Skip to content

Latest commit

 

History

History
288 lines (194 loc) · 10.5 KB

README.md

File metadata and controls

288 lines (194 loc) · 10.5 KB
Raw

English | 中文

FATE-Flow

Introduction

FATE-Flow is a end to end pipeline platform for Federated Learning. Pipeline is a sequence of components which is specifically designed for highly flexible, high performance federated learning tasks. That includes data processing, modeling, training, verification, publishing and serving inference.

FATE-Flow Federated Learning Pipeline

federated_learning_pipeline

FATE-Flow now supports

  • DAG define Pipeline
  • Describe DAG using FATE-DSL in JSON format
  • FATE has a large number of default federated learning components, such as Hetero LR/Homo LR/Secure Boosting Tree and so on.
  • Developers can easily implement custom components using Basic-API and build their own Pipeline through DSL.
  • Federated Modeling Task Life-cycle Manager, start/stop, status synchronization and so on.
  • Powerful Federated Scheduler, support multiple scheduling strategies for DAG Job and Component Task.
  • Real-time tracking of data, parameters, models, and metric during the experiment.
  • Federated Model Manager, model binding, versioning and deployment tools.
  • Provide HTTP API and Command Line Interface.
  • Data and interface support for modeling visualization on FATE-Board.

Build Pipeline

DSL Example

fate_flow_dsl

Write a DSL

Only one step is required to configure a component for pipeline.

  • define the module of this component
  • define the input, includes data, model or isometric_model(only be used for FeatureSelection)
  • define the output, includes data and model fate_flow_component_dsl

Pipeline Run Example

fate_flow_dag

Architecture

fate_flow_arch

Deploy

Fate-flow is deployed in $PYTHONPATH/fate_flow/. It depends on two configuration files: $PYTHONPATH/arch/conf/server.conf, $PYTHONPATH/fate_flow/settings.py

server.conf

Server.conf configures the address of all FATE services. FATE-Flow of different deployment modes needs different fate services. For details, please refer to the following specific deployment mode.

settings.py

Key configuration item description:

Configuration item Configuration item meaning Configuration item value
IP listen address for FATE-Flow default 0.0.0.0
GRPC_PORT listen port for the grpc server of FATE-Flow default 9360
HTTP_PORT listen port for the http server of FATE-Flow default 9380
WORK_MODE the work mode of FATE-Flow 0 for standalone, 1 for cluster
USE_LOCAL_DATABASE Whether to use a local database(sqlite) False for no, True for yes
USE_AUTHENTICATION Whether to enable authentication False for no, True for yes
USE_CONFIGURATION_CENTER Whether to use zookeeper False for no, True for yes
MAX_CONCURRENT_JOB_RUN Pipeline jobs that are executed in parallel at the same time default 5
MAX_CONCURRENT_JOB_RUN_HOST Maximum running jobs default 10
DATABASE configuration for mysql database custom configuration
REDIS configuration for redis custom configuration
REDIS_QUEUE_DB_INDEX the redis db index of redis queue default 0

service.sh

Server start/stop/restart script

Configuration item Configuration item meaning Configuration item value
PYTHONPATH the python path Absolute path to the parent directory of the FATE-Flow
venv the python virtual environment custom configuration, such as /data/projects/fate/venv, not /data/projects/fate/venv/bin/activate

Simple Standalone

You only need to start the FATE-Flow service to run the federated learning modeling experiment.

Configuration
Configuration item Configuration item value
WORK_MODE 0
USE_LOCAL_DATABASE True
Features
  • Use Sqlite as database, db file is fate_flow_sqlite.db on the FATE-Flow root directory.
  • Use in-process queue as job queue.
  • Multi-Party communication loop back.

Standard Standalone

You need to deploy three service:

  • MySQL
  • FATE-Flow
  • FATE-Board
Docker version

FATE provides a standalone version of the docker for experience.please refer to docker version deploy guide at docker-deploy.

Configuration
Configuration item Configuration item value
WORK_MODE 0
USE_LOCAL_DATABASE False
DATABASE configuration for mysql database
Features
  • Use MySQL as database.
  • Use in-process queue as job queue.
  • Multi-Party communication loop back.
  • Support visualization by FATE-Board !

Cluster

FATE also provides a distributed runtime architecture for Big Data scenario. Migration from standalone to cluster requires configuration change only. No algorithm change is needed. To deploy FATE on a cluster, please refer to cluster deploy guide at cluster-deploy.

Configuration
Configuration item Configuration item value
WORK_MODE 1
DATABASE configuration for mysql database
REDIS configuration for redis
Features
  • Use MySQL as database.
  • Use redis queue as job queue.
  • Multi-Party communication use proxy.
  • Support visualization by FATE-Board !

Usage

FATE-Flow provide REST API and Command Line Interface. Let's start using the client to run a Federated Learning Pipeline Job example(Standalone).

Offline Modeling

Upload Data

python fate_flow_client.py -f upload -c examples/upload_guest.json
python fate_flow_client.py -f upload -c examples/upload_host.json

Submit Job

python fate_flow_client.py -f submit_job -d examples/test_hetero_lr_job_dsl.json -c examples/test_hetero_lr_job_conf.json

Command response example:

{
    "data": {
        "board_url": "http://localhost:8080/index.html#/dashboard?job_id=2019121910313566330118&role=guest&party_id=9999",
        "job_dsl_path": "xxx/jobs/2019121910313566330118/job_dsl.json",
        "job_runtime_conf_path": "xxx/jobs/2019121910313566330118/job_runtime_conf.json",
        "logs_directory": "xxx/logs/2019121910313566330118",
        "model_info": {
            "model_id": "arbiter-10000#guest-9999#host-10000#model",
            "model_version": "2019121910313566330118"
        }
    },
    "jobId": "2019121910313566330118",
    "retcode": 0,
    "retmsg": "success"
}

Some of the following operations will use these response information.

Query Job

python fate_flow_client.py -f query_job -r guest -p 10000 -j $job_id

And then, you can found so many useful command from CLI.

For more Federated Learning pipeline Job example, please refer at federatedml-1.x-examples and it's README

Online Inference

Publish model to FATE-Serving, and then using Serving's GRPC API to inference.

Modify service configuration

Modify the IP and end of FATE-Serving in arch/conf/server_conf.json (please note that many parties need to modify the actual deployment address of their respective FATE-Serving), the content is "servings ":["ip: port"], restart FATE-Flow after modification. server_conf.json format is as follows:

{
	"servers": {
		"servings": [
			"127.0.0.1:8000"
		]
	}
}

Publish Model

python fate_flow_client.py -f load -c examples/publish_load_model.json

Please replace the corresponding configuration in publish_load_model.json with your job configuration. After that, you can make online inference request to FATE-Serving by specifying the model id and model version used.

Publish Model Online Default

python fate_flow_client.py -f bind -c examples/bind_model_service.json

Please replace the corresponding configuration in publish_online_model.json with your job configuration. After that, the FATE-Serving uses the configuration you provided to set the party's default model id and the default model version that involves the model id. And then, will you can make online inference request to FATE-Serving by only specifying the party_id or the model id.

Logs

FATE-Flow Server log

$PYTHONPATH/logs/fate_flow/

Job log

$PYTHONPATH/logs/$job_id/

FAQ

What is the role of FATE FLOW in the FATE?

  • FATE Flow is a scheduling system that schedules the execution of algorithmic components based on the DSL of the job submitted by the user.

ModuleNotFoundError: No module named "arch"

  • Set PYTHONPATH to the parent directory of fate_flow.

Why does the task show success when submitting the task, but the task fails on the dashboard page?

  • Submit success just means that the job was submitted and not executed. If the job fails, you need to check the log.
  • You can view the logs through the board.

What meaning and role do the guest, host, arbiter, and local roles represent in fate?

  • Arbiter is used to assist multiple parties to complete joint modeling. Its main role is to aggregate gradients or models. For example, in vertical lr, each party sends half of its gradient to arbiter, and then arbiter jointly optimizes, etc.
  • Guest represents the data application party.
  • Host is the data provider.
  • Local refers to local, only valid for upload and download.

Error about“cannot find xxxx” when killing a waiting job

  • Fate_flow currently only supports kill on the job initiator, kill will report "cannot find xxx".

What is the upload data doing?

  • Upload data is uploaded to eggroll and becomes a DTable format executable by subsequent algorithms.

How to download the generated data in the middle of the algorithm?

  • You can use python fate_flow_client.py -f component_output_model -j $job_id -r $role -g $guest -cpn $component_name -o $output_path

If the same file upload is executed twice, will fate delete the first data and upload it again?

  • It will be overwritten if the keys are the same in the same table.

What is the reason for the failure of this job without error on the board?

  • The logs in these places will not be displayed on the board:$job_id/fate_flow_schedule.log, logs/error.log, logs/fate_flow/ERROR.log .

What is the difference between the load and bind commands?

  • Load can be understood as a model release, and bind is the default model version.