NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
Go modules, #320 by repository stars
Latest stable
v1.14.6
Last release today
20 Nov 2026
Ships on a steady schedule
a new release about every 8 days
Nearly every release is documented
notes for 60 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
9 years old
3513 releases · first in 2018
Drag across the ruler to compare, or tap a release to open it.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
pkg/machinery/v1.15.0-alpha.2
pkg/machinery/v1.15.0-alpha.2
Welcome to the v1.15.0-alpha.2 release of Talos!
This is a pre-release of Talos
Please try out the release binaries and report any issues at
https://github.com/siderolabs/talos/issues.
Talos now waits for the partition it was booted from to be discovered before declaring the system volumes (META, STATE) missing
and entering maintenance mode, so that disks enumerated late (SD cards, USB, some RAID controllers) are not mistaken for a fresh machine.
On UEFI systems booted via systemd-boot the boot partition comes from the LoaderDevicePartUUID EFI variable, while the GRUB configuration
now passes the talos.boot.partuuid=<UUID> kernel argument (probed by GRUB at boot). The detected boot partition is published as the bootpartition resource.
ContainerConfig documents can now cap the swap a container may use, independently of its memory limit:
apiVersion: v1alpha1
kind: ContainerConfig
name: collector
image: ghcr.io/example/collector:v1.0.0
resources:
limits:
swap: "0"The value maps onto cgroup v2 memory.swap.max. Zero keeps the container out of swap entirely, so its memory ceiling is enforced by the OOM killer rather than by swapping; unset leaves the node's swap available to it.
Talos now generates containerd configuration with version = 4 (previously version = 3).
containerd only migrates configuration from older versions, so any custom CRI configuration provided via
CRICustomizationConfig documents (or shipped by system extensions) is now interpreted as version 4, and settings
which moved in version 4 are silently ignored (only a Ignoring unknown key in TOML warning is logged by containerd).
In version 4, the top-level server sections moved into plugin configuration blocks, so the following patches need to be adjusted:
| version 3 | version 4 |
|---|---|
[metrics] address |
[plugins."io.containerd.server.v1.metrics"] address |
[metrics] grpc_histogram |
[plugins."io.containerd.metrics.v1.grpc-prometheus"] grpc_histogram |
[grpc] address, uid, gid, max_recv_message_size, max_send_message_size |
[plugins."io.containerd.server.v1.grpc"] (same keys) |
[grpc] tcp_address, tcp_tls_ca, tcp_tls_cert, tcp_tls_key, tcp_tls_common_name |
[plugins."io.containerd.server.v1.grpc-tcp"] address, tls_ca, tls_cert, tls_key, tls_common_name |
[ttrpc] address, uid, gid |
[plugins."io.containerd.server.v1.ttrpc"] (same keys) |
[debug] address, uid, gid |
[plugins."io.containerd.server.v1.debug"] (same keys) |
For example, the containerd metrics customization documented for previous versions of Talos:
apiVersion: v1alpha1
kind: CRICustomizationConfig
name: enable-metrics
content: |
[metrics]
address = "0.0.0.0:11234"should now be written as:
apiVersion: v1alpha1
kind: CRICustomizationConfig
name: enable-metrics
content: |
[plugins."io.containerd.server.v1.metrics"]
address = "0.0.0.0:11234"Note that [debug] level, format and log_trace_id stay at the top level, and that the ttrpc address is no longer
derived from the grpc address: if the grpc address is customized, the ttrpc plugin address has to be set explicitly.
talosctl containers, logs, stats and restart now select the containerd namespace through the same --namespace flag and vocabulary already used by talosctl image and talosctl debug: system (the default), cri for Kubernetes workloads, and taloscontainers for containers declared via a ContainerConfig document.
The --kubernetes/-k flag is deprecated in favor of --namespace cri.
talosctl image list also supports --namespace taloscontainers, to inspect images pulled for ContainerConfig containers. talosctl debug does not support the taloscontainers namespace.
CPU frequency scaling can now be configured declaratively with the CPUScalingConfig document, instead of through kernel arguments
or raw SysfsConfig keys:
apiVersion: v1alpha1
kind: CPUScalingConfig
name: all-cores
selector:
match: true
governor: performanceSettings apply to the cpufreq policies matched by the selector, which targets cores by property rather than by index, so a machine
whose cores are not uniform can be configured without hardcoding how many it has:
apiVersion: v1alpha1
kind: CPUScalingConfig
name: performance-cores
selector:
match: cpu.core_type == "performance"
governor: performance
---
apiVersion: v1alpha1
kind: CPUScalingConfig
name: efficiency-cores
selector:
match: cpu.core_type == "efficiency"
governor: powersaveThe selector matches against the fields of the CPUScalingStatus resource, and may also set energyPerformancePreference,
minFrequencyKhz and maxFrequencyKhz. Where several documents match the same policy, the one whose name sorts first wins, and the
document which claimed a policy is reported on the resulting CPUScalingSpec resource. Removing the configuration restores the
values the machine booted with.
Setting a cpufreq attribute through both CPUScalingConfig and SysfsConfig is rejected, as the two would otherwise overwrite each
other continuously.
Talos can now collect SMART (Self-Monitoring, Analysis and Reporting Technology) health information from disks,
exposed via the new SMARTStatus resource (one per disk, keyed by the disk device name).
Collection is controlled by the presence of the new DiskSMARTConfig document: while the document is part of
the machine configuration SMART status is collected, and removing the document disables collection (any
SMARTStatus resources already produced are removed as well). There is no enabled field.
apiVersion: v1alpha1
kind: DiskSMARTConfig
interval: 30m0stalosctl gen config and talosctl cluster create emit this document by default, so SMART status collection
is enabled by default for machine configurations generated by Talos 1.15. Clusters upgraded from an earlier
version keep their existing machine configuration, which carries no DiskSMARTConfig document, so SMART
status collection stays off there until the document is added.
Rotational disks are never spun up just to be probed: a disk in standby is detected via ATA
CHECK POWER MODE and its SMART data is left untouched until it is active again.
To retrieve SMART statuses:
talosctl get smartThe resource reports an overall health verdict, temperature, power-on hours and power cycles, NVMe wear
indicators (percent used, available spare, critical warning, media errors) and, for ATA/SATA disks, the raw
SMART attribute table with per-attribute thresholds.
A new talosctl diskfree command (aliased to df) reports storage and inode usage for mounted volumes.
The EthernetConfig document now supports configuring Ethernet link flow control (802.3x pause frames)
with the new flowControl field, similar to the ethtool -A command:
apiVersion: v1alpha1
kind: EthernetConfig
name: eno1
flowControl:
rx: true # accept incoming pause frames
tx: false # don't transmit pause frames
autoneg: true # autonegotiate flow controlOmitting the flowControl field leaves the link's flow control configuration unchanged.
Current flow control settings are reported on the EthernetStatus resource.
ExtensionServiceConfig is supported only for extension services whose service manifest explicitly declares a configuration dependency:
depends:
- configuration: trueUsing ExtensionServiceConfig with a service that does not declare this dependency is unsupported and has undefined startup behavior.
The service might start before its configuration is rendered, leaving config files or environment variables unavailable until a later service restart.
Extension authors must declare the dependency before documenting ExtensionServiceConfig support.
The following talosctl gen subcommands have been removed:
talosctl gen catalosctl gen crttalosctl gen csrtalosctl gen keytalosctl gen keypairThese commands where historically used to generate pieces of Talos PKI, but they are no longer needed
as talosctl gen config and talosctl gen secrets now generate all required PKI material automatically.
A virtual machine disk can now be provisioned as an empty volume in a storage pool, rather than only
as a CD-ROM attached from a content library:
apiVersion: v1alpha1
kind: VirtualMachineConfig
name: vm1
disks:
- name: data
pool: pool1
size: 100GiB
format: qcow2
provision:
blank: {}The pool must be declared by a StoragePoolConfig document in the same machine configuration; a disk
naming an undeclared pool is now a configuration error rather than a virtual machine that silently
never starts.
Four behaviours are worth knowing before you rely on this:
talosctl get storagepoolvolumestatuses listssize grows the volume the next time no virtual machine has it open,size is reported and ignored. Shrinking would truncate a filesystem the guest laidStoragePoolConfig document. Removing aProvisioning a disk from a content library image is not implemented yet; provision.fromImage is
still only supported on a cdrom.
Setting ignoreCtrlAltDelete: true in the SecurityProfileConfig document makes Talos log and ignore Ctrl-Alt-Delete instead of rebooting the machine.
Ctrl-Alt-Delete still reboots the machine until a configuration with this option is applied.
Following Kubernetes deprecation of IPVS support (kube-proxy), Talos Linux now ships IPVS as kernel modules which are not loaded by default.
If you require IPVS support, you can enable it by loading the following kernel modules:
ip_vs (core)ip_vs_rr, ip_vs_wrr, ip_vs_lc, ip_vs_sh or ip_vs_mh (scheduling algorithms)xt_ipvs (nftables/iptables match)The kernel can once again ask Talos to load a module on demand (request_module()), e.g. when a driver loads a helper module while probing a device,
or when DRBD loads its transport module.
Talos handles the kernel's modprobe usermode helper call and loads the module (shipped with Talos or a system extension) by the requested name,
applying the parameters from KernelModuleConfig if the module is listed in the machine config.
Such modules no longer need to be listed in the machine config, and the order of the listed modules is not important.
Requests via subsystem aliases (e.g. fs-*, net-pf-*, crypto-*), which userspace can trigger via mounts, sockets, etc., are refused,
so on-demand loading can't be used to load arbitrary modules from workloads.
The outcome of each request is logged to the kernel log (talosctl dmesg).
The KmsgLogConfig document now supports an extraTags field which attaches additional key-value pairs
to every kernel log message sent to that destination:
apiVersion: v1alpha1
kind: KmsgLogConfig
name: apiSink
url: https://kmsglog.api/logs
extraTags:
cluster: staging-west
node: worker-1Tags are scoped to the destination they are configured on, so different endpoints can receive different tags.
The keys facility, seq, clock, priority, msg, talos-time and talos-level are reserved and rejected,
as they are set by Talos itself.
Talos now defaults kubelet imageGCHighThresholdPercent to 75 and imageGCLowThresholdPercent to 65 (upstream defaults are 85 and 80).
The upstream image GC high threshold matches the default imagefs.available<15% hard eviction threshold, so the node was put
into DiskPressure (and tainted for evictionPressureTransitionPeriod) before image GC had a chance to run.
The defaults are applied only if neither of the thresholds is set in machine.kubelet.extraConfig.
On upgrade, nodes with disk usage above 75% will garbage collect unused images.
Talos now validates the kubelet CPU manager and memory manager state files (/var/lib/kubelet/cpu_manager_state, /var/lib/kubelet/memory_manager_state)
the same way kubelet does on startup, and removes a state file kubelet would refuse to load before starting kubelet.
Previously only a change of the manager policy was detected, so changing e.g. reservedSystemCPUs or reservedMemory, or
a change of the CPU/NUMA topology of the machine, left kubelet unable to start until the state file was removed manually.
Only the rendered KubeletConfiguration (machine.kubelet.extraConfig) is checked: the resource manager settings passed as kubelet
command line flags via machine.kubelet.extraArgs (e.g. --cpu-manager-policy, --reserved-cpus) are not taken into account.
KubeSpan now supports filtering the endpoints received from other peers via filters.peerEndpoints in the KubeSpanConfig document.
While filters.endpoints filters the addresses a node advertises to the whole cluster, filters.peerEndpoints filters the endpoints
received from other peers before the node connects to them, affecting only that node. This allows excluding peer-advertised endpoints
which are known to be unreachable from a node (e.g., addresses of a private network the node is not connected to), so that KubeSpan
endpoint rotation never attempts them.
Talos now listens for LLDP on physical Ethernet interfaces and reports neighbors through the LLDPNeighborStatus resource.
Reported information includes chassis and port IDs, system and port descriptions, system names, management addresses, and VLANs.
The built-in listener does not transmit advertisements; use the lldpd system extension when advertising is required.
Secure Boot images no longer have lockdown=confidentiality enabled by default.
This change was made to improve compatibility with eBPF tooling under default schematic.
This means that Secure Boot images will now have lockdown=integrity enabled by default (implicitly), which is the recommended setting for most users.
Users can override it by adding lockdown=confidentiality to the kernel command line through Image Factory if they require it.
Talos Linux now supports creating MACVLAN links with a new MacVLANConfig document:
apiVersion: v1alpha1
kind: MacVLANConfig
name: eth0.macvlan
parent: eth0
mode: bridge # optional, defaults to `bridge`
addresses:
- address: 192.168.1.100/24The source mode is not supported yet, as it requires a list of allowed source MAC addresses.
The multipath-tools system extension now reads /etc/multipath.conf from the Talos host instead of using ExtensionServiceConfig.
The multipathd service waits for this file and bind-mounts it read-only into its service container.
Before updating the extension, apply a machine config patch that deletes the existing ExtensionServiceConfig document and adds an EtcFileConfig document:
apiVersion: v1alpha1
kind: ExtensionServiceConfig
name: multipathd
$patch: delete
---
apiVersion: v1alpha1
kind: EtcFileConfig
name: multipath.conf
mode: 0o644
contents: |
defaults {
user_friendly_names yes
find_multipaths no
path_selector "round-robin 0"
}The extension does not provide a default configuration, so multipathd remains waiting until /etc/multipath.conf is present.
Talos now supports NFS as an ExternalVolumeConfig filesystem type alongside virtiofs, allowing nodes to mount network exports at /var/mnt/<name> through the existing block volume machinery.
apiVersion: v1alpha1
kind: ExternalVolumeConfig
name: shared
filesystemType: nfs
mount:
nfs:
server: 10.0.0.10
path: /export
version: "4.1"Supported version values are 3, 4, 4.1, and 4.2.
The mount.nfs spec configures the kernel NFS client:
server (hostname or IP; IPv6 is bracketed automatically) and absolute path are required; optional port, transport (tcp/udp), and NFSv3-only mountPort/mountTransport (otherwise rpcbind discovery is used).locking (local default / remote) controls nolock/lock; Talos does not run rpc.statd, so local (nolock) is the safe default. You need rpc.statd extension for remote locking.recovery (hard/soft/soft-error) maps to hard/soft/softerr.timeout (timeo, deciseconds), retransmissions (retrans), readSize/writeSize (rsize/wsize, 1024–1048576 multiples of 1024), connections (nconnect, 1–16, TCP only), reservedPort (resvport/noresvport), and security (none/sys; Kerberos flavors are unsupported as Talos does not run rpc.gssd).tcp transport.Common mount flags (readOnly, disableAccessTime, secure) apply as with other volume types.
Validation rejects mountPort/mountTransport/locking outside NFSv3 and udp for NFSv4.
Talos reports host NUMA nodes, CPU membership and total node memory through the NUMATopology resource.
Explicit VM CPU pins and memory-node selections are checked against the host before starting or replacing a guest.
Invalid placement leaves an existing guest running; CPUs and guest memory may intentionally use different nodes.
This validates placement, not available memory capacity or reservations. Host inventory is not exposed in container mode.
RouteSpec resource IDs now match the key the Linux kernel uses for routes: routing table, address family, destination and priority (metric),
e.g. inet4/10.0.0.0/8/1024 instead of inet4/10.0.0.1/10.0.0.0/8/1024.
The gateway, the output link and the multipath next-hops are no longer part of the ID, so changing them updates the route in place
(RTM_NEWROUTE with NLM_F_REPLACE) without the route ever being missing, e.g. when a BGP route switches between a single next-hop and ECMP.
Routes which differ only in the next-hops (which can't be installed into the kernel at the same time) are now merged into a single route,
with the higher configuration layer taking precedence.
RouteStatus resource IDs are not changed.
talosctl now writes every file which might contain secrets with 0600 permissions (owner read/write only),
and it restricts the permissions of the files it overwrites as well.
This affects the machine configuration (talosctl gen config, talosctl machineconfig patch,
talosctl cluster create --skip-injecting-config), talosconfig and kubeconfig.
The TimeService/TimeCheck API (talosctl time --check) is no longer available to the os:reader role, as it allows querying
arbitrary time servers from the node. The TimeService/Time API (querying the configured time server) is still available to os:reader.
PTP device paths are now strictly validated both in the machine configuration and in the API: the path should be directly under /dev,
the name should start with ptp (e.g. /dev/ptp0, /dev/ptp_kvm), and the device should be a PTP character device.
The FilesystemTrimConfig document (and the per-volume trim settings) now support chunkSize, chunkDelay and minLength options.
By default, the whole filesystem is trimmed at once (same as the fstrim command), which issues discards for all free space back-to-back
and might cause latency spikes for other workloads on the same disk (e.g. etcd).
Setting chunkSize splits the trim into multiple operations with chunkDelay in between, while minLength skips free ranges smaller than the specified size.
Linux: 7.3
Kubernetes: 1.38.0-alpha.2
containerd: 2.3.7
runc: 1.5.2
CoreDNS: 1.14.7
systemd-udevd: 262
etcd: 3.7.3
Talos is built with Go 1.27.3.
New talosctl hypervisor start|stop|reboot <vm-name> commands control the power state of Talos-managed virtual machines.
start and stop override the virtual machine's power state until the node reboots, without changing the machine configuration.
After a reboot, the virtual machine follows the powerState of its VirtualMachineConfig again.
reboot restarts a running guest and leaves the power state unchanged.
stop and reboot ask the guest to shut down or restart by default. Pass --graceful=false to destroy the domain instead.
The commands require the os:admin or os:operator role.
VirtualMachineConfig now accepts a profile field selecting the guest-visible CPU and machine type.
Talos ships three builtin profiles on amd64:
host (default when profile is omitted): q35 machine, host-model CPU.passthrough: q35 machine, host-passthrough CPU.legacy: pc-i440fx-9.2 machine, kvm64 CPU, for guests that need a traditional PC chipset.Custom profiles are defined with the new VirtualMachineProfileConfig document:
apiVersion: v1alpha1
kind: VirtualMachineProfileConfig
name: rack-a-compat
machine: pc-q35-9.2
cpu:
mode: custom
model: Haswell-noTSX-IBRS
features:
- name: pku
policy: disablemachine accepts q35, pc, pc-q35-* or pc-i440fx-*. cpu.mode is host-model, host-passthrough or custom;
model, fallback (forbid by default, or allow) and features (policies require, disable, optional, force, forbid)
are only allowed with custom. The names host, legacy, passthrough and the compat- prefix are reserved.
Profiles are validated structurally only: Talos does not check that the host CPU supports the model, nor guarantee migration compatibility.
Profile changes, including switching a VM to a different profile, are not applied to running guests;
they take effect on the next deliberate stop and start of the virtual machine.
Hardware profiles are currently available only on amd64.
Talos-managed virtual machines now expose their serial console over the Talos API, with no host shell access required.
talosctl hypervisor console <vm-name> attaches to the live serial console of a running VM.
Press Ctrl-] to detach without stopping the VM; Ctrl-C is forwarded to the guest.
Only one session can be attached to a VM at a time, and previous output is not replayed.
The command is supported on Unix clients only.
Serial output is also persisted on the host (with bounded retention), and can be read with talosctl logs --kind=vm <vm-name>,
including --tail and --follow. The new --kind flag (container, vm or service) selects the log source explicitly;
omitting it keeps the previous service/container routing.
Both the console and the VM logs require the os:admin role and a single target node.
The new talosctl hypervisor vnc <vm-name> command forwards a virtual machine's graphical console to a local loopback TCP port.
Enable console.vnc.enabled in its VirtualMachineConfig; access requires the os:admin role and exactly one target node.
The listener accepts one viewer at a time and supports reconnecting without restarting the command.
Use a viewer that supports unauthenticated RFB.
Talos authenticates the tunnel, and no VNC TCP port is opened on the node.
Talos Linux now supports creating VXLAN links with a new VXLANConfig document:
apiVersion: v1alpha1
kind: VXLANConfig
name: vxlan100
id: 100 # VXLAN network identifier (VNI)
parent: eth0
local: 10.255.0.1 # or `group:` for a multicast group, but not both
port: 4789 # optional, defaults to 4789
learning: false # optional, defaults to true
addresses:
- address: 192.168.1.100/24The new WorkloadResourceConfig document sets independent aggregate memory limits for Kubernetes pods, Talos containers and virtual machines:
apiVersion: v1alpha1
kind: WorkloadResourceConfig
kubepods:
memory:
limit: 16GiB
taloscontainers:
memory:
limit: 4GiB
virtualMachines:
memory:
limit: 32GiBEach root is optional. Limits are ceilings, not reservations, and do not cover swap or HugeTLB memory.
For kubepods, Talos derives systemReserved.memory so kubelet enforces the cap and adjusts Node Allocatable;
an explicit systemReserved.memory, conflicting kubelet flags or the Static memory manager cannot be combined with this limit.
Changing the Kubernetes cap restarts kubelet. Without Kubernetes, only the container and virtual machine limits apply.
Lowering a limit can OOM-kill workloads in that group, including QEMU guests or Kubernetes control-plane pods.
Talos exclusively owns memory.max on /taloscontainers and /virtualmachines.partition, even without this document.
Omitting their limits means unlimited aggregate memory; individual child limits still apply.
Kubelet system-reserved or kube-reserved memory enforcement on these roots is rejected regardless of this document;
CPU-only or compressible enforcement and reservations on unrelated roots remain supported.
Kubelet drop-in directories are rejected while any cap is active; without caps, their contents remain the operator's responsibility.
Removing the Kubernetes cap restores the original kubelet configuration. The document has no effect in container mode.
Note truncated.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
One column per quarter.