Inserts are outrunning merges
Clients are inserting single rows or very small batches, or the partitioning scheme creates thousands of parts per day and the merger cannot keep up with them.
Batched inserts, partitioning rethought
Running ClickHouse?
ClickHouse scales from a single node to petabytes without most teams noticing the transition, until it does not. We are called in when merges start stacking, when the mutation queue grows faster than it drains, when a replicated table splits brain, or when the storage bill for raw logs crosses what the analytics was supposed to save. We trim the schema, retune the merges and get the cluster back into a steady state.
What it is
ClickHouse is a columnar OLAP database first released as open source by Yandex in 2016 and now developed by ClickHouse, Inc., founded in 2021. It aims at analytical queries over very large tables: sum, count, group by and window functions on billions of rows, returned in milliseconds to seconds.
The storage engine is MergeTree, which writes sorted parts to disk and merges them in the background. Replication is coordinated through ZooKeeper or its drop in ClickHouse Keeper. There is a managed ClickHouse Cloud offering, and self hosted deployments on bare metal and Kubernetes are both common. Teams typically arrive at ClickHouse when Postgres analytics have run out of headroom and Snowflake or BigQuery look too expensive.
The honest version
No technology is all upside. This is where ClickHouse earns its place and where it tends to cause trouble.
Sound familiar?
The ClickHouse problems that bring people to us, and what you have once each one is fixed.
Clients are inserting single rows or very small batches, or the partitioning scheme creates thousands of parts per day and the merger cannot keep up with them.
Batched inserts, partitioning rethought
A GROUP BY over very high cardinality keys, or a JOIN whose right side no longer fits in memory, taking out whatever node happened to receive the query.
Query rewritten or spilled to disk
An UPDATE or DELETE issued across a huge partition is now rewriting parts in the background and blocking later mutations, slowing every replica in the process.
Mutation handled, pattern changed
ZooKeeper under network pressure, or Keeper underprovisioned, letting sessions expire and leaving replicated tables unable to accept writes until the session recovers.
Coordination layer healthy again
A disk hiccup, a crash during merge, or a copy that went wrong leaving a part the server cannot read. The table works, but queries that touch the part fail.
Part restored from a replica or dropped
Raw JSON events written without codecs, partitions kept forever with no TTL, and materialised views that duplicate the underlying data four or five times over.
Schema tightened, TTLs set, bill cut
Upstream producers emitting slightly different schemas, a null that was supposed to be a zero, or a TSV with an extra tab. The ingest stops and the queue backs up until someone fixes the input.
Pipeline resilient to bad rows
A previous team set up replication, chose a partitioning scheme and left. The schema works but nobody knows why the merges run at 3am or what the backup story is.
Cluster mapped and owned by your team
What we do
Clear one stuck merge queue, or let us run the whole cluster.
How it works
Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.
We fix the cause, not the symptom. If something is down, getting it back up comes first.
Backups, monitoring, updates and access, so the same problem can't quietly come back.
You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.
Often found together
We work on the whole stack, so the problem can be wherever it actually is.
This almost always means inserts are too small or too frequent. ClickHouse wants batches of at least several thousand rows per insert, ideally per partition. If a client is streaming rows one at a time, a buffer table or a batching layer in front of it will stop the problem cold. The second place to look is the partitioning scheme, which sometimes creates far more parts than intended.
Three levers. Apply codecs like Delta, DoubleDelta or ZSTD on columns where they fit, add TTL expressions so cold partitions roll off or move to object storage automatically, and consolidate materialised views that are quietly tripling the data footprint. We measure the saving per lever so you know what paid back.
For most product analytics and observability workloads, yes, and the per query cost model usually comes out cheaper. The trade off is that you now run or pay for the cluster. If your queries are ad hoc and bursty, a per query warehouse may still win. If they are predictable and high volume, ClickHouse almost always wins. We will cost the two against your actual traffic.
Cloud takes merges, upgrades and replication off your plate for a markup. Self hosted gives control and lower cost at scale, but the operational knowledge is real and the on call pager becomes yours. For small teams without a database specialist, the managed service pays for itself. For larger teams or specific residency needs, self hosted still wins.
We identify the tables that are actually driving the slowness, design the ClickHouse schema for how those queries read rather than how they write, build an ingest pipeline that keeps both stores in sync during the transition, and cut reports over one by one. Postgres stays for the transactional work it is good at, and ClickHouse takes the aggregations off its back.
Free estimates. Describe your cluster size and what it is doing, paste a slow query or an error from the server log, and a price lands before any work is booked.
Free estimate
Tell us what you're running and what's going wrong. A person reads it and comes back with an estimate.
Your message is with a real person, and we'll get back to you soon.
Rather talk now? Call us: