Databases · ClickHouse documentation read 1 November 2026
ClickHouse's own sizing guide and usage recommendations, turned into choices for a server you rent whole: memory per core, disks, network, Keeper and the Linux settings to change on day one.
Quick answer
What hardware does ClickHouse need on bare metal?
ClickHouse on bare metal needs SSD storage, many cores and generous memory. Its sizing guide pairs 4 GB of RAM per core for general use and 8 GB for data warehousing. Below about 200 GB of compressed data, match RAM to the data size. Lumenfer's Lumen 32 (32 cores / 64 threads, 384 GB DDR5 ECC) costs $749.67 per 30 days, paid in crypto, no KYC.
| Memory per CPU core | 2 GB for high-concurrency serving, 4 GB for general use, 8 GB for data warehousing |
|---|---|
| Small data sets | Up to about 200 GB compressed: as much RAM as data |
| Large data, interactive queries | 128 GB of RAM or more |
| Memory to storage | 1:30 to 1:50 for frequently queried data; 1:100 to 1:130 for long retention |
| Storage | SSD if the budget allows; software RAID with mdadm; RAID 0 on SSD only with replication |
| File system | ext4 with noatime; XFS also works; avoid NFS and compressed file systems |
| CPU | Always the performance scaling governor |
| Kernel | Transparent huge pages on madvise; never disable overcommit; no outdated kernels |
| Network | 10 GbE if possible; 1 Gb works but slows replica repair and distributed queries |
| ClickHouse Keeper | Dedicated hosts in production, about 4 GB of RAM each, at least three nodes |
| Lumenfer port and traffic | 150 TB a month at 1 Gbit/s on Lumen 32; each plan lists its own |
ClickHouse sizes memory by the core, not by the server. Its guide pairs 4 GB per core with general workloads, 8 GB with data warehousing and 2 GB with high-concurrency serving. The ratios come from ClickHouse's sizing guide, which maps them to AWS instance families. Its own example is 100 GB of memory for 25 cores.
The usage recommendations add two rules. Up to about 200 GB of compressed data, use as much memory as data. For large data with interactive queries, use 128 GB or more so the hot subset stays in the page cache.
Applied to core counts, the 8 GB data-warehouse ratio gives:
Compare those with the memory column of the plan table below. Every Lumenfer plan meets ClickHouse's 4 GB-per-core general ratio, and all but Lumen 6 reach the 8 GB data-warehouse ratio. For low data volumes, the guide's floor is 8 GB. Spilling joins and sorts to disk saves memory, but the guide warns it can hurt query speed significantly.
All of them: ClickHouse sets its default query threads, max_threads, from the hardware threads it detects. On x86 chips with fewer than 32 cores and SMT, it counts logical cores, twice the physical count. A 16-core Ryzen with SMT therefore runs a query on up to 32 threads by default.
The sizing guide gives no universal CPU target. For ad-hoc analytics and data warehousing, it suggests aiming for 10 to 20% CPU use. That headroom lets heavy queries finish fast. For 100 or more queries per second, it points to compute-optimized machines with less memory per core.
Bare metal removes one ClickHouse caveat entirely. Its documentation asks virtualised setups to pass the host CPU through, so ClickHouse reads correct cpuid information. With no hypervisor, ClickHouse sees the real AMD EPYC 9354 or Ryzen directly. Our guide to how many CPU cores a rented server needs covers the clock-versus-cores trade-off.
Local SSD if the budget allows, which on Lumenfer means NVMe on every plan. ClickHouse prefers many servers with local drives over a few servers with disk shelves. It recommends software RAID with mdadm on Linux, ext4 with noatime, or XFS. It advises against NFS and compressed file systems.
Two-drive plans ship with the operating system on a software RAID 1 mirror; ask support before delivery if you prefer RAID 0 or two separate drives. ClickHouse calls RAID 0 acceptable on SSDs, but only with replication for safety. With a replica elsewhere, RAID 0 doubles usable NVMe.
By ClickHouse's memory-to-storage ratios, every Lumenfer plan is memory-heavy and disk-light. At 1:30 to 1:50, 256 GB of RAM pairs with 7.7 to 12.8 TB of data. Compare that with the disks column. That balance suits hot, frequently queried data. For long retention, the sizing guide lists tiered storage and object storage such as S3.
For a single server or a small replicated pair, yes; for a large sharded cluster, ClickHouse recommends 10 GbE. Its usage recommendations say 1 Gb works, but repairing a replica with terabytes of data and running distributed queries are much slower.
On Lumenfer, the port and monthly traffic are fixed per plan: 150 TB a month at 1 Gbit/s on Lumen 32. One dedicated IPv4 address and an IPv6 /64 come with every server. Each plan lists its port speed and its monthly traffic, counted on outgoing data per 30 days; going over never triggers an automatic charge.
Other hosts are stronger here. On 30 October 2026, KoDDoS listed an EPYC 9354 server in Hong Kong at $475 a month. It came with 384 GB, one 3.84 TB NVMe, a 10 Gbit/s port and 20 TB of traffic. A cluster that ships terabytes between nodes every day belongs on that kind of network.
Only for replicated or distributed setups, and then yes: ClickHouse strongly recommends dedicated Keeper hosts in production. Keeper replaces ZooKeeper for coordinating replicas. ClickHouse suggests at least three Keeper nodes for quorum, and says 4 GB of RAM is generally enough for each.
A single ClickHouse server with plain MergeTree tables needs no Keeper. ClickHouse's sizing guide also says to scale replicas up before adding more of them. It recommends the largest server available, because re-sharding later takes significant compute.
On Lumenfer, the cheapest path is often one large server first. Even Lumen 6 is far bigger than Keeper needs, so a three-node Keeper ensemble here is costly. For a replicated cluster, ClickHouse recommends at least three replicas per shard.
Four, from ClickHouse's usage recommendations: the performance CPU governor, transparent huge pages on madvise, overcommit left on, and a current kernel. ClickHouse warns at startup when huge pages are set to always, which slows it down on kernels before 5.9.
The overcommit rule conflicts with PostgreSQL, whose manual recommends strict overcommit. Run the two on separate machines; see the best server hosting for PostgreSQL.
# CPU: always the performance governor
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
# transparent huge pages: madvise, never always
echo madvise > /sys/kernel/mm/transparent_hugepage/enabled
# overcommit must stay 0 or 1 (never 2)
sysctl vm.overcommit_memory
# data volume on ext4 with noatime, in /etc/fstab, e.g.:
# /dev/md2 /var/lib/clickhouse ext4 defaults,noatime 0 2Size by memory per core and by hot data, then check the disks. The EPYC plans add ECC memory, which ClickHouse's sizing pages do not discuss; we prefer it for a primary store. Memory generation is the other question; see whether DDR5 is worth paying for. For more cores than we sell, read where to rent EPYC 9004 or 9005 servers.
In Lumenfer's check of 30 October 2026, 13 of 18 no-KYC crypto hosts sold at least one server with a 2022+ CPU and NVMe only. That was 113 of 229 comparable plans. Every plan and term is on the Lumenfer pricing page.
| Plan | CPU | Cores | Memory | Disks | Per 30 days | ClickHouse fit |
|---|---|---|---|---|---|---|
| Lumen 6 | AMD EPYC 4244P | 6 cores / 12 threads | 32 GB DDR5 ECC | 2 × 960 GB NVMe | $91.65 | Dev, small log stores, a Keeper node |
| Lumen 8 | AMD Ryzen 7 9700X | 8 cores / 16 threads | 64 GB DDR5 | 2 × 1 TB NVMe | $109.37 | Data sets that fit in memory |
| Lumen 16 | AMD Ryzen 9 9950X | 16 cores / 32 threads | 128 GB DDR5 | 2 × 1.92 TB NVMe | $205.27 | Product analytics, dashboards |
| Lumen 24 | AMD EPYC 9254 | 24 cores / 48 threads | 256 GB DDR5 ECC | 2 × 1.92 TB NVMe | $356.25 | Data warehouse on ECC memory |
| Lumen 32 | AMD EPYC 9354 | 32 cores / 64 threads | 384 GB DDR5 ECC | 2 × 3.84 TB NVMe | $749.67 | One large node before sharding |
| Server | Specs | Price per month |
|---|---|---|
| Lumenfer Lumen 32 | AMD EPYC 9354 (2022, 32 cores / 64 threads), 384 GB DDR5 ECC, 2 × 3.84 TB NVMe in RAID 1, 150 TB a month at 1 Gbit/s, 1 IPv4 and an IPv6 /64 | $749.67 |
| MonoVM: EPYC 9354/384/2x3.84TB NVMe | EPYC 9354, 32 cores / 384 GB / 2x3840 GB | $999.57 |
| 1Gbits (second cheapest) | Comparable 32-core server | $1038.02 |
| Cherry Servers (third cheapest) | Comparable 32-core server | €989.00 ($1104.91) |
| Median of 6 hosts | Comparable 32-core servers | $1476.45 |
Prices checked on
Crypto-paying hosts with no ID check at signup; 2022+ CPU, NVMe only, at least the same cores, memory and NVMe, monthly billing. 1 EUR = 1.1172 USD (ECB, 30 October 2026). MonoVM and 1Gbits accept crypto through several processors, including Cryptomus/Heleket, a processor the UK designated for sanctions on 8 October 2026.
Whole machines with root access, billed per 30 days, 3 months, 6 months or 12 months from a crypto balance. Amsterdam (Netherlands), Reykjavik (Iceland), Zurich (Switzerland) and Bucharest (Romania): all five plans in all four cities, one price everywhere.
Flagship
16 cores · 128 GB DDR5
$205.27
3 months $554.22 (−10%)6 months $985.29 (−20%)1 year $1,724.26 (−30%)
24 cores · 256 GB DDR5
$356.25
3 months $961.87 (−10%)6 months $1,710 (−20%)1 year $2,992.50 (−30%)
32 cores · 384 GB DDR5
$749.67
3 months $2,024.10 (−10%)6 months $3,598.41 (−20%)1 year $6,297.22 (−30%)
In every plan: 2022+ CPU · DDR5 · NVMe only · RAID 1 · Dedicated IPv4 · No KYC
ClickHouse's sizing guide says total memory should not drop below 8 GB for low data volumes. Real analytics needs far more: as much RAM as compressed data up to about 200 GB, and 128 GB or more beyond that.
ClickHouse's sizing and usage pages do not mention ECC. For a primary analytics store, we would still pick it, as for any database. Lumen 6, Lumen 24 and Lumen 32 carry ECC memory.
Avoid it. ClickHouse says never to disable overcommit, while PostgreSQL's manual recommends strict overcommit. Both also want most of the RAM for their own caches. Separate machines keep both happy.
One big server. ClickHouse's sizing guide says to scale replicas vertically before adding more, and to choose the largest server available, because re-sharding needs significant compute. Add replicas when you need availability.
Top up a crypto balance in BTC, USDT, USDC, ETH, SOL, LTC (minimum $25) and order from it. No KYC: we never ask for an ID, a selfie, documents, a card or a phone number. Servers are prepared by a person once your payment is confirmed, within 72 hours rather than instantly; the dashboard shows progress and then the IP address when the machine is ready.
From $91.65/30 days · no KYC
Order bare metal