NeverBlink for ClickHouseClickHouse fails quietly. NeverBlink is listening.
ClickHouse is fast because it defers work. Merges, TTL deletes and replication all happen in the background, which is why the worst ClickHouse incidents don't start with an alert. NeverBlink watches that background machinery and tells you, with the fix attached, while the problem is still a warning.
Nobody gets paged for any of these.
ClickHouse exposes over a thousand metrics through system tables, and very few teams know which fifteen actually predict outages. These are the failures NeverBlink is built to catch.
Merge backlog piles up
Parts accumulate faster than the merge pool clears them, until ingestion fails with TOO_MANY_PARTS.
A materialized view drops data
It fails mid-insert. The source table gets the row, the target doesn't, and rollups under-count for weeks.
TTL stops keeping up
Deletes only happen inside merges, and ClickHouse deprioritizes them when the pool is busy, until a replica goes read-only.
Replicas quietly diverge
A Keeper hiccup leaves a replica stale, and by default ClickHouse serves reads from it rather than erroring.
Detached parts eat the disk
Broken parts land in detached, where ClickHouse never cleans them up, until the disk fills.
Shard skew
One shard carries far more data or traffic than the others, and its node becomes the cluster's ceiling.
Purpose-built watchers for ClickHouse
Watchers for merge backlog, replication delay, stuck mutations, Keeper flapping and Kafka consumer lag, ranked by severity. Each alert arrives with the investigation done, down to copy-paste SQL.
- Slack, email, PagerDuty and 14+ more destinations
- Shows what happened, what it affects and what to run next
See how data actually flows
NeverBlink maps data from source tables through materialized view cascades to TTL and retention. It reviews every table for wrong compression codecs, oversized types and partition keys that fight your queries, each with the ALTER statement and the estimated gain.
Find the queries behind the bill, then get the rewrite
Every query on the cluster is fingerprinted and ranked by runtime, memory, CPU and IO. NeverBlink explains each expensive one, such as a full scan against the wrong ORDER BY key or a memory-hungry GROUP BY. The optimizer then proposes a rewritten query next to the original, with the schema changes worth considering.
A health score you can act on.
Scored checks are re-run continuously by a lightweight collector that reads system tables only, never your data.
Stability
- MergeTree part counts
- Disk and memory headroom
- Node version alignment
Performance
- Merge pool saturation
- Rejected or delayed inserts
- Query latency and error rates
Resiliency
- Replication queue health
- DDL queue health
- Recent backups
Security
- Authentication configuration
- Server version and known CVEs
On ClickHouse Cloud? You're covered too.
ClickHouse Cloud keeps the cluster running. It can't fix unoptimized application usage, and managed pricing makes the easy fix, scaling up, the expensive one. NeverBlink adds query analytics, schema reviews and cost recommendations on top, so what you run on it is worth the bill.
NeverBlink provides continuous insights on queries and node sizing, and walks us through the changes required to optimize the cluster. Our dashboard queries dropped from 8 seconds to subsecond.
Running in minutes.
Create a read-only monitoring user, point the collector at your cluster, and the first health assessment arrives the same day. With auto-discovery on, the collector finds every node through system.clusters.
Find out what your ClickHouse cluster isn't telling you.
Your first cluster is free.