Skip to content

Sizing Profile for Pilot Installs ​

Reference profiles ​

Two reference host sizes for this release.

small-dev ​

Intended for: development, demos, single-tenant POC.

ResourceValue
CPU4 vCPU
RAM16 GB
Disk100 GB SSD
Max tenants1–2
SocTalk control plane reserved~2 GB RAM, 1 vCPU
Per-tenant budget~6–8 GB RAM, 1–1.5 vCPU

Boot times are slower here; <30 min to OSS stack healthy SLO applies.

pilot-prod ​

Intended for: MSSP running real pilot customers, 3–5 tenants.

ResourceValue
CPU8 vCPU
RAM32 GB
Disk500 GB SSD
Max tenants3–5
SocTalk control plane reserved~3 GB RAM, 1–2 vCPU
Per-tenant budget~5–7 GB RAM, 1–1.5 vCPU

Boot times on the <15 min to OSS stack healthy SLO.

Per-tenant footprint (estimates) ​

These are starting-point values for ResourceQuota and LimitRange in the tenant chart. pre-release validation measures actuals; actuals replace these in the final values.

ComponentRAM requestRAM limitCPU requestCPU limitDisk (PVC)
Wazuh manager512 MB1 GB200 m500 m20 GB
Wazuh indexer (OpenSearch fork)2 GB (heap 1 GB)4 GB (heap 2 GB)500 m2000 m50 GB
Wazuh dashboard512 MB1 GB100 m500 m
Filebeat128 MB256 MB50 m200 m
linux-ep (L2 endpoint agent)256 MB512 MB100 m500 m
SocTalk adapter128 MB256 MB50 m200 m
Per-tenant reserved budget~8 GB request, ~16 GB limit~2.2 vCPU request, ~7.7 vCPU limit~120 GB

TheHive and Cortex are external integrations, not bundled subcharts, so they run outside the tenant namespace and are not part of this per-tenant footprint; size them where they are hosted. The bundled in-namespace stack is Wazuh plus the linux-ep agent, so the reserved budget above carries headroom over the current in-namespace pods.

Note: limits are burst ceilings; sustained usage is closer to requests. Running 3 tenants on an 8-vCPU / 32 GB / 500 GB host means:

  • RAM: ~24 GB of requests (fits), ~48 GB of limits (requires careful overcommit tuning).
  • CPU: ~6.6 vCPU of requests (fits with control plane), bursts share total.
  • Disk: ~360 GB of tenant PVCs (fits with margin for control plane + SocTalk DB).

This is why pilot-prod caps at 5 tenants; beyond 5, memory limits start bumping into node capacity even accounting for overcommit.

Max-tenants-per-node formula ​

Approximation:

max_tenants = floor((node_total_RAM - control_plane_RAM - safety_margin) / per_tenant_RAM_request)
  • control_plane_RAM: 2 GB (small-dev) or 3 GB (pilot-prod) for SocTalk + Postgres + ingress controller + Cilium + cert-manager.
  • safety_margin: 10% of node RAM for K8s system pods, CNI, DNS, monitoring.
  • per_tenant_RAM_request: 8 GB baseline.

For 32 GB pilot-prod: floor((32 - 3 - 3.2) / 8) = floor(25.8 / 8) = 3 guaranteed tenants without overcommit. With overcommit, 4–5 is safe for typical alert volumes.

Disk sizing drivers ​

The dominant disk consumer is the Wazuh indexer (stores indexed events). Ingest rate determines growth:

Alert rateDaily index size (rough)Retention 30 daysRetention 90 days
10 alerts/sec sustained~5 GB/day150 GB450 GB
1 alert/sec sustained~500 MB/day15 GB45 GB
100 alerts/day~10 MB/day300 MB900 MB

Tenant PVC sizes in the chart default to 50 GB for the Wazuh indexer; MSSPs override per-tenant for high-volume customers.

Retention policy defaults to 30 days of hot data in indexer; older data is deleted or archived (doesn't implement hot→cold tiering; a future release adds it).

Sizing gates ​

Pre-provisioning check ​

When MSSP operator creates a new tenant, SocTalk controller runs a sanity check:

available_RAM = node.allocatable.memory - sum(ns.resourceQuota.requests.memory for ns in existing_tenant_namespaces) - control_plane_reserve
if (new_tenant.resourceQuota.requests.memory > available_RAM):
    refuse with "insufficient cluster capacity for new tenant"
    or
    prompt MSSP: "this will overcommit; proceed? [y/N]"

This gate is softer in this release (warn rather than hard-fail) since MSSPs may intentionally overcommit for light-use customers.

Per-tenant LimitRange enforcement ​

Every tenant namespace has a LimitRange:

yaml
apiVersion: v1
kind: LimitRange
metadata: { name: tenant-limits, namespace: tenant-acme }
spec:
  limits:
    - type: Container
      default:
        memory: "2Gi"
        cpu: "500m"
      defaultRequest:
        memory: "256Mi"
        cpu: "100m"
      max:
        memory: "6Gi"
        cpu: "2"

Prevents an accidentally-misconfigured pod from requesting 30 GB and starving the node.

Profiles beyond ​

Documented but not validated in this release:

ProfileCPURAMDiskMax tenants
mid-host16 vCPU64 GB1 TB10–15
large-host32 vCPU128 GB2 TB25–30
multi-node cluster3 nodes × large-50+ (a future release multi-install recommended instead)

Recommendation for MSSPs growing past pilot-prod capacity:

  • : add a second host, run a second SocTalk install (schema supports this, tooling is manual).
  • a future release: multi-install automation in Cloud layer.
  • a future release: clustered K3s with proper scheduling across nodes.

Measurement plan (pre-release validation) ​

The spike produces real numbers to replace the estimates in §2:

  1. Deploy soctalk-tenant with one tenant on k3d (dev-harness).
  2. Idle measurement: take kubectl top pod -n tenant-acme snapshot.
  3. Load test: inject 10 alerts/sec for 10 minutes; measure peak.
  4. Stop load; measure ~5 minutes later for "warm-idle" numbers.
  5. Repeat with three tenants in parallel to observe interference.
  6. Update this document's tables with measured values.

Released under the Apache 2.0 License.