openpencil/packages/docs/guide/comparison.md

293 lines
15 KiB
Markdown
Raw Normal View History

# Open Pencil vs Penpot: Architecture & Performance Comparison
Why compare? OpenPencil exists because closed design platforms control what's possible. Understanding architectural differences shows what an open, local-first alternative can do differently.
::: info Penpot's WASM renderer
Penpot 2.x includes a Rust/Skia WASM renderer (`render-wasm/v1`) that can be enabled via server flags or the `?wasm=true` URL parameter. The old SVG renderer remains the default. This page covers both.
:::
## 1. Scale & Codebase Size
| Metric | Open Pencil | Penpot |
|--------|-------------|--------|
| Total LOC | **~26,000** | **~299,000** |
| Source files | ~143 | ~2,900 |
| Languages | TypeScript, Vue | Clojure, ClojureScript, Rust, JS, SQL, SCSS |
| Rendering engine | ~3,200 LOC (TS, 10 files) | 22,000 LOC (Rust/Skia WASM) |
| UI code | ~4,500 LOC | ~175,000 LOC (CLJS + SCSS) |
| Backend | None (local-first) | 32,600 LOC + 151 SQL files |
| LOC ratio | **1×** | **~11×** |
Open Pencil is **~11× smaller** — and that's the whole point. It's not a simplification; it's a fundamentally different architecture.
## 2. Architecture
### Open Pencil: Monolithic Client
```
┌─────────────────────────────────┐
│ Tauri (native shell) │
│ ┌───────────────────────────┐ │
│ │ Vue 3 + TypeScript │ │
│ │ ┌─────────┐ ┌──────────┐│ │
│ │ │ Editor │ │ Kiwi ││ │
│ │ │ Store │ │ Codec ││ │
│ │ └────┬─────┘ └──────────┘│ │
│ │ │ │ │
│ │ ┌────▼────────────────┐ │ │
│ │ │ Scene Graph (TS) │ │ │
│ │ │ Map<string, Node> │ │ │
│ │ └────┬────────────────┘ │ │
│ │ │ │ │
│ │ ┌────▼────┐ ┌──────────┐│ │
│ │ │ Skia │ │ Yoga ││ │
│ │ │CanvasKit│ │ Layout ││ │
│ │ │ (WASM) │ │ (WASM) ││ │
│ │ └─────────┘ └──────────┘│ │
│ └───────────────────────────┘ │
└─────────────────────────────────┘
```
**Everything in one process.** No server, no database, no Docker. The scene graph is a flat `Map<string, SceneNode>` in TypeScript. Rendering calls Skia CanvasKit directly from TS. Layout is Yoga WASM called synchronously.
### Penpot: Distributed Client-Server
```
┌───────────────────────────────────────────────────────┐
│ Docker Compose │
│ ┌──────────────┐ ┌─────────────┐ ┌──────────────┐ │
│ │ Frontend │ │ Backend │ │ Exporter │ │
│ │ ClojureScript│ │ Clojure │ │ (Chromium) │ │
│ │ shadow-cljs │ │ JVM │ │ │ │
│ │ ┌─────────┐ │ │ ┌────────┐ │ └──────────────┘ │
│ │ │render- │ │ │ │Postgres│ │ │
│ │ │wasm │ │ │ │Valkey │ │ ┌──────────────┐ │
│ │ │(Rust→ │ │ │ │ MinIO │ │ │ MCP │ │
│ │ │ Skia │ │ │ └────────┘ │ │ Server │ │
│ │ │ WASM) │ │ │ │ └──────────────┘ │
│ │ └─────────┘ │ │ │ │
│ └──────────────┘ └─────────────┘ │
└───────────────────────────────────────────────────────┘
```
**5+ services minimum.** PostgreSQL for persistence, Redis (Valkey) for pub/sub and caching, MinIO for asset storage, a JVM backend, a Node.js exporter (headless Chromium for server-side rendering), plus the ClojureScript frontend. Dev setup requires Docker Compose with custom networking.
### Verdict: Architecture
Open Pencil's single-process architecture eliminates:
- Network latency between frontend and backend
- Serialization/deserialization overhead at service boundaries
- Container orchestration complexity
- Database query overhead for every operation
Penpot's architecture is optimized for **multi-user server-hosted deployments**. Open Pencil is optimized for **instant local performance**.
## 3. Rendering Pipeline
### Open Pencil: TS → CanvasKit WASM (direct)
```typescript
// renderer.ts — direct CanvasKit calls from TypeScript
renderSceneToCanvas(canvas, graph, pageId) {
// Iterate nodes, build Skia paths/paints, draw
this.fillPaint.setColor(...)
canvas.drawRRect(rrect, this.fillPaint)
}
```
- **1 boundary crossing:** TS → WASM (CanvasKit)
- Scene graph lives in JS heap — no serialization to render
- ~3,200 LOC renderer (split into 10 focused files: scene, overlays, fills, strokes, shapes, effects, rulers, labels)
### Penpot: JS (compiled from CLJS) → Rust WASM → Skia
Penpot 2.x includes a Rust/Skia WASM renderer (`render-wasm/v1`), opt-in via server flags or `?wasm=true`. When enabled, shapes are rendered through:
```
ClojureScript (compiled to JS)
→ decompose to primitives + binary-pack into WASM linear memory
→ Rust WASM (via Emscripten C FFI)
→ skia-safe (Rust Skia bindings)
→ Skia (WebGL)
```
When disabled (default), shapes render as an SVG DOM tree via React/Reagent — each shape is a DOM element.
- **1 boundary crossing** (JS → WASM), same as Open Pencil — but with explicit serialization overhead: UUIDs split to 4×u32, transforms to 6×f32, fills/strokes binary-packed, base props batched into a 104-byte struct per shape
- Tile-based rendering system with interest areas
- 11 separate render surfaces (fills, strokes, shadows, etc.)
- Global mutable state via `unsafe { STATE.as_mut() }` pattern
- 22,000 LOC Rust render engine
Penpot's tile system (`TileViewbox`, `TileTextureCache`, `TILE_SIZE_MULTIPLIER`) pre-renders tiles around the viewport and caches textures (up to 1024 entries).
Open Pencil re-renders the full viewport every frame because CanvasKit called directly from TS is fast enough to not need tiling.
### Verdict: Rendering
| Aspect | Open Pencil | Penpot |
|--------|-------------|--------|
| JS→WASM boundary | Direct (TS objects) | Binary-packed (104-byte base props struct) |
| Rendering model | Immediate/full redraw | Tile-cached |
| Surface management | 1 surface | 11 surfaces |
| Memory overhead | Low (no tile cache) | High (1024 tile cache) |
| Code complexity | ~3,200 LOC (10 files) | 22,000 LOC |
| Unsafe code | None | `unsafe` global state |
When Penpot's WASM renderer is enabled, both projects use Skia via JS→WASM. Open Pencil calls CanvasKit directly with TS objects. Penpot decomposes ClojureScript data into binary-packed structs, writes them to WASM linear memory, and renders through a 22,000 LOC Rust engine. When WASM is disabled (default), Penpot renders shapes as an SVG DOM tree. For small-to-medium documents, the direct CanvasKit path is faster. Penpot's tile system may win on extremely large canvases (100K+ shapes) where only a small viewport is visible — but the overhead is significant.
## 4. Scene Graph & Data Model
### Open Pencil
```typescript
// Flat map, O(1) lookup
nodes: Map<string, SceneNode>
// 29 node types from Figma's Kiwi schema
// ~390 fields per NodeChange (Figma-compatible)
```
- TypeScript interfaces with strict types
- GUIDs match Figma's `sessionID:localID` format
- Direct property access — no indirection layers
### Penpot
```clojure
;; 20+ type definition files in common/src/app/common/types/
;; shapes_builder.cljc, shapes_helpers.cljc
;; Separate type systems for: color, component, container, fills,
;; grid, modifiers, objects_map, page, path, etc.
```
- Data spread across `common/` (49,600 LOC of .cljc)
- Separate geometry modules for flex layout (~6 files), grid layout (~5 files), constraints, bounds, corners, effects
- Runtime schema validation (Malli)
- Data must cross CLJS→Rust boundary for rendering
### Verdict: Data Model
Open Pencil reuses Figma's proven schema (194 Kiwi definitions) directly in TypeScript — zero translation. Penpot maintains its own type system across Clojure/ClojureScript/Rust, requiring manual sync between all three.
## 5. Layout Engine
### Open Pencil: Yoga WASM (314 LOC)
```typescript
import Yoga from 'yoga-layout'
// Direct mapping: Figma stack* fields → Yoga flex properties
const root = Yoga.Node.create()
root.setFlexDirection(FlexDirection.Row)
root.calculateLayout()
applyYogaLayout(graph, frame, yogaRoot)
```
314 lines total. Synchronous, in-process.
### Penpot: Dual Implementation
1. **ClojureScript** (common): `flex_layout/` (6 files), `grid_layout/` (5+ files) — custom implementations
2. **Rust WASM**: `flex_layout.rs` (741 LOC), `grid_layout.rs` (843 LOC) — reimplemented from scratch
Penpot maintains **two independent layout engines** (CLJS and Rust) that must produce identical results.
### Verdict: Layout
Open Pencil delegates to a battle-tested library (Yoga, used by React Native on billions of devices) in 314 lines. Penpot maintains ~3,000+ LOC of custom layout code duplicated across two languages.
## 6. File Format & Figma Compatibility
### Open Pencil
- **Native Kiwi binary format** — same serialization as Figma uses internally
- Direct `.fig` file import via extracted Kiwi codec (2,178 LOC schema + 551 LOC codec)
- Figma clipboard paste support (reads Figma's Kiwi binary from the clipboard)
- Wire-compatible with Figma's multiplayer protocol
### Penpot
- **ZIP archive** (`.penpot` files) containing JSON manifests, per-file JSON data, binary assets, and thumbnails (v3 format)
- SVG used for default rendering and export (opt-in WASM renderer available)
- No native `.fig` import
- Three format versions (v1 legacy Transit, v2, v3 JSON-in-ZIP) with migration system
### Verdict: File Format
Open Pencil has a significant advantage — it can read Figma files natively and even paste Figma clipboard data. Penpot requires manual export/import and cannot open `.fig` files.
## 7. State Management & Undo
### Open Pencil
```typescript
// 110 LOC — inverse command pattern
class UndoManager {
apply(entry: UndoEntry) { entry.forward(); this.undoStack.push(entry) }
undo() { entry.inverse(); this.redoStack.push(entry) }
}
```
110 lines. Forward/inverse closures that capture minimal state. Batch support for multi-step operations.
### Penpot
State management uses Potok (a Redux-like library for ClojureScript atoms). Events implement `UpdateEvent` (pure state→state) or `WatchEvent` (side effects via RxJS). Undo stores inverse change vectors (max 50 entries), with transactions to group rapid changes and auto-expiry after 20 seconds.
### Verdict: State
Open Pencil's approach is simpler and lower overhead. Penpot's approach is more suitable for collaboration (changes are serializable), but at the cost of complexity.
## 8. Developer Experience
| Metric | Open Pencil | Penpot |
|--------|-------------|--------|
| Dev setup | `bun install && bun dev` | Docker Compose + JVM + Node + Rust toolchain |
| Hot reload | Vite HMR (~50ms) | shadow-cljs (seconds) |
| Type checking | TypeScript (strict) | Runtime (Malli schemas) |
| Build time | <5s (Vite) | Minutes (JVM startup + CLJS compile + Rust WASM) |
| First contribution barrier | Low (TS/Vue) | High (Clojure + Rust + Docker) |
| Desktop | Tauri v2 (~5MB) | N/A (browser-only) |
| Hiring pool | Massive (TS/Vue devs) | Tiny (ClojureScript + Rust) |
## 9. Performance Characteristics
| Scenario | Open Pencil | Penpot |
|----------|-------------|--------|
| Cold start | <2s (WASM load) | 10s+ (server + client + WASM) |
| Operation latency | <1ms (in-process) | 10-50ms (network round-trip) |
| Render frame | Direct Skia call | CLJS→JS→WASM FFI→Skia |
| Memory baseline | ~50MB (browser tab) | ~300MB+ (JVM + Postgres + Valkey + browser) |
| Offline capability | Full (local-first) | None (server-dependent) |
| 10K shapes render | One pass, no caching | Tile-based with 11 surfaces |
## 10. What Penpot Does Better
1. **Server-side collaboration** — centralized multi-user editing with WebSockets, user accounts, and access control (Open Pencil uses P2P via Trystero + Yjs — no server, but also no access control or persistence beyond the session)
2. **PDF export** — headless Chromium export service for PDF rendering (OpenPencil exports SVG but not PDF yet)
3. **Plugin system** — full plugin API with sandboxed execution
4. **Design tokens** — native design token support
5. **CSS Grid layout** — custom implementation (Open Pencil waiting for Yoga Grid)
6. **Self-hosting** — Docker-based deployment for teams
7. **Maturity** — years of production usage, battle-tested at scale
## 11. Scripting & Extensibility
OpenPencil ships with an [`eval` command](/eval-command) that provides a Figma-compatible Plugin API for headless scripting — batch operations, automated testing, and AI-driven modifications all run without the GUI. On top of that, **78 AI tools** are available via built-in chat, MCP server (stdio + HTTP), and the CLI — covering read, create, modify, structure, variables, vector path, analyze (color/typography/spacing/clusters), diff, boolean operations, and arrangement. Penpot has a plugin system with sandboxed execution but no headless scripting API or MCP integration.
## Summary
| Dimension | Winner | Why |
|-----------|--------|-----|
| **Architecture simplicity** | Open Pencil | Single process vs 5+ services |
| **Rendering performance** | Open Pencil | Direct CanvasKit vs SVG DOM (default) or binary-packed WASM |
| **Code maintainability** | Open Pencil | ~26K LOC in 1 language vs 299K in 4+ languages |
| **Figma compatibility** | Open Pencil | Native Kiwi codec vs no .fig support |
| **Developer onboarding** | Open Pencil | TS/Vue vs Clojure/Rust/Docker |
| **Desktop experience** | Open Pencil | Tauri native vs browser-only |
| **Layout engine** | Open Pencil | Yoga (proven) vs custom dual implementation |
| **Collaboration** | Tie | Penpot: server-based with access control; Open Pencil: P2P via Trystero + Yjs, zero hosting |
| **Self-hosting** | Penpot | Docker-ready vs desktop-only |
| **Ecosystem maturity** | Penpot | Years of production vs early stage |
Open Pencil is architecturally leaner — a single-process CanvasKit renderer in ~26K LOC of TypeScript, Figma-compatible by design. Penpot is a full-stack platform with ~299K LOC across Clojure, ClojureScript, Rust, and SCSS, plus a Docker service fleet. Both now offer real-time collaboration (different architectures: P2P vs server). Penpot has a plugin ecosystem and server-side PDF export; Open Pencil has Figma-compatible headless scripting, **78 AI/MCP tools**, SVG export, and a native desktop app.