ClickHouse is the database people reach for when a dashboard query starts taking eleven seconds and no amount of Postgres indexing fixes it. The reason it works is not magic: it stores data by column instead of by row, so a query that touches three columns of a fifty-column table reads three columns' worth of disk. Add aggressive compression and vectorised execution and you get the aggregate scans that made it famous.
I'm Ajay, I build PandaStack. We use ClickHouse ourselves — every sandbox lifecycle event on the platform lands in it — and we run managed Postgres as a product, not managed ClickHouse. So this is a genuine survey rather than a pitch, including the part where the honest recommendation for a lot of readers is 'you don't need ClickHouse yet'.
The one architectural split that matters
Every ClickHouse hosting option falls into one of two camps, and this determines your cost curve far more than the vendor's per-unit price does.
Shared-nothing is classic ClickHouse: data lives on local NVMe attached to the nodes that query it. It is the fastest configuration that exists, because there is no network between the CPU and the bytes. It is also the one where storage and compute are welded together — needing more disk means buying more compute, and scaling out means physically resharding.
Shared-storage keeps data in object storage (S3, GCS) with the compute nodes holding a local cache. Storage becomes effectively unlimited and cheap, nodes become stateless and disposable, and you can scale compute up for a heavy query and back down afterwards. The cost is latency on a cold cache: a query touching data nobody has read recently pays an object-storage round trip. This is the architecture ClickHouse Cloud is built on, and it is where the market has moved.
The options
ClickHouse Cloud
The first-party managed service, built on the shared-storage architecture with separated compute. It has the features that matter operationally — automatic scaling, idle suspension for dev services, native S3 tiering, and the newest engine version without you doing anything. If you are choosing on capability alone this is the reference and everything else is compared against it. The reasons to look elsewhere are the usual ones: cost at scale, cloud/region availability, and a hard requirement that data stay in your own account.
Altinity
The long-standing independent option, and the one that specialises in running ClickHouse inside your own cloud account rather than theirs. That distinction matters a lot in regulated environments: the data never leaves your VPC, but you still get an operator that handles upgrades, backups and cluster topology. Altinity also maintains a stable long-term-support build, which is a real consideration given how fast upstream ClickHouse moves. Pick it when compliance says the bytes stay in your account.
Self-hosted on Kubernetes (Altinity or Bitnami operators)
The Kubernetes operators make cluster topology declarative — shards, replicas, ZooKeeper or ClickHouse Keeper, storage classes, the lot. If you already run a Kubernetes platform team, this is a well-trodden path and the cheapest per byte you will find. If you don't, understand what you're signing up for: ClickHouse's distributed mode has real operational depth, and replication misconfiguration produces subtly wrong query results rather than loud errors.
Cloud marketplace and adjacent services (Aiven, DoubleCloud, Tinybird)
Aiven runs managed ClickHouse alongside its Kafka and Postgres services, which is convenient if you want one vendor for the whole pipeline. Tinybird is a different animal worth knowing about: it wraps ClickHouse in an API-first product where you define queries as endpoints and it handles ingestion, schema and HTTP serving. If what you actually want is 'analytics API' rather than 'a database I administer', Tinybird removes a whole layer of work — at the price of a more opinionated model.
A single node you run yourself
The most underrated option. ClickHouse on one machine with good NVMe handles a genuinely large amount of data — tens of billions of rows is unremarkable — and single-node means no ZooKeeper, no replication topology, no distributed query debugging. Most teams that think they need a cluster need a bigger disk. Take the backup problem seriously (BACKUP TO S3 on a schedule, restore-tested) and this configuration will carry you a very long way.
ClickHouse in an app or sandbox VM
For development, CI and short-lived analysis, running ClickHouse inside a microVM alongside the code that queries it is fast to set up and disposable by design. On PandaStack an app runs in its own Firecracker microVM with a full Linux userspace, so this is an install line:
# Build command: install ClickHouse into the app's own microVM.
curl -fsSL https://clickhouse.com/ | sh && ./clickhouse install --noninteractive
# Start command: server in the background, your app in the foreground.
clickhouse start && node server.jsThere is a useful accident of protocol design here. ClickHouse's native protocol on 9000 is raw TCP, but it also speaks a complete HTTP interface on 8123 — you can run any query with a POST body. That means it works over an HTTP-only path, which is exactly what PandaStack's per-port URLs are. A sandbox's 8123 is reachable; its 9000 is not.
# The HTTP interface takes SQL as a POST body -- no native client needed.
curl -s "https://8123-$SANDBOX_ID.$PREVIEW_SUFFIX/" \
--data-binary "SELECT count() FROM events WHERE day >= today() - 7"The question to ask before any of this
A large fraction of ClickHouse adoptions are premature, and the tell is consistent: the team has a slow dashboard, not an analytics workload. Before you add a second database engine with its own ingestion, schema, backup and operational story, check whether Postgres can still do the job.
- Under roughly 50–100 million rows in the fact table, a well-indexed Postgres with pre-aggregated rollup tables is usually fast enough and infinitely simpler.
- If your slow queries are all the same shape, a materialised view refreshed on a schedule beats a new database engine.
- If you need one specific analytics capability — time bucketing, retention windows — a Postgres extension may cover it without a second system.
- ClickHouse earns its place when you are scanning hundreds of millions to billions of rows, with high-cardinality group-bys, over data that arrives continuously.
The costs of adding ClickHouse are not just the hosting bill. You now have two schemas that must agree, a pipeline between them that can lag or break, and a second thing to back up and restore-test. That's worth paying when the query speed is genuinely the constraint on the product. It is not worth paying to make a report load in 400ms instead of 3 seconds.
A short evaluation checklist
- Size the working set — the data your queries actually touch, not your total data. That number decides shared-nothing versus shared-storage.
- Check whether your ingestion path exists. ClickHouse ingests best in large batches; a naive row-at-a-time insert loop will produce thousands of tiny parts and grind merges to a halt.
- Confirm the backup story is BACKUP TO an object store on a schedule, and restore-test it once. A columnar database with no tested restore is a spreadsheet with better marketing.
- Decide whether idle suspension matters. For dev and staging services, a provider that sleeps them is a large fraction of the bill.
- Read the pricing page for the storage line specifically. Compute is usually the headline; storage is usually what surprises you.
The summary
ClickHouse Cloud if you want the reference implementation with separated storage and compute and you're happy for the data to live in their account. Altinity if compliance requires it to live in yours. A Kubernetes operator if you already have a platform team and want the lowest cost per byte. Tinybird if what you want is an analytics API rather than a database. A single well-specified node if you're earlier than you think — it goes much further than people expect. And a sandbox-hosted instance for dev and CI, where the ephemerality is a feature and the HTTP interface makes it reachable without any special networking.