Page tree
Skip to end of metadata
Go to start of metadata

You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 9 Next »

Abstract

As of 5.0-alpha3, trees and lists are hardly usable, refreshing all the time, jumping up and down... So far this has even prevented us from completely dropping the old adminInterface module, because it is still much more usable, responsive. The goal of this concept is to analyze and highlight the performance hiccups in the magnolia trees and lists.

 

1. Rows with different heights

  • Cause: This is due to the activation-status icon having a bigger font-size, introducing a bigger row-height.
  • Effect: This is then likely to produce scroll jumps when the table is repainted, because this is based on calculated row height.
  • (tick) Fix: CSS only.

2. Tables set 'immediate'

  • Cause: What immediate means for Vaadin components is that value changes are immediately propagated to server.
  • Effect: For tables, the value is the selection, so every time one presses an arrow key and changes selection, we get a server roundtrip, seemingly repainting the table.
  • (warning) However, setting tables non-immediate would no longer update location fragment in the URL or action bar preview as soon as selection changes, but in the next roundtrip.
  • Fix: Set table non-immediate, and implement an asynchronous scheduled server-side update for selection value changes, so that it does not block client-side selection.

3. Lazy loading in trees

  • Effect: When getting down a big amount of nodes, at some point selection cannot keep up.
  • Cause: That is because only a subset of rows (visible row count * cache ratio) has effectively been loaded (row-based lazy-loading).
  • (info) We don't need row-based lazy loading in trees.
    • There is already depth-based lazy-loading provided by the hierarchical container.
      • This ensures child nodes are loaded only when parent is expanded.
      • Chance is minimal to run into huge loading times when expanding.
        • JCR itself is not optimized for big flat data structures.
        • It was the same in old admin tree.
  • (warning) How do I turn it off?
    • Row-based lazy loading is provided by the Table component.
    • Documentation tells to use setPageLength(0) to disable paging...
      • ... until client-side updates it automatically!
  • Fix: Patch the Vaadin Table class to always use the getPageLength() getter from the class' internal code (no direct field reference).
    • This enables our TreeTable to override getPageLength() and return 0 all the time.
  • (info) Some lazy loading might still occur when collapsing nodes down the hierarchy, but that's acceptable, given how tightly lazy loading is integrated within the Table component.
  • (info) There will be a new Table component in a future Vaadin version, but not by the time we ship 5.0.

4. Partial updates

  • Cause: Vaadin TreeTable component normally supports partial updates if container is an ItemSetChangeNotifier (not sure why that condition).
    • But this is currently disabled with // FIXME: This disables partial updates until TreeTable is fixed so it does not change component hierarchy during paint
      • Forcing partial updates seemed to work basically fine.
      • Vaadin commit said: Currently TreeTable changes its child components during paint, which is too late for AbstractCommunicationManager to take into account
      • see http://dev.vaadin.com/ticket/8628
  • Effect: The whole table is repainted when expanding one node.
    • This is not the most harmful factor for user interaction, but it surely has its share of impact when many nodes are expanded.
  • (info) Inplace editing is not yet using partial updates, although that was also done in the standalone PoC (ticket already exists).
  • Fix: Patch TreeTable#setContainerDataSource() method to set containerSupportsPartialUpdates = true.

5. Column expand ratios

  • Effect: When comparing M5's config tree with a Vaadin prototype using the same JCR containers, there was still quite a big latency in the config tree when expanding a node.
    • (error) Column widths are actually updated for all the table, when expanding/collapsing nodes!
      • Firebug clearly shows that table cells' width value is bouncing several times within 1-2 pixels of its original value; this is a very expensive CPU operation.
  • Cause n°1: We set column's expand ratio systematically, even when it's not configured.
  • Cause n°2: Table seems to send visible columns and their expand ratio even if values have not changed.
  • Fix: First, do not set expand ratio systematically when building columns.
    • Maybe also patch TreeTable class so that it not send visible columns and expand ratios again to client if they have not changed.

6. Scroll position

  • Effect: The tree might still jump slightly when expanding/collapsing nodes, just for the time it repaints.
  • Cause: It is likely that there is some scroll position hack in connectors that we should assert how useful it is.
    • FOUND! line 138 in TableConnector: getWidget().updateFirstVisibleAndScrollIfNeeded(uidl);
    • It is documented in VScrollTable as /** For internal use only. May be removed or replaced in the future. */
      • as well as all other methods originating from legacy updateFromUIDL mechanism.
  • Fix: Patch TableConnector to remove the call or VScrollTable to remove the body of that method.

 

Disculped

  • It is interesting to note that our JCR containers are fine; they are not causing performance flaws.
    • Some methods however should be left unsupported since the container should be read-only, in other words, persistance is not managed by the container.
    • There is interesting insight on this in the Vaadin book section about SQL container.
  • Tried without config app's inplace editing (uses a FieldFactory) => no significant improvement
  • Tried without our own MagnoliaTable and MagnoliaTreeTable => no significant improvement
  • Tried without CellStyleGenerator => no significant improvement
  • Tried without our client-side patches that force width and height in the post-layout phase => no significant improvement

 

  • No labels