reflex-xy

xy figures as first-class Reflex components: WebGL rendering, million-point interactivity, and streaming updates — with chart data riding the app's existing websocket, not a sidecar API.

How it works

• Control plane (Reflex-native). The only chart state is a token string, minted by a @reflex_xy.figure computed var. Semantic events — on_point_hover(event), on_point_click(event), on_select_end(event), and on_view_change(event) — arrive as ordinary Reflex event handlers with small JSON payloads.
• Data plane (xy-native). A second socket.io namespace (/_xy) multiplexed onto the app's own engine.io websocket ships the spec as JSON and every data column as a binary frame (no JSON numbers, no base64, no extra endpoints to reverse-proxy). Pan/zoom/hover round-trips go straight to the figure kernel and never touch Reflex state.
• No figure server. Figures live in a per-process registry as rebuildable caches: the token encodes (client, state, var), so any backend worker can re-run the builder against Reflex state (already distributed via redis in prod) when a reconnect lands on it.

Usage

# rxconfig.py
import reflex as rx
import reflex_xy

config = rx.Config(app_name="dash", plugins=[reflex_xy.XYPlugin()])

# dash/dash.py
import numpy as np
import reflex as rx
import xy
import reflex_xy


class Dash(rx.State):
    points: int = 200_000
    hovered: dict = {}

    @reflex_xy.figure
    def chart(self) -> xy.Chart:
        rng = np.random.default_rng(7)
        xs = rng.normal(size=self.points)
        ys = xs * 0.6 + rng.normal(scale=0.6, size=self.points)
        return xy.scatter_chart(xy.scatter(xs, ys), width="100%", height=460)

    @rx.event
    def on_hover(self, row: dict):
        self.hovered = row


def index() -> rx.Component:
    return rx.vstack(
        reflex_xy.chart(Dash.chart, on_point_hover=Dash.on_hover, height="460px"),
        rx.text(Dash.hovered.to_string()),
        width="100%",
    )


app = rx.App()

Tailwind classes on live charts

Add Reflex's Tailwind plugin when chart chrome uses utility classes. A direct fixed xy.Chart is scanned automatically. A token or Var is resolved only at runtime, so expose its possible complete utility names through the adapter:

# rxconfig.py
config = rx.Config(
    app_name="dash",
    plugins=[
        # Keep dark: utilities synchronized with Reflex's .dark color mode.
        rx.plugins.TailwindV4Plugin(config={"darkMode": "selector"}),
        reflex_xy.XYPlugin(),
    ],
)

# dash.py
LIVE_CHART_CLASSES = [
    "rounded-2xl border border-slate-200 dark:border-slate-800",
    "bg-white text-slate-900 dark:bg-slate-950 dark:text-slate-100",
]

reflex_xy.chart(
    Dash.chart,
    tailwind_classes=LIVE_CHART_CLASSES,
)

The inventory is present only in generated source for Tailwind's build-time scan. It never reaches the DOM; the runtime chart receives the same class names from its XY payload. Use a string or ordered iterable (not a mapping or set), and include every complete class that any state-driven version of the figure can emit. Quotes, escaped arbitrary-selector underscores, and Unicode content are preserved verbatim:

LIVE_CHART_CLASSES = [
    "before:content-['✓']",
    r"[&_[data-xy-slot=legend\_label]]:font-semibold",
]

When a recomputed live figure changes root or slot classes, the browser rebuilds its DOM chrome so the new classes take effect without changing the figure token. Every named-axis range and durable box/range/lasso geometry survives that rebuild without replaying semantic callbacks.

Change points in an event handler and the chart re-publishes itself to every subscriber — the token never changes, so nothing re-renders except pixels.

Builders can be async def (they become reflex AsyncComputedVars, same rule as rx.var) — await a database, HTTP endpoint, or dataframe store:

    @reflex_xy.figure
    async def remote(self) -> xy.Chart:
        rows = await fetch_rows(self.query)
        return xy.line_chart(xy.line(rows.t, rows.value))

Streaming: reflex_xy.append(token, x=[...], y=[...]) from any handler or background task pushes an incremental update over the same socket.

Events and cross-filtering

Live token-backed charts emit versioned, bounded dictionaries with a stable token and canonical row IDs. Selection rows can update ordinary State, which causes dependent @reflex_xy.figure builders to republish without changing their tokens or remounting the chart:

class Dash(rx.State):
    selected_groups: list[str] = []

    @rx.event
    def filter_groups(self, event: dict):
        selection = event["selection"]
        self.selected_groups = [] if selection["cleared"] else sorted({
            row["color_category"] for row in selection["rows"]
        })

def index():
    return rx.grid(
        reflex_xy.chart(Dash.groups, on_select_end=Dash.filter_groups),
        reflex_xy.chart(Dash.filtered_detail),
    )

Fixed-data charts

For a chart that doesn't depend on state, skip the state var entirely — pass the Chart straight in:

def index() -> rx.Component:
    return reflex_xy.chart(
        xy.line_chart(xy.line(t, np.sin(t)), width="100%", height=220),
        height="220px",
    )

That compiles the figure to a content-addressed binary asset at page build and renders it with zero backend involvement — client-side hover, pan/zoom, and density re-bin, same as Figure.to_html() exports; works under reflex export. When a fixed chart still needs kernel round-trips (deep drilldown into millions of points), register it once at module scope instead:

cloud = reflex_xy.inline(xy.scatter_chart(xy.scatter(x, y)))  # module scope

def index() -> rx.Component:
    return reflex_xy.chart(cloud, height="460px")

inline() tokens are content-addressed, so every backend worker derives the same one — no state, no coordination. The escalation path is: direct Chart (static) → inline() (fixed data, live kernel) → @reflex_xy.figure (per-session, state-driven).