udev-manager

Kubernetes device plugin that exposes udev-managed devices (disk partitions, network bandwidth, RDMA) as allocatable resources.

Quick start

1. Build an image for deployment

docker build -f udev-manager.Dockerfile --network=host -t <tag> .

2. Deploy as a DaemonSet with a config file:

domain: 'ydb.tech'
partitions:
  - matcher: 'ydb_disk_(.*)'

3. Matching partitions appear as allocatable resources on the node:

Allocatable:
  ydb.tech/part-ssd_01:  1
  ydb.tech/part-ssd_02:  1
  ydb.tech/part-ssd_03:  1

Configuration

The binary accepts a --config flag with one of:

• file:<path> — read from a file
• env:<VAR> — read from an environment variable
• stdin — read from standard input

Config reference

Field	Type	Description
domain	string	Required. Resource domain (e.g. ydb.tech).
disable_topology_hints	bool	Disable NUMA topology hints for partition devices.
health_check_port	uint16	Port for /healthz endpoint (default: 8080).
partitions	list	Expose each matching partition as its own resource.
batchPartitions	list	Group matching partitions into a single resource.
networkBandwidth	list	Expose network bandwidth shares as resources.
networkRdma	list	Expose RDMA device resources.
numaAffinity	list	Expose statically-configured resources that fake NUMA affinity.

Partitions

Each partition matching the regexp becomes its own Kubernetes resource named {domain}/part-{label}, where {label} comes from the first capture group.

partitions:
  - matcher: 'ydb_disk_(.*)'
  - matcher: 'ssd_(.*)'
    domain: storage.example.com  # optional domain override

Batch partitions

All partitions matching the regexp are grouped into a single resource. A pod requesting one unit receives device specs and env vars for every matching partition at once. This is useful when a workload must own a full set of disks (e.g. a striped volume).

batchPartitions:
  - name: nvme-data          # resource: {domain}/batch-nvme-data
    matcher: 'nvme_data_.*'
    count: 2                  # up to 2 pods can hold this simultaneously
  - name: ssd-wal
    matcher: 'ssd_wal_.*'    # count defaults to 1 (exclusive access)

Network bandwidth

Exposes bandwidth shares for a network interface. Each share represents mbpsPerShare Mbps.

networkBandwidth:
  - matcher: '(^eth0$)'
    mbpsPerShare: 1000

Network RDMA

Exposes RDMA character devices for a network interface.

networkRdma:
  - matcher: '(^ib0$)'
    resourceCount: 4

Numa affinity

Exposes statically-configured resources that exist purely to fake NUMA affinity. Unlike the other types these resources are not backed by udev: the devices are created directly from configuration and are always healthy. Each device advertises a NUMA topology hint for the configured node, so the kubelet topology manager co-locates a requesting pod with that node. Allocating one mounts nothing and sets no env vars.

numaAffinity:
  - name: node0              # resource: {domain}/numa-node0
    numaNode: 0              # NUMA node the devices report affinity to
    count: 4                 # number of devices (default 1)
  - name: node1
    numaNode: 1              # count defaults to 1
    domain: accel.example.com  # optional domain override

Development

Requires Docker for building and testing (the project depends on libudev, which is Linux-only).

make build   # compile
make test    # run tests with -race

Examples

See the examples/ directory for sample configurations.