config/hostmgr/base.yaml

storage:
  cassandra:
    max_parallel_batches: 1000
    max_updates_job: 10
    connection:
      contactPoints: ["127.0.0.1"]
      port: 9042
      consistency: LOCAL_QUORUM
      serialConsistency: LOCAL_SERIAL
      hostPolicy: TokenAwareHostPolicy
      # Need to increase timeout from 10s to 20s to avoid recovery code from timing out
      # We saw recovery code timing out when peloton was recovering from a
      # Cassandra latency spike issue.
      timeout: 20s
    store_name: peloton_test
    migrations: pkg/storage/cassandra/migrations/
  use_cassandra: true
  auto_migrate: false

host_manager:
  http_port: 5291
  grpc_port: 5391
  offer_hold_time_sec: 864000
  offer_pruning_period_sec: 3600
  taskupdate_ack_concurrency: 10
  taskupdate_buffer_size: 100000
  task_reconciler:
    initial_reconcile_delay_sec: 60
    reconcile_interval_sec: 1800
    explicit_reconcile_batch_interval_sec: 5
    explicit_reconcile_batch_size: 1000
  hostmap_refresh_interval: 10s
  host_pruning_period_sec: 120s
  host_placing_offer_status_sec: 300s
  held_host_pruning_period_sec: 180s
  hostmgr_backoff_retry_count: 3
  hostmgr_backoff_retry_interval_sec: 15
  host_drainer_period: 900s
  # scarce_resource_types are resources, which are exclusively reserved for specific task requirements,
  # and to prevent every task to schedule on those hosts such as GPU.
  # Resource Types are case sensitive, supported resource types are "CPU", "GPU", "Mem" and "Disk"
  # As of now, GPU is the only supported scarce resource type, adding other resource types are no-op.
  scarce_resource_types:

  # slack_resource_types are resource types exposed as revocable resources by Mesos.
  # usage slack = allocated resources - actual resources consumed.
  # Mesos supports cpus & mem as revocable resources.
  # Peloton only supports [cpus] as revocable/slack resources.
  slack_resource_types:

  # bin_packing represents the strategy hostmanager is going to use in order
  # to pack the tasks in the host. By default it was FIRST_FIT, we are changing
  # it to DEFRAG.
  bin_packing: FIRST_FIT # DEFRAG/FIRST_FIT

  # bin packing refresh interval represents the time interval in which
  # we can refresh the list of hosts based on bin packing algorithm
  bin_packing_refresh_interval: 30s
  enable_host_pool: false
  host_pool_reconcile_interval: 10s

mesos:
  encoding: "x-protobuf"
  framework:
    gpu_supported: true
    task_killing_state: false
    partition_aware: false
    revocable_resources: false
    user: "root"
    name: "Peloton"
    # TODO : add roles for other components
    role: "peloton"
    principal: "peloton"
    # ~100 weeks to failover
    failover_timeout: 60000000
    max_connections_to_mesos_master: 1024

election:
  root: "/peloton"

health:
  heartbeat_interval: 5s

metrics:
  runtime_metrics:
    enabled: true
    interval: 10s