yangzhang75 opened a new issue, #8580:
URL: https://github.com/apache/texera/issues/8580

   ### Task Summary
   
   Switching between a workflow's two views -- the operator canvas 
(`/user/workflow/:id`) and the Form View (`/user/workflow/:id/form`) -- is a 
full browser page load in both directions (`window.location.href` in 
`MenuComponent.openFormViewPage` and `WorkflowFormComponent.openCanvasPage`). 
The two views are views of one open workflow, so everything that makes the 
workflow live is thrown away on the way out and rebuilt on the other side:
   
   * both `ngOnDestroy`s call `clearWorkflow()` (which destroys the Yjs shared 
document and leaves its co-editing room), 
`computingUnitStatusService.disconnect()`, `resetExecutionAndWorkers()`, and 
clear the console and results;
   * the arriving view then re-bootstraps Angular (`APP_INITIALIZER` blocks on 
`/api/config`), re-parses the whole bundle, re-fetches operator metadata that a 
`providedIn: "root"` singleton had already cached for the page, fetches the 
workflow, opens a new shared document and room, reconnects the computing unit, 
and waits for the backend to report the execution state again.
   
   The visible cost is seconds of blank page on every switch, and a run that 
was already going takes that long to say so again. Nothing about it is 
necessary: `WorkflowWebsocketService`, `ComputingUnitStatusService`, 
`WorkflowActionService` and the joint graph it owns are all root-provided and 
survive a route change.
   
   Routing between the two was tried during the Form View work and reverted; 
the code comments record three symptoms, two of which have identified causes:
   
   1. **A ghost coeditor of yourself.** The arriving view built a second shared 
document for the same workflow, so the page joined the room it was already in 
and saw its own other client.
   2. **Undraggable operators.** A JointJS paper binds to the joint graph, 
which is root-provided and outlives the component that created it, and no paper 
is ever disposed -- neither `WorkflowEditorComponent`'s nor 
`MiniMapComponent`'s. Every mount left another paper listening to that graph 
from a detached DOM node, and whichever paper answered a pointer event decided 
whether an operator could be dragged. Harmless while every mount followed a 
page load.
   3. **Broken runs.** No independent cause found; likely a consequence of the 
first two, but this is the part that most needs checking in a browser.
   
   Proposed: route in both directions, have the departing view hand the session 
to the arriving one instead of tearing it down, have the arriving view attach 
to what it was given instead of rebuilding it, and dispose both papers on 
destroy. Leaving the workspace altogether keeps today's teardown, including on 
unload.
   
   Follow-up to the Form View feature (parent issue #8011); the switch itself 
landed in #8456.
   
   ### Task Type
   
   - [x] Performance
   - [x] Refactor / Cleanup
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to