) (an explicit order list), .reverse(), .bin(thresholds=None), and .drop_nulls() decide which groups exist and in what order.
• As a mark or size channel value (a value slot): .sum(), .mean(), .count(), and .distinct() fold a group's rows to one number, overriding the channel's own default aggregation (sum for size, mean for position).

Mixing the two — an aggregate op on by, or a domain op on a value channel — raises.

Sort a stack's groups by another field's total, instead of data order:

# Bars ordered ascending by their own total `value`
chart(data).flow(
    spread(by=field("category").sort("value"), dir="x")
).mark(rect(h="value"))

Omit the by argument to sort by the group key itself; pass order="desc" for descending.

Sort groups by an explicit order, for a domain-specific sequence no aggregate expresses (severity, calendar order, a fixed ranking):

# Weather categories in a fixed order, not alphabetical or by aggregate
chart(data).flow(
    stack(by=field("weather").sort(["sun", "fog", "drizzle", "rain", "snow"]), dir="y")
)

Groups whose key isn't in the list are appended after, in natural sort order.

Bin a numeric field into groups — a histogram, with no precomputed bins:

# One bar per ~10 auto-computed bins of `age`, height = row count per bin
chart(data).flow(
    spread(by=field("age").bin(), dir="x")
).mark(rect(h=field("age").count()))

Empty bins are dropped, like an ordinary group-by. Pass field("age").bin(5) (a count) or explicit thresholds (a list) to control the binning.

Drop rows with a missing/null grouping field, before grouping:

# Rows with a null/missing "genre" are excluded entirely, instead of
# forming their own "null" group
chart(data).flow(
    treemap(by=field("genre").drop_nulls(), size="worldwideGross")
)

Override a channel's default aggregation:

# The bar height is each species' MEAN weight, not the sum
chart(data).flow(
    spread(by="species", dir="x")
).mark(rect(h=field("weight").mean()))

.count() and .distinct() report measure "count" (they're counts, not the source field's own units) unless you annotate the accessor explicitly: field("id", measure="my-measure").distinct().

Space-filling spines (mosaic / marimekko)

size=field(<name>).normalize() turns a stack's entries into a space-filling spine: each entry's size becomes its SHARE of the window (the operator's own split entries, v_e / Σv_e), so the axis reads as a local 0–100% conditional distribution. It replaces the removed normalize=True layout flag — .normalize() is a data transform on the size channel (a windowed share, computed once up front), not a layout mode, so the same field can drive both a cross-axis marginal size and the normalized fill with no preprocessing.

Nest two stacks on alternating axes for a mosaic: the outer sizes each column by its raw total (size="count" — the marginal), the inner sizes by each entry's share (the conditional):

(
    chart(passengers, axes=True)
    .flow(
        # columns by class — width ∝ each class's count (marginal)
        stack(by="pclass", dir="x", size="count"),
        # survival share within each column (conditional), filling height
        stack(by="survived", dir="y", size=field("count").normalize()),
    )
    .mark(rect(fill="survived", stroke="white", strokeWidth=1))
    .render(w=400, h=300)
)

.normalize() composes to any depth: a third alternating level gives a nested mosaic (class → sex → survived). Because each level's stacking axis is a local self-scaling region and the count is read raw everywhere, the marginal × conditional × … area factorization holds all the way down.

WARNING

Inner conditional axes are local scopes, so they don't yet render 0–1 tick labels — only the outermost marginal axis is labeled.

Notes

• dir is required — spread() raises a ValueError without it.
• Use stack when you want groups touching edge-to-edge with no gap.
• Data order determines group order; sort your data first if order matters.