Charting Roadmap

Current status

Built-in line, scatter, step, clustered bar, box, and real OHLC candlestick charts have shipped, together with extension-side sampling and unified zoom/pan viewport refinement. See Charting for the supported controls and validation rules. This page tracks architectural boundaries and future work rather than promising additional chart types.

Current architecture

The built-in renderer is uPlot-first and optimized for numeric or temporal x data from the kdb results panel. The extension host owns eligibility checks, sorting, grouping, validation, and type-aware data reduction before bounded chart data enters the webview.

Line, scatter, and step use generic selected numeric y series. Bar uses a dedicated clustered canvas path so selected and grouped series render side by side with positive clamped widths and a zero baseline. Box uses selected numeric y series to compute min, q1, median, q3, and max. Candlestick uses explicit, distinct numeric Open, High, Low, and Close roles and a dedicated wick/body canvas path; it is not four line series.

Group by is implemented only for line, scatter, step, and bar. Box and candlestick do not expose it. Every type requires numeric or temporal x; categorical group values are not silently reused as x.

Renderer roles

Tool	Role	Notes
uPlot	Shipped built-in VS Code chart renderer	Keeps the webview dependency small and provides axes, cursor, legend, zoom state, canvas composition, and PNG export. Dedicated hooks draw clustered bars, box summaries, and candlesticks.
ECharts	Prototype fallback only	Consider only if a future verified requirement cannot be implemented safely with the existing uPlot architecture. It is not a current dependency.
Plotly and plotly-resampler	External Python/pandas workflow	Use the opt-in local data server for richer external analysis rather than adding Plotly.js to the built-in webview.

No additional renderer dependency should be added without a concrete requirement, package review, CSP/offline verification, and performance testing against representative kdb results.

Shipped data reduction

The extension guards the number of source rows scanned, derives the initial sample target from plot width, and retains the raw result extension-side.

• Generic line, scatter, and step data uses min/max bucket sampling to preserve local extremes better than uniform stride.
• Bar data is consolidated into complete distinct-x clusters and, when necessary, evenly thinned without dropping individual series from a retained cluster.
• Box data keeps each numeric y series together long enough to compute min, q1, median, q3, and max for each x value or bucket.
• Candlestick data targets roughly one candle per horizontal pixel and uses financial aggregation: first valid open, maximum high, minimum low, and last valid close per x bucket. The four roles are never sampled independently.
• A completed horizontal pan uses the exact unchanged drag-zoom range-loading/resampling decision for the same clamped absolute viewport, including zoom/pan/zoom composition; Refine view applies that current-range path explicitly. Plain drag zooms, Shift+drag grabs the content, buttons and arrow keys pan by 20%, and y remains automatic. Reset invalidates stale refinement responses and restores the original full sample/range without backend I/O.

The webview receives sampled arrays plus source/eligible/sample counts, the algorithm name, and warnings. It does not rescan data on cursor movement.

Current validation boundaries

The current chart contract intentionally stays narrow:

• x must be numeric or temporal.
• Generic y and every OHLC role must be numeric.
• Nested lists, objects, and mixed incompatible columns are rejected.
• Candlestick requires four distinct OHLC columns and rejects retained rows with missing/non-finite values or inconsistent high/low bounds.
• Bar width must come from usable positive x spacing. Compatible duplicate-x rows align into one cluster, multiple finite values for the same generated series and x are rejected, and clusters too dense to distinguish safely are skipped with a status rather than overlapped.
• Category and generated-series caps keep grouped charts bounded.

These rules prevent the renderer from inventing semantics for malformed or ambiguous data.

Future priorities

Responsiveness and cancellation

Continue hardening chart request versioning and cancellation so stale sorting, sampling, zoom refinement, reruns, and panel disposal do not waste extension-host work. Any CPU-heavy additions should yield or move off the critical path where practical.

Sampling quality

Benchmark the shipped type-aware reducers on representative sorted and unsorted kdb results, duplicate timestamps, sparse and dense ranges, gaps, infinities, mixed temporal units, positive/negative bars, and large OHLC data. Consider alternative generic reducers such as LTTB only if measurements show a real improvement without weakening predictable bounds.

Candlestick aggregation must remain financial. A future reducer must preserve first open, maximum high, minimum low, and last close for every bucket.

Visual and accessibility QA

Keep dense-axis label thinning, tooltip precision, keyboard-usable controls, light/dark theme contrast, high-DPI canvas export, and clustered bar/candle width behavior under regression coverage. Visual review should include overlapping timestamps, nearly identical x values, zero/negative bars, flat candles, and large candle counts.

External workflows

The tokenized localhost data server remains the path for pandas, Plotly, plotly-resampler, notebooks, and other richer analysis. Future server work may add bounded metadata or chunking where a concrete external workflow requires it.

Non-goals

• General BI dashboards.
• Invented or placeholder chart types.
• Streaming live updates in the current chart panel.
• Plotly.js as the default built-in dependency.
• Sending complete million-row series to the webview by default.
• Treating categorical, nested, or object data as numeric chart values.