From 0623107214d80317d3e2701b26e29ff2cf7d06f0 Mon Sep 17 00:00:00 2001 From: Michal Wozniak Date: Fri, 6 Sep 2024 13:44:42 +0200 Subject: [PATCH] Review remarks --- keps/2724-topology-aware-schedling/README.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/keps/2724-topology-aware-schedling/README.md b/keps/2724-topology-aware-schedling/README.md index 08483e4c2b..2518c9709b 100644 --- a/keps/2724-topology-aware-schedling/README.md +++ b/keps/2724-topology-aware-schedling/README.md @@ -127,8 +127,7 @@ follow naturally we may as well support them in alpha. One complication we noticed for JobSet is that the proposed design assumes injecting the dedicated scheduling gate, called `kueue.x-k8s.io/topology` into -the PodTemplate. However, currently JobSet's `spec.replicatedJob` field is -immutable, even if the JobSet is suspended. This has been relaxed in JobSet 0.6 +the PodTemplate. This required a JobSet fix which is released in 0.6 (see [PR](https://github.com/kubernetes-sigs/jobset/pull/623)). #### Support for the "auto" mode @@ -166,7 +165,8 @@ replacement node. In order to mitigate this risk we propose to extend the waitForPodsReady mechanism with a new timeout, called replacement timeout, which defines the timeout for all pods to be ready again. The time is computed since the last -transition to the PodsReady=false, more details in the KEP PR. This mechanism +transition to the `PodsReady=false`, more details in the +[KEP PR](https://github.com/kubernetes-sigs/kueue/pull/2737). This mechanism will also be helpful for regular workloads. A more involving approach would be to recompute the TopologyAdmission, however,