package yggdrasil_decision_forests.distribute.proto

Get desktop application:
View/edit binary Protocol Buffers messages

GRPC service.

• rpc Run (Query, Answer)

  grpc.proto:72

  [AUTOMATION]: Internal rpc option goes here.

  message Query

  grpc.proto:89

  • optional bytes blob = 1

  • optional uint64 manager_uid = 3

  • optional int32 worker_idx = 4

  • optional WorkerConfig worker_config = 5

  message Answer

  grpc.proto:98

  • optional bytes blob = 1

  • optional string error = 2

• rpc WorkerRun (WorkerQuery, WorkerAnswer)

  grpc.proto:75

  [AUTOMATION]: Internal rpc option goes here.

  message WorkerQuery

  grpc.proto:107

  • optional bytes blob = 1

  • optional uint64 manager_uid = 3

  message WorkerAnswer

  grpc.proto:112

  • optional bytes blob = 1

  • optional string error = 2

• rpc Shutdown (ShutdownQuery, Empty)

  grpc.proto:78

  [AUTOMATION]: Internal rpc option goes here.

  message ShutdownQuery

  grpc.proto:103

  • optional bool kill_worker_manager = 1

• rpc Ping (Empty, Empty)

  grpc.proto:81

  [AUTOMATION]: Internal rpc option goes here.

• rpc UpdateWorkerAddress (UpdateWorkerAddressQuery, Empty)

  grpc.proto:84

  [AUTOMATION]: Internal rpc option goes here.

  message UpdateWorkerAddressQuery

  grpc.proto:117

  • optional int32 worker_idx = 1

  • optional string new_address = 2

Used in: GRPCImp

• optional string prefix = 1

• optional int32 num_workers = 2

Configuration for the distribution manager and workers.

Used in: model.proto.DeploymentConfig, utils.distribute_cli.proto.Config

• optional string implementation_key = 1

  Identifier of the distribution strategy. The registration strategy are registered with "REGISTER_Distribution_Manager" method. Make sure to link the implementation you are using. The "MULTI_THREAD" implementation is inefficient and only exist for unit testing purpose. For non-distributed training, rely on the multi-threading directly implemented in the learning algorithm.

• optional int32 verbosity = 5

  Amount of verbose. 0: No information. 1: Small amount of information (e.g. creating of the workers). 2: Information at every query.

• optional string working_directory = 6

  Working directory accessible by the manager and workers. Can be used for communication by distribution implementations. If you are training a model with distributed training, do not specify this path directly. Instead, specify the "cache_path" of the deployment configuration (and this path will be set automatically).

Used as request type in: Server.Ping

Used as response type in: Server.Ping, Server.Shutdown, Server.UpdateWorkerAddress

(message has no fields)

Computation distribution over multiple jobs communicating over GRPC. A worker in an instance of ":grpc_worker_main"/

• oneof worker_address

  • SocketAddresses socket_addresses = 1

  • Bns bns = 2

  • GrpcAddresses grpc_addresses = 5

• optional bool use_loas = 3

• optional int32 key = 4

  Optional global identifier of the distribute session (shared among the manager and the workers). Useful to differentiate between multiple distribute sessions running in the same process. At a given time, only one manager can be instantiated with the same key in a given process. Note that many managers can be instantiated without a key.

Used in: GRPCImp

• repeated string addresses = 1

  Address of the workers. Example: "127.0.0.1:2345". See https://grpc.github.io/grpc/cpp/md_doc_naming.html for the supported formats.

In process, non-parallelized distribution. For debugging and pipeline development.

• optional int32 num_workers = 1

  Number of workers / threads.

List of socket addresses.

Used in: GRPCImp

• repeated SocketAddresses.SocketAddress addresses = 1

Used in: SocketAddresses

• optional string ip = 1

  Human readable IP notation e.g. 127.0.0.1.

• optional int32 port = 2

Configuration of the workers.

Used in: Query

• optional bytes welcome_blob = 1

  Welcome message.

• repeated string worker_addresses = 2

  Addresses of the workers.

• optional uint64 manager_uid = 3

  Unique ID of the manager.

• optional string worker_name = 4

  Job done by the worker.

• optional int32 parallel_execution_per_worker = 5

  "parallel_execution_per_worker" parameter of the manager. How many request can a worker process in parallel.