Kubeflow Pipelines experiment run pagination default pageSize ordering gap
0 reputation · 22 Oct 2020, 18:00 UTC
Kubeflow Pipelines experiment run listing now enforces `pageToken`-bounded pagination for datasets exceeding 100 results, phasing out the default-unrestricted `pageSize` across compatible client versions.
However, the default `pageSize` provides no guaranteed ordering, which risks duplicate or missed runs during concurrent modifications.
Across Kubeflow sub-projects, pagination primitives are inconsistent: Katib trial listing, Notebook file browsing, and Pipelines run traversal each employ distinct token or offset strategies, preventing a unified dataset traversal approach.
No cursor-based pagination alternative is officially provided, forcing implementers to rely on offset arithmetic that degrades performance and reliability beyond 10,000 item datasets. What ordering mechanism, if any, does the `pageToken` enforce across consecutive `listRuns` calls? How do disparate Kubeflow sub-project pagination implementations differ in token lifecycle and result consistency? Is there a deployment‑specific `pageSize` maximum that should be verified before scaling traversal beyond 10,000 runs?
kubeflow