From af2bfd8a632335eb8cccd1f4f113a8b940a9f450 Mon Sep 17 00:00:00 2001 From: Erik Sundell Date: Wed, 27 Mar 2024 14:20:09 +0100 Subject: [PATCH] basehub: rely on a single user-scheduler replica --- helm-charts/basehub/values.yaml | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/helm-charts/basehub/values.yaml b/helm-charts/basehub/values.yaml index 362991ca5d..d0d8ae2862 100644 --- a/helm-charts/basehub/values.yaml +++ b/helm-charts/basehub/values.yaml @@ -179,6 +179,15 @@ jupyterhub: replicas: 0 userScheduler: enabled: false + # replicas default value is 2 in z2jh, but its an optimization that seldom + # is needed so let's settle for 1 to preserve max number of pods available + # per node on EKS clusters where they are somewhat limited. Only one + # replica is actively doing work to schedule pods, so at worst during + # disruptions of pods we end up waiting longer for a replica to reschedule + # on a new core node than it took for a non-leader replica to become + # leader. If that takes more time, it could could delay scheduling a newly + # created user server assuming the hub pod wasn't also disrupted. + replicas: 1 # FIXME: We should think about these resource requests/limits, see # https://github.com/2i2c-org/infrastructure/issues/2127. #