You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Ensure deterministic graph calculation with consistent layer, node, and edge ordering in Kedro-Viz (#2185)
Resolves#2057.
Based on @astrojuanlu's comment, I realized that Kedro-Viz might be causing the randomness. To investigate, I checked if the inputs to graph calculations—nodes, edges, and layout—were consistent. While the order of nodes and layout remained stable, the order of edges varied with each backend data request. To address this, we now sort the layers, the edges and nodes to ensure consistency.
Update -
As @ravi-kumar-pilla noted, there was an issue with layer ordering: changing a node name could alter the layer order, especially for layers with identical dependencies, like model_input and feature in the example he shared. This issue stemmed from the layer_dependencies dictionary in services/layers.py, where layers with the same dependencies weren’t consistently ordered. To fix this, I added alphabetical sorting for layers with identical dependencies to ensure stability in toposort.
For nodes and edges, I now sort them immediately upon loading from the backend API, ensuring they are consistently ordered in the Redux state.
For testing, I initially considered using Cypress/backend e2e testing with screenshot comparison of the flowchart, but it proved too complex. Instead, I created a new mock dataset, reordered_spaceflights, with reordered nodes and edges. I added tests in normalise-data-test and actions/graph-test. The first test verifies that nodes and edges are consistently sorted by their ids, regardless of backend order. The second test compares the x and y coordinates in the flowchart, confirming that the graph layout is the same between the two mocks.
0 commit comments