open-nomad/website/pages/docs/job-specification/multiregion.mdx

273 lines
8.3 KiB
Plaintext

---
layout: docs
page_title: multiregion Stanza - Job Specification
sidebar_title: multiregion <sup>Beta</sup>
description: |-
The "multiregion" stanza specifies that a job will be deployed to multiple federated
regions.
---
# `multiregion` Stanza
<Placement
groups={[
['job', 'multiregion'],
]}
/>
~> **Enterprise Only!** This functionality only exists in Nomad
Enterprise. This is not present in the open source version of Nomad.
The `multiregion` stanza specifies that a job will be deployed to multiple
[federated regions]. If omitted, the job will be deployed to a single region
— the one specified by the `region` field or the `-region` command line
flag to `nomad job run`.
Federated Nomad clusters are members of the same gossip cluster but not the
same raft cluster; they don't share their data stores. Each region in a
multiregion deployment gets an independent copy of the job, parameterized with
the values of the `region` stanza. Nomad regions coordinate to rollout each
region's deployment using rules determined by the `strategy` stanza.
```hcl
job "docs" {
multiregion {
strategy {
max_parallel = 1
on_failure = "fail_all"
}
region "west" {
count = 2
datacenters = ["west-1"]
meta {
my-key = "my-value-west"
}
}
region "east" {
count = 5
datacenters = ["east-1", "east-2"]
meta {
my-key = "my-value-east"
}
}
}
}
```
## Multiregion Deployment States
A single region deployment using one of the various [upgrade strategies]
begins in the `running` state, and ends in the `successful` state, the
`canceled` state (if another deployment supersedes it before it it's
complete), or the `failed` state. A failed single region deployment may
automatically revert to the previous version of the job if its `update`
stanza has the [`auto_revert`][update-auto-revert] setting.
In a multiregion deployment, regions begin in the `pending` state. This allows
Nomad to determine that all regions have accepted the job before
continuing. At this point up to `max_parallel` regions will enter `running` at
a time. When each region completes its local deployment, it enters a `blocked`
state where it waits until the last region has completed the deployment. The
final region will unblock the regions to mark them as `successful`.
## `multiregion` Parameters
- `strategy` <code>([Strategy](#strategy-parameters): nil)</code> - Specifies
a rollout strategy for the regions.
- `region` <code>([Region](#region-parameters): nil)</code> - Specifies the
parameters for a specific region. This can be specified multiple times to
define the set of regions for the multiregion deployment. Regions are
ordered; depending on the rollout strategy Nomad may roll out to each region
in order or to several at a time.
~> **Note:** Regions can be added, but regions that are removed will not be
stopped and will be ignored by the deployment. This behavior may change before
multiregion deployments are considered GA.
### `strategy` Parameters
- `max_parallel` `(int: <optional>)` - Specifies the maximum number
of region deployments that a multiregion will have in a running state at a
time. By default, Nomad will deploy all regions simultaneously.
- `on_failure` `(string: <optional>)` - Specifies the behavior when a region
deployment fails. Available options are `"fail_all"`, `"fail_local"`, or
the default (empty `""`). This field and its interactions with the job's
[`update` stanza] is described in the [examples] below.
Each region within a multiregion deployment follows the `auto_revert`
strategy of its own `update` stanza (if any). The multiregion `on_failure`
field tells Nomad how many other regions should be marked as failed when one
region's deployment fails:
- The default behavior is that the failed region and all regions that come
after it in order are marked as failed.
- If `on_failure: "fail_all"` is set, all regions will be marked as
failed. If all regions have already completed their deployments, it's
possible that a region may transition from `blocked` to `successful` while
another region is failing. This successful region cannot be rolled back.
- If `on_failure: "fail_local"` is set, only the failed region will be marked
as failed. The remaining regions will move on to `blocked` status. At this
point, you'll need to manually unblock regions to mark them successful
with the [`nomad deployment unblock`] command or correct the conditions
that led to the failure and resubmit the job.
~> For `system` jobs, only [`max_parallel`](#max_parallel) is enforced. The
`system` scheduler will be updated to support `on_failure` when the the
[`update` stanza] is fully supported for system jobs in a future release.
### `region` Parameters
The name of a region must match the name of one of the [federated regions].
- `count` `(int: <optional>)` - Specifies a count override for task groups in
the region. If a task group specifies a `count = 0`, its count will be
replaced with this value. If a task group specifies its own `count` or omits
the `count` field, this value will be ignored. This value must be
non-negative.
- `datacenters` `(array<string>: <optional>)` - A list of
datacenters in the region which are eligible for task placement. If not
provided, the `datacenters` field of the job will be used.
- `meta` - `Meta: nil` - The meta stanza allows for user-defined arbitrary
key-value pairs. The meta specified for each region will be merged with the
meta stanza at the job level.
As described above, the parameters for each region replace the default values
for the field with the same name for each region.
## `multiregion` Examples
The following examples only show the `multiregion` stanza and the other
stanzas it might be interacting with.
### Max Parallel
This example shows the use of `max_parallel`. This job will deploy first to
the "north" and "south" regions. If either "north" finishes and enters the
`blocked` state, then "east" will be next. At most 2 regions will be in a
`running` state at any given time.
```hcl
multiregion {
strategy {
max_parallel = 2
}
region "north" {}
region "south" {}
region "east" {}
region "west" {}
}
```
### Rollback Regions
This example shows the default value of `on_failure`. Because `max_parallel = 1`,
the "north" region will deploy first, followed by "south", and so on. But
supposing the "east" region failed, both the "east" region and the "west"
region would be marked `failed`. Because the job has an `update` stanza with
`auto_revert=true`, both regions would then rollback to the previous job
version. The "north" and "south" regions would remain `blocked` until an
operator intervenes.
```hcl
multiregion {
strategy {
on_failure = ""
max_parallel = 1
}
region "north" {}
region "south" {}
region "east" {}
region "west" {}
}
update {
auto_revert = true
}
```
### Override Counts
This example shows how the `count` field override the default `count` of the
task group. The job the deploys 2 "worker" and 1 "controller" allocations to
the "west" region, and 5 "worker" and 1 "controller" task groups to the "east"
region.
```hcl
multiregion {
region "west" {
count = 2
}
region "east" {
count = 5
}
}
}
group "worker" {
count = 0
}
group "controller" {
count = 1
}
```
### Merging Meta
This example shows how the `meta` is merged with the `meta` field of the job,
group, and task. A task in "west" will have the values
`first-key="regional-west"`, `second-key="group-level"`, whereas a task in
"east" will have the values `first-key="job-level"`,
`second-key="group-level"`.
```hcl
multiregion {
region "west" {
meta {
first-key = "regional-west"
second-key = "regional-west"
}
}
region "east" {
meta {
second-key = "regional-east"
}
}
}
}
meta {
first-key = "job-level"
}
group "worker" {
meta {
second-key = "group-level"
}
}
```
[federated regions]: https://learn.hashicorp.com/nomad/operating-nomad/federation
[`update` stanza]: /docs/job-specification/update
[update-auto-revert]: /docs/job-specification/update#auto_revert
[examples]: #multiregion-examples
[upgrade strategies]: https://learn.hashicorp.com/nomad?track=update-strategies#update-strategies
[`nomad deployment unblock`]: /docs/commands/deployment/unblock