GitHub user CaroFernando added a comment to the discussion: Proposal: Ditch 
Angular and Switch to a new frontend framework

Ahead of the vote, I wanted to add data on two questions raised here: how hard 
it would be to migrate the editor and shared editing (@chenlica), and whether 
React holds up under Texera's real-time load (@Ma77Ball). Code and full results 
are on my fork: 
https://github.com/CaroFernando/texera/tree/proto/frontend-migration-landing/frontend-experiments

**How I tested it:**
- **Code audit:** I went through every frontend file involved in the editor and 
in shared editing and counted what is tied to Angular, to JointJS, or to 
neither.
- **Benchmark:** the same workflow rendered by four canvases:
  - JointJS (what we use today, hosted in React)
  - React Flow
  - Svelte Flow
  - solid-flow
- **Realistic session:**
  - Operators that mirror `joint-ui.service.ts`
  - A Yjs doc shaped like our `SharedModel`, with two co-editors sending 
updates from a Web Worker (as y-websocket would)
  - One execution sending statistics for every operator every 500 ms (our real 
interval)
  - The local user dragging and panning, on workflows of 30 to 300 operators
- **Measurement:**
  - Headless Chrome with CPU throttling to emulate a mid-range laptop
  - Frame times, input blocking and memory
  - CPU profiles split into Yjs, framework, canvas library and rendering
  - Medians of 3 runs

## 1. Migrating shared editing and the editor

- **The core of shared editing is already framework-neutral.**
  - Covered: `SharedModel`, `WorkflowGraph`, `SharedModelChangeHandler`, 
`WorkflowActionService`, co-editor presence, undo/redo and UDF parameter sync.
  - Together that's about 4,000 lines, with ~8 lines of Angular in total 
(`@Injectable` and the like).
  - This code can move to React almost unchanged.
- **The canvas code carries over too if we keep JointJS,** which is what 
@aglinxinyuan's prototype (#9389) does. JointJS is framework-agnostic. 
`joint-ui.service.ts` and the graph wrapper (~2,200 lines) have no Angular, and 
the existing editor tests can verify the port.
- **What has to be rebuilt is the Angular-bound UI around it.** About 2,800 
lines:
  - The editor component and minimap (~2,400 more lines)
  - The collaborative property panel, built on ngx-formly with its own Yjs 
wrapper
  - The code editor and the comment box

  I think the property panel is the real hard part of migrating shared editing, 
more than the canvas or the Yjs layer. It needs a plan of its own.
- **This fits the incremental approach @aglinxinyuan proposed (#9472):** keep 
JointJS, move the Yjs services as plain TypeScript, port the UI component by 
component, and treat any canvas change as a separate effort later.

Write-up: 
https://github.com/CaroFernando/texera/blob/proto/frontend-migration-landing/frontend-experiments/EDITOR-MIGRATION.md

## 2. Performance under real-time load

**Realistic session, mid-range machine (30 s session):**

| | 100 operators | 300 operators |
| --- | --- | --- |
| JointJS as Texera runs it today | 38 fps, input blocked 9.7 s | 10 fps, 
statistics 24 s late |
| JointJS with two fixes | 59 fps, no blocking | 56 fps, 0.5 s blocked |
| React Flow | 60 fps | 60 fps, 5.8 s CPU |
| Svelte Flow | 60 fps | 60 fps, 5.1 s CPU |
| solid-flow | 60 fps | 57 fps, visible stutters, 10.7 s CPU |

- **Real-time collaboration itself is cheap.** Yjs plus the canvas binding is 
under 5% of CPU in every case; redrawing operators is the cost.
- **React's re-rendering was not the bottleneck.** React itself was 0.6 s of 
React Flow's 5.8 s of CPU at 300 operators. At realistic sizes React Flow also 
used less memory than Svelte Flow or solid-flow without culling.
- **Svelte Flow was ~12% cheaper than React Flow,** not something users would 
notice. solid-flow was the weakest.
- **Today's editor slowness comes from how we use JointJS, not from Angular.** 
We use a synchronous paper, and every 500 ms each operator is re-rendered three 
times for its statistics. Two fixes close most of the gap, and they can be 
applied today, independent of the migration:
  - Async rendering with culling
  - Writing statistics in one `attr()` call
  - Plus the `jointjs` → `@joint/core` 4 upgrade @zyratlo and @Xiao-zhen-Liu 
mentioned. Against our code it type-checks after a handful of small changes.

Results: 
https://github.com/CaroFernando/texera/blob/proto/frontend-migration-landing/frontend-experiments/EDITOR-BENCHMARK.md

## 3. Ecosystem for what we build next

To add to @kunwp1's and @zyratlo's point about community size and coding 
agents: it shows most clearly in agent features, which are where Texera is 
heading. Weekly npm downloads:

| | React | Angular |
| --- | --: | --: |
| Framework | 274M | 8.5M |
| Vercel AI SDK UI hooks | 14.3M | 39k |
| CopilotKit | 811k | 23k |
| assistant-ui | 3.5M | no Angular version |

On Angular, more of the streaming chat, tool-call rendering and agent state UI 
would be built by hand.

## Conclusion

The data supports moving to React:

- **The real-time concern doesn't show up as a React cost.**
- **The editor and the Yjs core carry over with low risk** if we keep JointJS 
and migrate incrementally.
- **The ecosystem gap matters most for the agent features we plan to build.**
- **The editor's current performance problem should be fixed now either way,** 
since it isn't caused by Angular.
- **The collaborative property panel is the piece I'd plan most carefully.**

Caveats: one emulated machine profile and a synthetic workflow layout; the 
JointJS fixes were measured in the benchmark, not yet in Texera; the property 
panel and code editor were audited, not prototyped.

GitHub link: 
https://github.com/apache/texera/discussions/9387#discussioncomment-19771518

----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]

Reply via email to