Metrics and Alerts Reference - ClickHouse Documentation

Key metrics and recommended alert rules for ClickHouse Private components. This is not exhaustive — your environment may expose additional metrics depending on configuration.

Operator Metrics

Metric	Type	Description
`last_cluster_reconcile`	Gauge	Timestamp of the last reconciliation for a given CR (labeled by `app`, e.g., `c-default-xx-01`). Use to determine if reconciles are occurring regularly.
`controller_runtime_reconcile_errors_total`	Counter	Total number of reconciliation errors per controller. Use in conjunction with `controller_runtime_reconcile_total` to calculate the reconciliation error rate.
`controller_runtime_reconcile_total`	Counter	Total number of reconciliation attempts per controller.

ClickHouse Server Metrics

Server metrics are exposed via the :8123/metrics Prometheus endpoint on each server pod. This endpoint requires authentication and should use a dedicated user with read-only privileges. Note: The prometheus.io/* annotations on server pods expose some metrics, but do not include the ClickHouse_CustomMetrics_* metrics listed below. Set up a PodMonitor or equivalent scrape target for :8123/metrics.

Metric	Type	Description
`ClickHouse_CustomMetric_NumberOfBrokenDetachedParts`	Gauge	Number of broken detached parts.
`ClickHouse_CustomMetric_LostPartCount`	Gauge	Number of lost parts, indicating data loss. False positives are possible.
`ClickHouseErrorMetric_CANNOT_WRITE_TO_FILE_DESCRIPTOR`	Counter	Count of CANNOT_WRITE_TO_FILE_DESCRIPTOR errors. Mostly indicates a full cache disk.
`ClickHouseErrorMetric_CHECKSUM_DOESNT_MATCH`	Counter	Count of checksum mismatch errors. May indicate a bug after upgrade.
`ClickHouseErrorMetric_CORRUPTED_DATA`	Counter	Count of corrupted data errors. May indicate a bug after upgrade.
`ClickHouseErrorMetric_LOGICAL_ERROR`	Counter	Count of logical errors. Often a bug in the codebase.
`ClickHouseErrorMetric_NOT_ENOUGH_SPACE`	Counter	Count of NOT_ENOUGH_SPACE errors. Could indicate full PVC, misconfiguration, or large temp data reservation.
`ClickHouseErrorMetric_POTENTIALLY_BROKEN_DATA_PART`	Counter	Count of potentially broken data part errors. Indicates data loss on SELECT.
`ClickHouseErrorMetric_REPLICA_ALREADY_EXISTS`	Counter	Count of replica already exists errors. Usually a bug in the Replicated database engine.
`ClickHouseMetrics_IsServerShuttingDown`	Gauge	`1` if the ClickHouse server is in the process of shutting down.
`ClickHouse_CustomMetric_TableReadOnlyDurationSeconds`	Gauge	Duration in seconds that a table has been in READONLY mode.

ClickHouse Keeper Metrics

Keeper metrics are exposed via :8001/metrics. A PodMonitor should be created for keeper pods to capture these metrics.

Note: Specific keeper metric definitions are pending documentation (TODO in source). The metrics endpoint is available but individual metric descriptions have not yet been cataloged.

The following metric is referenced by the sizing guidance:

Metric	Type	Description
`ClickHouseAsyncMetrics_KeeperApproximateDataSize`	Gauge	Approximate size of the logical Keeper metadata in bytes, exposed by the Keeper metrics endpoint. Use it to watch metadata growth trends.

Keeper Sizing Alerts

Keeper holds its dataset in memory and a fixed allocation can be outgrown silently (see Keeper sizing: node vs pod), so alert on memory headroom and watch metadata growth. Example rule (adjust the threshold and label filters to your environment; the container_* and kube_* series are standard cAdvisor / kube-state-metrics metrics):

alert: KeeperMemoryHeadroomLow
expr: |
  max by (namespace, pod) (container_memory_working_set_bytes{container=~".*-keeper"})
    /
  max by (namespace, pod) (kube_pod_container_resource_limits{resource="memory", container=~".*-keeper"})
  > 0.8
for: 15m

Purpose: Fires when a Keeper pod’s working set exceeds 80% of its memory limit: the pod is running out of headroom, and the pod allocation should be raised before Keeper is OOM-killed.

Alert Rule Definitions

Operator Alerts

ClickhouseOperatorNotReconciling

alert: ClickhouseOperatorNotReconciling
expr: avg(increase(last_cluster_reconcile[90m])) by (app) == 0
for: 120m

Purpose: Alerts when the operator has not reconciled within 2 hours. Recommended action:

Check that the operator pod is running and healthy.
Check operator logs to see what is preventing reconciliation.
Verify there is no clickhouse.com/skip-reconcile annotation on the CR.

ClickhouseOperatorReconcileErrors

alert: ClickhouseOperatorReconcileErrors
expr: |
  (
    sum(rate(controller_runtime_reconcile_errors_total{namespace="clickhouse-operator-system"}[5m])) by (controller, namespace)
    /
    sum(rate(controller_runtime_reconcile_total{namespace="clickhouse-operator-system"}[5m])) by (controller, namespace)
  ) > 0.05
for: 15m

Purpose: Alerts when reconciliation errors exceed 5% of total reconciliation attempts. Recommended action: Check operator logs for error messages to find the underlying cause.

ClickHouse Server Alerts

ClickHouseBrokenDetachedParts

alert: ClickHouseBrokenDetachedParts
expr: ClickHouse_CustomMetric_NumberOfBrokenDetachedParts > 100
for: 60m

Purpose: Triggered when broken detached parts exceed 100 for at least 60 minutes. Recommended action:

Wait to see if ClickHouseDataLoss also triggers.
For SMT/RMT tables: if data loss alert also fires, investigate and mitigate data loss first.
For local metadata tables (s3disk or s3diskWithCache): some small number of broken detached parts may not indicate an incident (files may be created but not written during hard restarts).
Contact ClickHouse support if the issue persists.

ClickHouseDataLoss

alert: ClickHouseDataLoss
expr: ClickHouse_CustomMetric_LostPartCount > 0

Purpose: Indicates potential data loss. Fires immediately when lost parts are detected. Recommended action: Contact ClickHouse support. Initial investigation steps:

Query lost parts:

SELECT database, table, lost_part_count AS value
FROM system.replicas
WHERE value > 0

Search logs for lost parts:

SELECT hostName(), event_time, logger_name, message
FROM clusterAllReplicas(default, system.text_log)
WHERE message_format_string = 'Part {} is lost forever.'
ORDER BY hostName(), event_time

Investigate history of a specific lost part:

SELECT event_time, message
FROM system.text_log
WHERE message LIKE '%<part name>%'
  AND hostName() = '<host where lost forever log was created>'
ORDER BY event_time ASC

Check for false positives:
- Check if the table has TTL and the lost part should have been dropped by TTL.
- Check system.query_log for TRUNCATE or DROP PARTITION queries.

ClickHouseCannotWriteToFileDescriptor

alert: ClickHouseCannotWriteToFileDescriptor
expr: |
    increase(ClickHouseErrorMetric_CANNOT_READ_FROM_FILE_DESCRIPTOR[30s]) > 0
    or (
        ClickHouseErrorMetric_CANNOT_READ_FROM_FILE_DESCRIPTOR > 0
        and
        ignoring (table) ClickHouseAsyncMetrics_Uptime < 3 * 60 * 60
    )

Note: The alert expression checks CANNOT_READ_FROM_FILE_DESCRIPTOR, not CANNOT_WRITE. This is intentional in the upstream alert definition — the read metric fires in the same failure scenario (full cache disk). Purpose: Mostly indicates a full cache disk (“no space left on device”). Recommended action:

If node size or type recently changed, may be related to misconfiguration.
Known issue: cache disk usage tracking can be incorrect when join_algorithm = 'partial_merge' is used.
Diagnostic steps:
1. kubectl exec into the pod and run df -h to check cache disk size.
2. Run SELECT path, max_size FROM system.filesystem_cache_settings to check required cache size.
3. If actual disk is smaller than configured, the issue is misconfiguration.
Contact ClickHouse support.

ClickHouseChecksumsMismatch

alert: ClickHouseChecksumsMismatch
expr: |
    increase(ClickHouseErrorMetric_CHECKSUM_DOESNT_MATCH[30s]) > 0
    or (
        ClickHouseErrorMetric_CHECKSUM_DOESNT_MATCH > 0
        and
        ignoring (table) ClickHouseAsyncMetrics_Uptime < 3 * 60 * 60
    )

Purpose: Checksums of data parts don’t match. May indicate a bug after upgrade. Recommended action: Contact ClickHouse support.

ClickHouseCorruptedData

alert: ClickHouseCorruptedData
expr: |
    increase(ClickHouseErrorMetric_CORRUPTED_DATA[30s]) > 0
    or (
        ClickHouseErrorMetric_CORRUPTED_DATA > 0
        and
        ignoring (table) ClickHouseAsyncMetrics_Uptime < 3 * 60 * 60
    )

Purpose: Data parts are corrupted. May indicate a bug after upgrade. Recommended action: Contact ClickHouse support.

ClickHouseLogicalErrors

alert: ClickHouseLogicalErrors
expr: |
    increase(ClickHouseErrorMetric_LOGICAL_ERROR[30s]) > 0
    or (
        ClickHouseErrorMetric_LOGICAL_ERROR > 0
        and
        ignoring (table) ClickHouseAsyncMetrics_Uptime < 3 * 60 * 60
    )

Purpose: Logical errors occurred. Often a bug in the codebase. Recommended action: Contact ClickHouse support.

ClickHouseNotEnoughSpaceErrors

alert: ClickHouseNotEnoughSpaceErrors
expr: |
    increase(ClickHouseErrorMetric_NOT_ENOUGH_SPACE[30s]) > 0
    or (
        ClickHouseErrorMetric_NOT_ENOUGH_SPACE > 0
        and
        ignoring (table) ClickHouseAsyncMetrics_Uptime < 3 * 60 * 60
    )

Purpose: NOT_ENOUGH_SPACE errors emitted. Could indicate full PVC, misconfiguration, or large temporary data reservation. Recommended action: Contact ClickHouse support.

ClickHouseBrokenPartDetectedOnSelect

alert: ClickHouseBrokenPartDetectedOnSelect
expr: |
    increase(ClickHouseErrorMetric_POTENTIALLY_BROKEN_DATA_PART[30s]) > 0
    or (
        ClickHouseErrorMetric_POTENTIALLY_BROKEN_DATA_PART > 0
        and
        ignoring (table) ClickHouseAsyncMetrics_Uptime < 3 * 60 * 60
    )

Purpose: SELECT failed due to POTENTIALLY_BROKEN_DATA_PART error, indicating data loss. Recommended action: See ClickHouseDataLoss action. Examine logs for POTENTIALLY_BROKEN_DATA_PART exception. Check system.errors if not found in logs. Contact ClickHouse support.

ClickHouseReplicaAlreadyExists

alert: ClickHouseReplicaAlreadyExists
expr: |
    increase(ClickHouseErrorMetric_REPLICA_ALREADY_EXISTS[30s]) > 0
    or (
        ClickHouseErrorMetric_REPLICA_ALREADY_EXISTS > 0
        and
        ignoring (table) ClickHouseAsyncMetrics_Uptime < 3 * 60 * 60
    )

Purpose: Replica creation failed because a replica already exists at the path. Usually a bug in the Replicated database engine or Shared Catalog. Recommended action: Unlikely to be user error. Contact ClickHouse support.

ClickHouseServerShutdownStuck

alert: ClickHouseServerShutdownStuck
expr: ClickHouseMetrics_IsServerShuttingDown == 1
for: 70m

Purpose: ClickHouse server has been shutting down for over an hour. Recommended action: Check threads via system.stack_trace or GDB.

ClickHouseTableReplicasReadOnly

alert: ClickHouseTableReplicasReadOnly
expr: ClickHouse_CustomMetric_TableReadOnlyDurationSeconds > 3600

Purpose: A table has been in READONLY mode for more than an hour. Recommended action:

Check server logs filtered by affected table name(s).
Check keeper logs for potential issues.

Query current read-only tables:

SELECT dateDiff('second', readonly_start_time, now()) AS readonly_duration_seconds, database, table, hostname()
FROM clusterAllReplicas(default, system.replicas)
WHERE is_readonly = 1

Try running SYSTEM RESTART REPLICA for affected tables.
A simple replica restart may resolve the issue.

Diagnostic SQL Queries

Replication Queue Size per Table

Trigger: Alert if count exceeds 100 for any table.

SELECT
    concat(database, '.', table),
    count()
FROM system.replication_queue
GROUP BY database, table

Alternative:

SELECT
    concat(database, '.', table),
    queue_size
FROM system.replicas

Replication Queue Oldest Entry per Table

Trigger: Alert if oldest entry is older than 1 day.

SELECT
    concat(database, '.', table),
    min(create_time)
FROM system.replication_queue
GROUP BY database, table

Read-Only Tables

SELECT
    dateDiff('second', readonly_start_time, now()) AS readonly_duration_seconds,
    database,
    table,
    hostname()
FROM clusterAllReplicas(default, system.replicas)
WHERE is_readonly = 1

Lost Parts by Table

SELECT database, table, lost_part_count AS value
FROM system.replicas
WHERE value > 0

Lost Parts Log Search

SELECT hostName(), event_time, logger_name, message
FROM clusterAllReplicas(default, system.text_log)
WHERE message_format_string = 'Part {} is lost forever.'
ORDER BY hostName(), event_time

Part History Investigation

SELECT event_time, message
FROM system.text_log
WHERE message LIKE '%<part name>%'
  AND hostName() = '<host>'
ORDER BY event_time ASC

Cache Disk Configuration Check

SELECT path, max_size
FROM system.filesystem_cache_settings

General Alerting Recommendations

Standard infrastructure alerts should be configured for:

Crashlooping pods
Unschedulable pods
Pod OOM kills
PVC capacity
Node health

These are environment-specific and not covered by the ClickHouse-specific alerts above. The Grafana ClickHouse mixin provides a prebuilt dashboard for many ClickHouse metrics.

​Operator Metrics

​ClickHouse Server Metrics

​ClickHouse Keeper Metrics

​Keeper Sizing Alerts

​Alert Rule Definitions

​Operator Alerts

​ClickhouseOperatorNotReconciling

​ClickhouseOperatorReconcileErrors

​ClickHouse Server Alerts

​ClickHouseBrokenDetachedParts

​ClickHouseDataLoss

​ClickHouseCannotWriteToFileDescriptor

​ClickHouseChecksumsMismatch

​ClickHouseCorruptedData

​ClickHouseLogicalErrors

​ClickHouseNotEnoughSpaceErrors

​ClickHouseBrokenPartDetectedOnSelect

​ClickHouseReplicaAlreadyExists

​ClickHouseServerShutdownStuck

​ClickHouseTableReplicasReadOnly

​Diagnostic SQL Queries

​Replication Queue Size per Table

​Replication Queue Oldest Entry per Table

​Read-Only Tables

​Lost Parts by Table

​Lost Parts Log Search

​Part History Investigation

​Cache Disk Configuration Check

​General Alerting Recommendations