Who needs Graphviz when you can build it yourself?

Original Article ↗ Hacker News Discussion ↗

Domain-Specific vs Generic Layout Algorithms

  • Many commenters see iongraph as a strong example of specializing a generic algorithm to a narrow domain (JIT control-flow graphs) to beat long‑mature generic tools like Graphviz.
  • Application-specific design can yield huge performance and quality gains, but often isn’t worth the engineering cost unless the use case is central.
  • A suggested middle ground: detect when input fits “simple/fast” layouts and fall back to heavier, more general algorithms otherwise.

Iongraph’s Approach, Strengths, and Limits

  • Iongraph benefits from domain knowledge: reducible control flow, known nesting depth, explicit loop backedges, and SSA structure, enabling more readable CFG layouts.
  • Commenters like the readability (especially edge routing) and want it generalized into a broader control-flow graph viewer and debugger.
  • Others note that pushing the algorithm to fully general graphs may be hard; its assumptions are tightly tied to compiler IR structures.

Graphviz, dot, and the Tooling Ecosystem

  • Several point out the article conflates Graphviz (framework), dot (language), and dot (one layout engine) — Graphviz has multiple engines (dot, neato, sfdp, etc.).
  • Dot’s text format is widely praised as a portable, simple standard worth using even when rendering with other tools.
  • Alternatives discussed: Mermaid, D2, ELK/elkjs, OGDF, TikZ, nomnoml, and custom stacks combining JS layout libraries with SVG renderers.
  • Opinions diverge on Graphviz quality: some find it “not very good” beyond tiny graphs; others think it’s fine and value its robustness and longevity.

Scaling, Readability, and Diagram Purpose

  • Strong agreement that large graphs quickly become unreadable, regardless of engine; more control over tradeoffs and interactive tools (collapse, search, focus, links between diagrams) are seen as necessary.
  • Some argue diagrams are best for small, well-defined subgraphs and communication, not exhaustive visualization of complex systems.
  • Comparison with UML: one camp sees it as overgeneralized and harmful; another pushes for many small, domain-specific diagram “languages.”

Implementation, Web Stack, and AI

  • The demo’s timeouts are attributed to a convoluted SpiderMonkey-on-WASM stack needed for JIT behavior, not the layout itself.
  • A few see AI as lowering the barrier to building custom tools, but others stress the article mainly showcases domain expertise, not LLMs.
  • Experiments with generative models for layout exist, but reported results are inconsistent and hard to control aesthetically.