Task volumes
Every task gets its own volume, sized to what it asks for and deleted after it.
- One
PersistentVolumeClaimper task, shared by all its containers. - Sized by the task's
resources.disk_gb. OvermaxSizeGi, the request gets400before any Pod starts. - Owned by the task's Job, so Kubernetes deletes both together.
Settings
All keys are under taskpods.persistence:
| Key | Default | What it does |
|---|---|---|
defaultSizeGi | 10 | Size when the task asks for none. |
maxSizeGi | 100 | Largest a task may ask for. The schema refuses an install without it. |
storageClass | cluster default | "" omits it (cluster default); "-" sets storageClassName: "" (no dynamic provisioning, needs a pre-bound PV); any other value names a StorageClass. |
accessModes | [ReadWriteOnce] | Access modes of every task volume. |
retainAfterSeconds | 3600 | ttlSecondsAfterFinished of the task's Job. null keeps the Job and volume until deleted by hand. |
labels | {} | Labels on every task volume: cost allocation, backup selectors, data classification. |
annotations | {} | Annotations on every task volume, e.g. for Velero or Kasten. |
taskpods:
persistence:
defaultSizeGi: 20
maxSizeGi: 500
storageClass: fast-ssd
labels: {backup: velero}A task runs on one node, so ReadWriteOnce is enough. If your storage only offers ReadWriteMany (NFS, CephFS), set accessModes: [ReadWriteMany].
Task paths
The volume is mounted at /transfer in every container but ack: there TIF stages inputs and TOF collects outputs. Executors also see the volume at the paths the task declares: each declared path is a mount of the same place on the volume, so a file an executor writes under /data is the file TOF uploads.
What is mounted:
volumesandDIRECTORYinputs and outputs: the path itself.FILEinputs: the file itself, a single-file mount. It can be read and written in place, but not renamed or replaced (mv tmp in.txtfails), and nothing can be created below it.FILEoutputs: their directory.- Outputs with wildcards (
FILEorDIRECTORY): the directory before the first wildcard (/datafor/data/*/x.txt), whateverpath_prefixis.
A path below another mounted path is covered by that mount. Only data under these paths persists between executors and reaches TOF; anything else an executor writes stays in its own container and is lost.
Each of these directories exists, empty, before the first executor runs: the kubelet creates the mount points, and TIF creates the ones below another mounted path (a volume /mnt/task/usage when /mnt/task is mounted), after staging inputs. Both get the mode of the volume root, so an executor running as another user than TIF can write there. Paths that already exist, such as staged inputs, are left as they are.
A task is refused with 400 Bad Request if a path:
- is relative;
- has a
..component (.and repeated slashes are fine); - would need a mount at
/(for example/as a volume, or a wildcard in the first component,/*.txt); - is in
/transfer, where Poiesis stages task files; - is below a
FILEinput; - is at or below an operator mount (
taskpods.extraVolumeMounts), such as a read-only reference volume.
Reference data
Operators can mount shared data, such as a read-only reference genome, into every task with taskpods.extraVolumes and taskpods.extraVolumeMounts. The mounts reach executors too:
taskpods:
extraVolumes:
- name: ref-grch38
persistentVolumeClaim:
claimName: ref-grch38-ro
readOnly: true
extraVolumeMounts:
- name: ref-grch38
mountPath: /ref/grch38
readOnly: trueThe names task-data, postgres-ca and tmp are reserved; the API refuses to start with them. See Helm values.
Cleanup and OpenShift
The volume's ownerReference to its Job sets blockOwnerDeletion: false, so the API's Role needs no jobs/finalizers permission, which OpenShift project admins can't grant. Kubernetes' garbage collector deletes the volume when the Job goes.
