ClickHouse Clusters & Load Balancers

Sticky routing with X-CH-UI-Session for multi-node ClickHouse deployments behind chproxy, HAProxy, or Kubernetes

If your ClickHouse runs as a single server, you can skip this page — CH-UI works out of the box. This page is for deployments where CH-UI talks to a load balancer in front of multiple ClickHouse pods: chproxy, HAProxy, an nginx ingress, a Kubernetes Service, etc.

The problem

ClickHouse's system.* tables are node-local. system.query_log, system.processes, system.metrics, system.parts — each only contains data for the specific pod that handled that request.

Without sticky routing, every page refresh in CH-UI can hit a different pod. Query history flickers between pods. Running queries vanish. Part counts contradict each other. The UI becomes unusable.

The fix is session affinity: every request from a given CH-UI session must land on the same pod.

How CH-UI signals affinity

Every request CH-UI sends to ClickHouse carries this header:

X-CH-UI-Session: ch-ui-<32 hex>

The value is a stable per-process identifier — generated once when the CH-UI server (self-hosted) or the connector agent (cloud) starts, sent verbatim on every subsequent request. Configure your load balancer to hash on this header for upstream selection.

Why not ClickHouse's native `session_id`?

ClickHouse's built-in session_id URL parameter would seem like the obvious fit — but it triggers a server-side mutex. Only one query can run at a time per session. A dashboard with 8 concurrent panel queries deadlocks instantly with SESSION_IS_LOCKED.

The custom header gives load balancers something to hash on without invoking ClickHouse's session locking. Pods serve concurrent queries normally; only the routing is sticky.

Load balancer configuration

chproxy

[server.http]
listen_addr = ":8123"

[server.proxy]
# Hash on the session header for sticky upstream selection.
[server.proxy.hash]
header_name = "X-CH-UI-Session"

[[users]]
name = "chui"
password = "..."
to_cluster = "ch_cluster"

HAProxy

frontend ch_in
    bind *:8123
    default_backend ch_pool

backend ch_pool
    balance roundrobin
    # Stick on the session header — store hash for 1h.
    stick-table type string len 64 size 5k expire 1h
    stick on req.hdr(X-CH-UI-Session)
    server ch01 10.0.0.11:8123 check
    server ch02 10.0.0.12:8123 check
    server ch03 10.0.0.13:8123 check

Kubernetes — nginx ingress

The default sessionAffinity: ClientIP on a Service does not work for cloud CH-UI, because all traffic comes from a single agent IP. Use an L7 ingress instead:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: clickhouse
  annotations:
    # Hash on the session header. Header → variable: lowercase + dashes→underscores → prefix $http_
    nginx.ingress.kubernetes.io/upstream-hash-by: "$http_x_ch_ui_session"
spec:
  rules:
    - host: clickhouse.example.com
      http:
        paths:
          - path: /
            pathType: Prefix
            backend:
              service:
                name: clickhouse
                port: { number: 8123 }

Kubernetes — Traefik

apiVersion: traefik.containo.us/v1alpha1
kind: Middleware
metadata:
  name: chui-affinity
spec:
  plugin:
    headerHash:
      header: X-CH-UI-Session

(or use a ServersTransport with disableHTTP2: false plus the stickyCookie approach if you don't have a hashing plugin available — fall back to a cookie strategy and copy the value from the header in middleware).

How to verify it's working

Sidebar — the bottom of the CH-UI sidebar shows the hostname of the pod currently serving your session. Refresh a few times. The hostname should not change.
Admin → Overview → Cluster Topology — shows the full cluster: every shard, every replica, host + address + port. The pod serving you has a green dot in the Local column.
Backend log on the LB — confirm the hash header is being read. For chproxy: --log-level=debug shows the selected upstream per request. For HAProxy: show table mytable shows the stick-table populated.

What you see if it's not working

Hostname in the sidebar changes between page loads → sticky routing isn't active. The header is being sent (always is), but the LB isn't hashing on it.
Dashboards show inconsistent counts on refresh → same root cause.
SESSION_IS_LOCKED errors → you've configured ClickHouse's native session_id somewhere instead of (or in addition to) the header. Remove that — use only the header.

Single-node setups

If you're on a single ClickHouse server, the header is still sent — ClickHouse ignores unknown headers, so it's a no-op. The sidebar shows the hostname; the Admin cluster topology shows "Single-node setup detected". Nothing to configure.