Modes de cohérence du cluster

Verify your mode#

# Prometheus metric: always 1 with the mode label.
curl -s http://127.0.0.1:9000/metrics | grep bouine_cluster_mode_info

# Admin API: /v1/cluster/peers returns all members in every mode.
curl -s http://127.0.0.1:9000/v1/cluster/peers \
  -H "Authorization: Bearer ${BOUINE_ADMIN_TOKEN}"

# CLI
bouine cluster peers --server 127.0.0.1:9000 --token "${BOUINE_ADMIN_TOKEN}"

Expected: every node reports the same mode label. If they disagree, the cluster has a configuration drift — every pod must use the same mode.

Per-mode expectations#

strong (default)#

When things go wrong:

• Purge didn’t propagate. Check bouine_cluster_invalidations_http_total. If zero, the admin port may be unreachable. Verify cluster.tls config and network policies. The gossip broadcast queue provides a secondary delivery path (check bouine_cluster_invalidations_gossip_total).

• Peer fetch is slow or failing. Check bouine_peer_fetch_duration_seconds (should be < 2 ms on LAN). If elevated, check cluster network health (kubectl get endpoints bouine-headless). Increase hop_limit if node churn is high.

eventual#

When things go wrong:

• Purge didn’t propagate within 5 seconds. Check gossip membership:

  curl -s http://127.0.0.1:9000/v1/cluster/peers

  Should show 3+ nodes on every pod. If a node is missing, check its cluster.join list and DNS resolution. The headless Service must have publishNotReadyAddresses: true.

• Stale reads on one node. That node may be partitioned. Check bouine_cluster_invalidations_gossip_total — if the counter hasn’t incremented recently, the gossip link is broken. Restart the node.

• Hit rate lower than expected. Each node cold-starts independently in eventual mode. Over time, hit rate naturally plateaus as each node fills its cache from origin traffic. If load is unevenly distributed across nodes (e.g. session affinity), some nodes may have much lower hit rates.

Memory and bandwidth budget#

strong mode#

• Memory per node: working set ÷ N (where N = cluster size).
• Bandwidth: minimal. Peer-fetch RPCs are small (key hashes, HTTP headers). HTTP fan-out is infrequent (only on invalidation).

eventual mode#

• Memory per node: 1–N× depending on traffic overlap. With round-robin load balancing, expect ~1× (each node caches ~1/N of the working set).
• Bandwidth: minimal. Gossip invalidations only.

Switching modes#

Mode changes require a full cluster restart. The procedure differs per mode:

From strong to eventual#

1. Update cluster.mode in your ConfigMap.
2. Rolling restart all pods one by one (kubectl rollout restart statefulset/bouine).
3. Verify bouine_cluster_mode_info{mode="..."} on every pod.

No data migration needed — each node starts with an empty cache.

From eventual to strong#

1. Update cluster.mode: strong in your ConfigMap.
2. Add hop_limit field if missing.
3. Rolling restart. The consistent-hash ring forms within seconds of all nodes joining.
4. Cache state from eventual is not preserved — nodes start with empty caches. Expect elevated miss rates for the first few minutes until caches warm up.

Alerts#

# Alert if cluster mode differs across pods (configuration drift).
- alert: ClusterModeMismatch
  expr: count(count by (mode) (bouine_cluster_mode_info == 1)) > 1
  for: 2m
  labels:
    severity: critical
  annotations:
    summary: "Cluster mode mismatch — pods running different consistency modes"

Troubleshooting quick reference#