Scheduler
Configure Scheduler YAML File
The default path for the scheduler yaml configuration file is /etc/dragonfly/scheduler.yaml in linux,
and the default path is $HOME/.dragonfly/config/scheduler.yaml in darwin.
# Server scheduler instance configuration.
server:
# # Access ip for other services,
# # when local ip is different with access ip, advertiseIP should be set.
# advertiseIP: 127.0.0.1
# # Access port for other services,
# # when local ip is different with access port, advertisePort should be set.
# advertisePort: 8002
# # Listen ip.
# listenIP: 0.0.0.0
# Port is the ip and port scheduler server listens on.
port: 8002
# # Server host.
# host: localhost
# WorkHome is working directory.
# In linux, default value is /usr/local/dragonfly.
# In macos(just for testing), default value is /Users/$USER/.dragonfly.
workHome: ''
# logDir is the log directory.
# In linux, default value is /var/log/dragonfly.
# In macos(just for testing), default value is /Users/$USER/.dragonfly/logs.
logDir: ''
# cacheDir is dynconfig cache directory.
# In linux, default value is /var/cache/dragonfly.
# In macos(just for testing), default value is /Users/$USER/.dragonfly/cache.
cacheDir: ''
# pluginDir is the plugin directory.
# In linux, default value is /usr/local/dragonfly/plugins.
# In macos(just for testing), default value is /Users/$USER/.dragonfly/plugins.
pluginDir: ''
# dataDir is the directory.
# In linux, default value is /var/lib/dragonfly.
# In macos(just for testing), default value is /Users/$USER/.dragonfly/data.
dataDir: ''
# Scheduler policy configuration.
scheduler:
# Algorithm configuration to use different scheduling algorithms,
# default configuration supports "default" and "ml"
# "default" is the rule-based scheduling algorithm,
# "ml" is the machine learning scheduling algorithm
# It also supports user plugin extension, the algorithm value is "plugin",
# and the compiled `d7y-scheduler-plugin-evaluator.so` file is added to
# the dragonfly working directory plugins.
algorithm: default
# backToSourceCount is single task allows the peer to back-to-source count.
backToSourceCount: 3
# retryBackToSourceLimit reaches the limit, then the peer back-to-source.
retryBackToSourceLimit: 5
# Retry scheduling limit times.
retryLimit: 10
# Retry scheduling interval.
retryInterval: 50ms
# GC metadata configuration.
gc:
# pieceDownloadTimeout is the timeout of downloading piece.
pieceDownloadTimeout: 30m
# peerGCInterval is the interval of peer gc.
peerGCInterval: 10s
# peerTTL is the ttl of peer. If the peer has been downloaded by other peers,
# then PeerTTL will be reset
peerTTL: 24h
# taskGCInterval is the interval of task gc. If all the peers have been reclaimed in the task,
# then the task will also be reclaimed.
taskGCInterval: 30m
# hostGCInterval is the interval of host gc.
hostGCInterval: 6h
# hostTTL is time to live of host. If host announces message to scheduler,
# then HostTTl will be reset.
hostTTL: 1h
# Database info used for server.
database:
# Redis configuration.
redis:
# Redis addresses.
addrs:
- redis-servive:6379
# Redis sentinel master name.
masterName: ''
# Redis username.
username: ''
# Redis password.
password: ''
# Redis broker DB.
brokerDB: 1
# Redis backend DB.
backendDB: 2
# Network topology DB.
networkTopologyDB: 3
# Dynamic data configuration.
dynConfig:
# Dynamic config refresh interval.
refreshInterval: 1m
# Scheduler host configuration.
host:
# idc is the idc of scheduler instance.
idc: ''
# location is the location of scheduler instance.
location: ''
# Manager configuration.
manager:
# addr is manager access address.
addr: manager-service:65003
# schedulerClusterID cluster id to which scheduler instance belongs.
schedulerClusterID: 1
# keepAlive keep alive configuration.
keepAlive:
# KeepAlive interval.
interval: 5s
# Seed peer configuration.
seedPeer:
# Scheduler enable seed peer as P2P peer,
# if the value is false, P2P network will not be back-to-source through
# seed peer but by peer and preheat feature does not work.
enable: true
# Machinery async job configuration,
# see https://github.com/RichardKnop/machinery.
job:
# Scheduler enable job service.
enable: true
# Number of workers in global queue.
globalWorkerNum: 500
# Number of workers in scheduler queue.
schedulerWorkerNum: 500
# Number of workers in local queue.
localWorkerNum: 1000
# Network topology to collect configuration.
networkTopology:
# enable network topology service, including probe, network topology collection.
enable: true
# collectInterval is the interval of collecting network topology.
collectInterval: 2h
probe:
# queueLength is the length of probe queue.
queueLength: 5
# count is the number of probing hosts.
count: 10
# Trainer service configuration.
trainer:
# enable trainer service.
enable: false
# addr is trainer service address.
addr: trainer-service:9090
# interval is the interval of training.
interval: 168h
# uploadTimeout is the timeout of uploading dataset to trainer.
uploadTimeout: 1h
# Store task download information.
storage:
# maxSize sets the maximum size in megabytes of storage file.
maxSize: 100
# maxBackups sets the maximum number of storage files to retain.
maxBackups: 10
# bufferSize sets the size of buffer container,
# if the buffer is full, write all the records in the buffer to the file.
bufferSize: 100
# Enable prometheus metrics.
metrics:
# Scheduler enable metrics service.
enable: true
# Metrics service address.
addr: ':8000'
# Enable host metrics.
enableHost: false
security:
# autoIssueCert indicates to issue client certificates for all grpc call.
# If AutoIssueCert is false, any other option in Security will be ignored.
autoIssueCert: false
# caCert is the root CA certificate for all grpc tls handshake, it can be path or PEM format string.
caCert: ''
# tlsVerify indicates to verify certificates.
tlsVerify: false
# tlsPolicy controls the grpc shandshake behaviors:
# force: both ClientHandshake and ServerHandshake are only support tls
# prefer: ServerHandshake supports tls and insecure (non-tls), ClientHandshake will only support tls
# default: ServerHandshake supports tls and insecure (non-tls), ClientHandshake will only support insecure (non-tls)
# Notice: If the drgaonfly service has been deployed, a two-step upgrade is required.
# The first step is to set tlsPolicy to default, and then upgrade the dragonfly services.
# The second step is to set tlsPolicy to prefer, and then completely upgrade the dragonfly services.
tlsPolicy: 'prefer'
certSpec:
# dnsNames is a list of dns names be set on the certificate.
dnsNames:
- 'dragonfly-scheduler'
- 'dragonfly-scheduler.dragonfly-system.svc'
- 'dragonfly-scheduler.dragonfly-system.svc.cluster.local'
# ipAddresses is a list of ip addresses be set on the certificate.
ipAddresses:
# validityPeriod is the validity period of certificate.
validityPeriod: 4320h
network:
# Enable ipv6.
enableIPv6: false
# Console shows log on console.
console: false
# Whether to enable debug level logger and enable pprof.
verbose: false
# Listen port for pprof, only valid when the verbose option is true
# default is -1. If it is 0, pprof will use a random port.
pprof-port: -1
# Jaeger endpoint url, like: http://jaeger.dragonfly.svc:14268/api/traces.
jaeger: ''