During rolling cluster upgrades from etcd 3.5.19 to 3.6.x, a membership data propagation issue occurs where wrong membership data may be propagated to new members during snapshot
  transfers. This results in cluster upgrade failures with the error etcdserver: too many learner members in cluster.

    4.19-4.21

    Consistently reproducible during multi-stage cluster upgrades involving etcd <=3.5.19 and subsequent upgrades to 3.6.x when adding learner members.

  1. Upgrade cluster from a version using etcd 3.5.19 to etcd 3.5.21
  2. Attempt subsequent upgrade to etcd 3.6.4 (or other 3.6.x version)
  3. Add new learner members during the upgrade process
  4. Observe membership data propagation during snapshot transfer     

    Wrong membership data is propagated to new members during snapshot transfers, causing the error etcdserver: too many learner members in cluster and blocking the cluster upgrade from
  completing successfully.

    Membership data should be correctly propagated to new members during snapshot transfers, allowing the cluster upgrade to complete successfully without membership-related errors.

    Upstream issue: https://github.com/etcd-io/etcd/issues/20793