{ { < include "compose/deploy.md" > } }

Attributes#

endpoint_mode#

endpoint_mode specifies a service discovery method for external clients connecting to a service. The Compose Deploy Specification defines two canonical values:

  • endpoint_mode: vip: Assigns the service a virtual IP (VIP) that acts as the front end for clients to reach the service on a network. Platform routes requests between the client and nodes running the service, without client knowledge of how many nodes are participating in the service or their IP addresses or ports.

  • endpoint_mode: dnsrr: Platform sets up DNS entries for the service such that a DNS query for the service name returns a list of IP addresses (DNS round-robin), and the client connects directly to one of these.

services:
  frontend:
    image: example/webapp
    ports:
      - "8080:80"
    deploy:
      mode: replicated
      replicas: 2
      endpoint_mode: vip

labels#

labels specifies metadata for the service. These labels are only set on the service and not on any containers for the service. This assumes the platform has some native concept of "service" that can match the Compose application model.

services:
  frontend:
    image: example/webapp
    deploy:
      labels:
        com.example.description: "This label will appear on the web service"

mode#

mode defines the replication model used to run the service on the platform. Either global, exactly one container per physical node, or replicated, a specified number of containers. The default is replicated.

services:
  frontend:
    image: example/webapp
    deploy:
      mode: global

placement#

placement specifies constraints and preferences for the platform to select a physical node to run service containers.

constraints#

constraints defines a required property the platform's node must fulfill to run the service container. It can be set either by a list or a map with string values.

deploy:
  placement:
    constraints:
      - disktype=ssd

deploy:
  placement:
    constraints:
      disktype: ssd

preferences#

preferences defines a property the platform's node should fulfill to run service container. It can be set either by a list or a map with string values.

deploy:
  placement:
    preferences:
      - datacenter=us-east

deploy:
  placement:
    preferences:
      datacenter: us-east

replicas#

If the service is replicated (which is the default), replicas specifies the number of containers that should be running at any given time.

services:
  frontend:
    image: example/webapp
    deploy:
      mode: replicated
      replicas: 6

resources#

resources configures physical resource constraints for container to run on platform. Those constraints can be configured as:

  • limits: The platform must prevent the container to allocate more.
  • reservations: The platform must guarantee the container can allocate at least the configured amount.
services:
  frontend:
    image: example/webapp
    deploy:
      resources:
        limits:
          cpus: '0.50'
          memory: 50M
          pids: 1
        reservations:
          cpus: '0.25'
          memory: 20M

cpus#

cpus configures a limit or reservation for how much of the available CPU resources, as number of cores, a container can use.

memory#

memory configures a limit or reservation on the amount of memory a container can allocate, set as a string expressing a byte value.

pids#

pids tunes a container’s PIDs limit, set as an integer.

devices#

devices configures reservations of the devices a container can use. It contains a list of reservations, each set as an object with the following parameters: capabilities, driver, count, device_ids and options.

Devices are reserved using a list of capabilities, making capabilities the only required field. A device must satisfy all the requested capabilities for a successful reservation.

capabilities#

capabilities are set as a list of strings, expressing both generic and driver specific capabilities. The following generic capabilities are recognized today:

  • gpu: Graphics accelerator
  • tpu: AI accelerator

To avoid name clashes, driver specific capabilities must be prefixed with the driver name. For example, reserving an nVidia CUDA-enabled accelerator might look like this:

deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["nvidia-compute"]
driver#

A different driver for the reserved device(s) can be requested using driver field. The value is specified as a string.

deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["nvidia-compute"]
          driver: nvidia
count#

If count is set to all or not specified, Compose reserves all devices that satisfy the requested capabilities. Otherwise, Compose reserves at least the number of devices specified. The value is specified as an integer.

deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["tpu"]
          count: 2

count and device_ids fields are exclusive. Compose returns an error if both are specified.

device_ids#

If device_ids is set, Compose reserves devices with the specified IDs provided they satisfy the requested capabilities. The value is specified as a list of strings.

deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["gpu"]
          device_ids: ["GPU-f123d1c9-26bb-df9b-1c23-4a731f61d8c7"]

count and device_ids fields are exclusive. Compose returns an error if both are specified.

options#

Driver specific options can be set with options as key-value pairs.

deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["gpu"]
          driver: gpuvendor
          options:
            virtualization: false

restart_policy#

restart_policy configures if and how to restart containers when they exit. If restart_policy is not set, Compose considers the restart field set by the service configuration.

  • condition. When set to:
  • none, containers are not automatically restarted regardless of the exit status.
  • on-failure, the container is restarted if it exits due to an error, which manifests as a non-zero exit code.
  • any (default), containers are restarted regardless of the exit status.
  • delay: How long to wait between restart attempts, specified as a duration. The default is 0, meaning restart attempts can occur immediately.
  • max_attempts: How many times to attempt to restart a container before giving up (default: never give up). If the restart does not succeed within the configured window, this attempt doesn't count toward the configured max_attempts value. For example, if max_attempts is set to '2', and the restart fails on the first attempt, more than two restarts must be attempted.
  • window: How long to wait before deciding if a restart has succeeded, specified as a duration (default: decide immediately).
deploy:
  restart_policy:
    condition: on-failure
    delay: 5s
    max_attempts: 3
    window: 120s

rollback_config#

rollback_config configures how the service should be rollbacked in case of a failing update.

  • parallelism: The number of containers to rollback at a time. If set to 0, all containers rollback simultaneously.
  • delay: The time to wait between each container group's rollback (default 0s).
  • failure_action: What to do if a rollback fails. One of continue or pause (default pause)
  • monitor: Duration after each task update to monitor for failure (ns|us|ms|s|m|h) (default 0s).
  • max_failure_ratio: Failure rate to tolerate during a rollback (default 0).
  • order: Order of operations during rollbacks. One of stop-first (old task is stopped before starting new one), or start-first (new task is started first, and the running tasks briefly overlap) (default stop-first).

update_config#

update_config configures how the service should be updated. Useful for configuring rolling updates.

  • parallelism: The number of containers to update at a time.
  • delay: The time to wait between updating a group of containers.
  • failure_action: What to do if an update fails. One of continue, rollback, or pause (default: pause).
  • monitor: Duration after each task update to monitor for failure (ns|us|ms|s|m|h) (default 0s).
  • max_failure_ratio: Failure rate to tolerate during an update.
  • order: Order of operations during updates. One of stop-first (old task is stopped before starting new one), or start-first (new task is started first, and the running tasks briefly overlap) (default stop-first).
deploy:
  update_config:
    parallelism: 2
    delay: 10s
    order: stop-first