Storage requirements - Poolside
Partitioning overview
This page shows recommended partition sizing for single-node Poolside on-premises model deployments running RKE2 and explains how model checkpoints affect storage requirements. The checkpoint estimates cover these model artifacts:
| Model | Checkpoint size (GB) |
|---|---|
| laguna_m_fp8_fp8kv_re_06_2026 | 214 |
| laguna_xs_fp8_fp8kv_re_06_2026 | 33 |
Partition recommendations for STIG-hardened hosts
For STIG-hardened Linux nodes, Poolside recommends the following partition layout and sizing as a minimum for the listed models:
| Mount | Size | Filesystem | Mount options | Purpose |
|---|---|---|---|---|
/boot/efi |
600M | fat32 |
umask=0077 |
UEFI system partition |
/boot |
1G | xfs |
nodev,nosuid |
Kernel, initramfs |
/ |
50G | xfs |
defaults | OS root |
/tmp |
2G | xfs |
nodev,nosuid,noexec |
Temporary files |
/var |
5G | xfs |
nodev,nosuid |
Cache, spool, backups |
/var/tmp |
10G | xfs |
nodev,nosuid,noexec |
Temporary files |
/var/log |
50G | xfs |
nodev,nosuid,noexec |
System + audit logs |
/var/lib |
400G | xfs |
nodev,nosuid |
kubelet, rancher, containers |
/home |
50G | xfs |
nodev,nosuid |
Terraform working directories |
/opt/poolside |
350G | xfs |
nodev,nosuid |
Model artifact storage |
| swap | 4G | swap |
none | Swap space |
The total recommended disk space for the listed model artifacts is 1 TB.
Dedicated /opt/poolside partition for model artifacts
STIG controls recommend dedicated partitions to prevent system partitions from filling up and causing denial of service. Poolside recommends placing model artifact storage on a dedicated /opt/poolside partition rather than the root filesystem. Without separation, model artifact growth can render the OS unusable.
STIG mount option recommendations
/tmpand/var/tmpshould includenoexecin the mount options to reduce the risk of executing files from temporary directories./var/logshould includenoexecin the mount options to reduce the risk of executing files from log directories./var/libcannot includenoexec. The container runtime runs binaries from overlay layers stored there. Addingnoexechere breaks Kubernetes pods.
Inference storage impact
When you run inference workloads, each inference pod loads model checkpoints from SeaweedFS into a storage-backed emptyDir volume under /var/lib/kubelet. This means you store at least two copies of the model checkpoint on disk: one copy in SeaweedFS and one copy in the emptyDir volume for each running pod. Each inference workload consumes disk space equal to its model checkpoint size:
| Model | Checkpoint size (GB) | emptyDir consumption (GB) | Total disk usage (GB) |
|---|---|---|---|
| laguna_m_fp8_fp8kv_re_06_2026 | 214 | 214 | 428 |
| laguna_xs_fp8_fp8kv_re_06_2026 | 33 | 33 | 66 |
Plan for sufficient free space in /var/lib to accommodate your largest model’s checkpoint size for each running inference pod.
Scaling prioritization guidance
If you need to scale storage, allocate additional space in this priority order:
/var/lib: Scales per model deployed in-cluster/opt/poolside: Scales with model checkpoints/home: Scales with Terraform files. The Poolside Terraform bundle packages container upgrade images.
Core OS partitions (/, /boot, /tmp, /var, /var/tmp, /var/log) can remain fixed regardless of total disk size.