[1.15.4] Raft leader election unexpected behavior

We're encountering an odd case where a Vault leader is being replaced by a single follower that is restarted in a cluster of 3. I'm running vault 1.15.4 running on GKE and deployed by helm version 0.26.1.

The leader was vault-0, I deleted one of the followers vault-1, and once vault-1 pod was created it became the current leader.
I then deleted vault-2 (follower), and pod vault-0 became the leader.
I deleted the current leader vault-0 and the new vault-1. became the leader (expected behavior)

I was expecting a real HA scenario where on steps 1 and 2 (follower deletion), the leader should continue to be the same without triggering the leader election.

Here's my config:

        listener "tcp" {
          tls_disable = false
          address = "0.0.0.0:8200"
          cluster_address = "0.0.0.0:8201"
          http_read_timeout = "600s"
          tls_cert_file = "tls.crt"
          tls_key_file  = "tls.key"
          tls_client_ca_file = "tls.ca"
        }

        telemetry {
          prometheus_retention_time = "12h"
          disable_hostname = true
          enable_hostname_label = true
        }

        seal "gcpckms" {
          project     = "abc"
          key_ring    = "abc"
          crypto_key  = "abc"
        }

        storage "raft" {
          path = "/vault/data"
          retry_join {
            leader_api_addr = "https://vault-0:8200"
            leader_ca_cert_file = "ca.crt"
            leader_client_cert_file = "tls.crt"
            leader_client_key_file = "tls.key"
          }
          retry_join {
            leader_api_addr = "https://vault-1:8200"
            leader_ca_cert_file = "ca.crt"
            leader_client_cert_file = "tls.crt"
            leader_client_key_file = "tls.key"
          }
          retry_join {
            leader_api_addr = "https://vault-2:8200"
            leader_ca_cert_file = "ca.crt"
            leader_client_cert_file = "tls.crt"
            leader_client_key_file = "tls.key"
          }
          performance_multiplier = 1

        }
        service_registration "kubernetes" {}

The autopilot config:

vault operator raft autopilot get-config
Key                                   Value
---                                   -----
Cleanup Dead Servers                  false
Last Contact Threshold                10s
Dead Server Last Contact Threshold    24h0m0s
Server Stabilization Time             10s
Min Quorum                            0
Max Trailing Logs                     1000
Disable Upgrade Migration             false

And probes:

Liveness:   http-get https://:8200/v1/sys/health%3Fstandbyok=true delay=60s timeout=3s period=5s #success=1 #failure=2
Readiness:  http-get https://:8200/v1/sys/health%3Fstandbyok=true&sealedcode=204&uninitcode=204 delay=5s timeout=3s period=5s #success=1 #failure=2

Seems to be related https://github.com/hashicorp/vault/issues/14153

hashicorp / vault

[1.15.4] Raft leader election unexpected behavior #24995