Task API
Nomad's Task API provides every task managed by Nomad with a Unix Domain Socket (UDS) to access the local agent's HTTP API. Regardless of agent configuration the Task API does not require mTLS, but always requires authentication. See below for details.
The Unix Domain Socket is located at ${NOMAD_SECRETS_DIR}/api.sock
.
Rationale
Nomad's HTTP API is available on every agent at the configured
bind_addr
. While this is convenient for user access, it is not
always accessible to workloads running on Nomad. These workloads may have a
network configuration that makes it impossible to access the agent HTTP
address, or the agent's HTTP address may be difficult for workloads to discover
in a way that's portable between Nomad nodes and clusters.
A Unix Domain Socket is a way to expose network services that works with most runtimes and operating systems and adds minimal complexity or runtime overhead to Nomad.
Security
Unlike the agent's HTTP API, the Task API always requires authentication even if ACLs are disabled. This allows Nomad to always make the Task API available even if the workload is untrusted. If ACLs are enabled, the anonymous policy is not available via the Task API.
Both ACL Tokens and Workload Identities are accepted. Once the Task API has authenticated the credentials, the normal endpoint-specific authorization is applied when ACLs are enabled.
The Workload Identity should be used by tasks accessing the Task API.
An ACL Token should be used when an operator is accessing the Task API via
nomad alloc exec
or when a task is proxying Nomad HTTP requests
on behalf of an authenticated user. The Task API could be used by a proxy
presenting Nomad's UI with a standard TLS certificate for browsers.
If task.user
is set in the jobspec, the Task API will only be
usable by that user. Otherwise the Unix Domain Socket is accessible by any
user.
mTLS is never enabled for the Task API since traffic never leaves the node.
Using the Task API
The following jobspec will use the Task API to set [Dynamic Node Metadata][dnm] and exit.
If the job was able to run successfully after about 10 seconds you can observe the outcome by searching for the updated Node's metadata: