Benchmarks · v2.0.29

A full SOC day, measured. Rerun it yourself.

We installed nano fresh, licensed it, and ran what a security team actually does: steady ingest, analysts hunting, 200 detections, AI agents and wall dashboards, all at once. Then we pushed one cluster until something broke. The kit, the seed and every raw result are public.

Get the benchmark kit Raw results (JSON)
SOC day, 100 GB/day 0 events lost, 0 query errors
Search p95 during the SOC day 253 to 318 ms
Event to alert, p50 19 s
Cost per GB ingested $0.058 at 100 GB/day
Ceiling, 3 ingest hosts 22,150 events/s (1.9 TB/day)
Ceiling, 4 ingest hosts 27,047 events/s (2.3 TB/day)
Release nano v2.0.29, stock config
Reproducible Open kit, fixed seed, raw JSON
Benchmark 1 · A normal SOC day

Four hours of real work, nothing dropped.

One small cluster (4 vCPU app, 8 vCPU ClickHouse, 8 vCPU ingest) ran for 4 hours at each sizing tier, 50 and 100 GB/day, on two identical clusters side by side. During every minute of it, 8 analysts replayed 94 real investigations, 200 scheduled detections ran in alerting mode, 2 AI agents hunted through Agent Lane, and 4 wall displays refreshed an 8-panel dashboard with the cache bypassed.

50 GB/day100 GB/day
Measured ingest 586 events/s, 49.7 GB/day 1,173 events/s, 99.5 GB/day
Events sent / stored 8,444,268 / 8,444,268 16,885,890 / 16,885,890
Markers stored exactly once 57,444 of 57,444 114,870 of 114,870
Search p50 / p95: triage 44 / 318 ms 46 / 253 ms
Search p50 / p95: hunt 46 / 318 ms 49 / 291 ms
Search p50 / p95: retro IR 53 / 310 ms 55 / 263 ms
Dashboard refresh (8 panels) p95 733 ms 692 ms
Agent Lane query p50 / p95 15 / 82 ms (3,201 queries) 18 / 127 ms (3,175 queries)
Query errors (analysts, dashboards, agents) 0 / 0 / 0 (20,031 queries) 0 / 0 / 0 (20,029 queries)
Event to alert p50 / p95: all rules 19.1 / 32.2 s 19.0 / 28.9 s
Event to alert p50 / p95: every-minute rules 18.2 / 25.2 s 19.2 / 25.3 s
Ingest lag p50 / p99 6.2 / 11.1 s 6.4 / 11.5 s
Busiest host (ClickHouse) CPU p95 26% (74% headroom) 33% (67% headroom)
Lowest memory / disk headroom 86% / 87% 84% / 83%
Cost per GB/day ingested $0.116 per GB ($5.76/day) $0.058 per GB ($5.76/day)
SOC day results, nano v2.0.29, 4 hours per tier

No search took longer than 1.2 s on the server, no dashboard refresh took longer than 1.4 s, and no query failed. At 100 GB/day the cluster used a third of its ClickHouse CPU. Event-to-alert time includes each rule's schedule by design: a rule that runs every 5 minutes can wait up to 5 minutes, so the every-minute rules are shown on their own.

Benchmark 2 · Ceiling

How far one cluster goes.

A role-split Compose install on plain VMs, the way a customer runs it: one app host, two ClickHouse shards and 3 or 4 ingest hosts. We ramp events per second in steps of 5 minutes each. The ceiling is the highest step that keeps lag p99 under 60 s, loses nothing, keeps sender errors under 0.1% and keeps up with at least 95% of the target rate.

3 ingest hosts4 ingest hosts
Sustained events/s, median (range) 22,150 (15,345 to 23,097; 6 ramps) 27,047 (25,079 to 30,882; 5 ramps)
Sustained GB/day, median 1,879 2,295
Ingest lag p50 / p99 at the ceiling 4.7 s / 9.1 s 5.2 s / 25.0 s
Search p95 at peak, hunt / triage 676 ms / 245 ms 959 ms / 513 ms
What saturated first Ingest: Vector at 62 to 78% CPU on the ingest hosts (72 to 78% in 5 of 6 ramps) The smaller ClickHouse shard (8 vCPU): 86 to 95% CPU, with ingest at 64 to 77%
Cluster cost per hour (spot market / our bid) $0.81 / $1.43 $0.89 / $1.59
Cost per GB/day sustained $0.0104 per GB $0.0093 per GB
Ceiling results, nano v2.0.29, medians over every ramp that reached a passing step

No confirmed loss. Past its ceiling, nano doesn't drop events or fall behind: the ingest tier pushes back on the senders, so they slow down and lag stays flat. Every step that failed, failed on "keeps up" or on lag, never on loss. Two steps were still draining in-flight events at the 60 s check. We recounted one afterwards and every event and marker was stored exactly once; the other's ClickHouse host was lost before it could be recounted.

Adding a fourth ingest host moved the ceiling about 22%. With 3 hosts, ingest saturated first. With 4, the smaller ClickHouse shard became the limit.

Read this before you quote a number

What went wrong, and what it means.

01 · Two installs

The ceiling medians pool two fresh installs.

Partway through, five of the cloud VMs froze at the same moment across all three clusters and never came back. It was a problem in the provider's infrastructure, not in nano: the load on every host was ordinary, the freezes hit three separate clusters at the same instant, and the provider still showed the VMs as running while they were unreachable. We replaced them, wiped every host and reinstalled all three clusters from scratch, then reran both SOC days in full and every ceiling ramp. The five ceiling ramps that finished before the freeze are kept, so the ceiling medians cover ramps from both installs: 6 with 3 ingest hosts and 5 with 4.

02 · Smaller shard

ClickHouse shard 2 ran on an 8 vCPU host.

The provider had no second 16 vCPU host available that day, on spot or on-demand. Shard 2 ran on an 8 vCPU / 64 GB host instead, and that shard is what capped the 4-ingest-host ceiling. The v2.0.25 runs used two matching 16 vCPU shards.

03 · Costs

Costs are spot prices.

Every cost is the spot market price of the nano hosts at run time; the load generators are excluded. We also record our bid, the most we could have paid. On-demand or your own hardware will cost more or less; the hardware shapes are listed below so you can price them.

04 · Version

These are v2.0.29 numbers.

Stock enterprise bundle, fresh licensed installs, no tuning. The workload is synthetic but realistic; your mix of sources and queries will differ, so rerun the kit with your own. Not measured this time: real-time rules (the set is 200 scheduled rules) and more than 2 agents (Agent Lane allows 2 active tokens on this ClickHouse size).

Method

How we measure.

nano is a SIEM, so we measure what a security team feels: does it keep up, how long until an event is searchable, how long until a detection fires, do hunts stay fast under load, and what does it cost. The full method is in METHODOLOGY.md.

Ingest to searchable

For every stored event, the time it was inserted into ClickHouse minus its event timestamp. The generator stamps events as it sends them, so this covers the network, Vector, parsing, batching and the insert.

Zero loss

Senders count every event the ingest tier acknowledged, and each batch carries one sequence-numbered marker. A run passes only if the stored count is at least the acknowledged count and every marker is stored exactly once.

Search latency

The server-side execution time nano reports, and the round trip the load tool sees, per query class: triage (15 min to 1 h), hunting (6 to 24 h) and retro IR (up to 7 days). Analysts pause 5 to 20 s between queries to read results.

Event to alert

nano writes every detection as an OCSF Detection Finding. For each one, the time it was stored minus the timestamp of the newest event it covers.

Headroom

CPU, memory and disk on every host, sampled every 5 s. Headroom is 100 minus the p95 for CPU, or the maximum for memory and disk.

Cost

The spot market price of the nano hosts at run time. Cost per GB is the cluster's daily cost divided by the GB/day it sustained.

Data and hardware

The dataset is a seeded generator.

We don't publish a pile of logs. The dataset is log-blaster with a fixed seed (20261114): a simulated company of 5,000 hosts (20,000 for the ceiling) with users, IP plans and process trees. The same seed always produces the same fleet and the same events. GB/day is measured, not estimated: the raw bytes nano stored, divided by elapsed time. Events average about 1.1 KB raw.

SourceShareOCSF classes
Windows Sysmon 60% Process, Network, File activity
Windows Security events 10% Authentication
Web proxy 12% HTTP Activity
Apache access logs 10% HTTP Activity
AWS CloudTrail 8% API Activity
Generated sources

Every event goes through the real ingest path: HTTP to nginx and Vector on an ingest host, a per-source ingest token, the source's OCSF parser, then ClickHouse. Nothing is written directly to the database. The install is what a customer runs: the signed enterprise bundle, one inventory file, private TLS between hosts and the installer on each host for its role, on Ubuntu 24.04 spot-priced cloud VMs in one US region.

SOC day cluster

RoleHostsVM classvCPU / RAM
app 1 gp.large 4 / 16 GB
ClickHouse 1 mh.xlarge 8 / 64 GB
ingest 1 gp.xlarge 8 / 32 GB
load generator (not in the cost) 1 ch.large 4 / 8 GB

Ceiling cluster

RoleHostsVM classvCPU / RAM
app (UI, API, search, detections, PostgreSQL, Keeper) 1 gp.xlarge 8 / 32 GB
ClickHouse shard 1 (also Keeper) 1 gp.2xlarge 16 / 64 GB
ClickHouse shard 2 (also Keeper) 1 mh.xlarge 8 / 64 GB
ingest (Vector behind nginx) 3, then 4 ch.xlarge 8 / 16 GB
load generators (not in the cost) 2 gp.xlarge 8 / 32 GB
Reproduce it

Run every number on your own hardware.

The kit is open source under the same license as nano's open core: the generator and its seed, the query corpus, the detection set, the topology, the run scripts and the raw JSON from these runs. We don't publish comparisons with other products; the method and data are public, so you can.

Start with nano.

From US$149/month at 5 GB/day, US$279/month at 10 GB/day (list price, billed monthly; 10% off billed annually). Every feature in one license, and one license covers two installs.

Buy nano
Talk to sales