Collector service (Alpha)

Deployment scenario for near real-time monitoring.

Overview

In this scenario, you deploy an Evidently Collector service for near real-time monitoring.

Evidently Collector is a service that allows you to collect online events into batches, create Reports or Test Suites over batches of data, and save them as snapshots into the workspace.

You will need to POST the predictions from the ML service to the Evidently Collector service. You can POST data on every prediction or batch them. The Evidently collector service will perform asynchronous computation of monitoring snapshots based on the provided config.

You can also pass the path to the optional reference dataset.

If you receive delayed ground truth, you can also later compute and log the model quality to the same project. You can run it as a separate process or batch monitoring job.

Code example

Refer to this example:

https://github.com/evidentlyai/evidently/tree/main/examples/integrations/collector_service

Collector configuration

Before sending events, you must configure the collector and start the service.

You can choose either of the two options:

• Create configuration via code, save it to a JSON file, and run the service using it.

• Run the service first and create configuration via API.

The collector service can simultaneously run multiple “collectors” that compute and save snapshots to different workspaces or projects. Each one is represented by a CollectorConfig object.

CollectorConfig Object

You can configure the following parameters:

Parameter	Type	Description
trigger	CollectorTrigger	Defines when to create a new snapshot from the current batch.
report_config	ReportConfig	Configures the contents of the snapshot: Report or TestSuite computed for each batch of data.
reference_path	Optional[str]	Local path to a .parquet file with the reference dataset.
cache_reference	bool	Defines whether to cache reference data or re-read it each time.
api_url	str	URL where the Evidently UI Service runs and snapshots will be saved to.
api_secret	Optional[str]	Evidently UI Service secrets.
project_id	str	ID of the project to save snapshots to.

You can create a ReportConfig object from Report or TestSuite objects. You must run them first so that all Metrics and Tests are collected (including when you use Presets or Test/Metric generators).

report = Report(...) 
report.run(...) 
report_config = ReportConfig.from_report(report) 

# or 

test_suite = TestSuite(...) 
test_suite.run(...) 
report_config = ReportConfig.from_test_suite(test_suite)

CollectorTrigger

Currently, there are two options available:

• IntervalTrigger: triggers the snapshot calculation each interval seconds

• RowsCountTrigger: triggers the snapshot calculation every time the configured amount of rows have been sent to the collector service

Note: we are also working on CronTrigger and other triggers. Would you like to see additional scenarios? Please open a GitHub issue with your suggestions.

Setup via file

You can define the configuration and save it as a JSON file. Example:

config = CollectorServiceConfig(collectors={
        "main": CollectorConfig(
            trigger=IntervalTrigger(interval=60 * 60),
            report_config=ReportConfig.from_report(report),
            reference_path="reference_data.parquet",
            project_id="834ec9a0-ee58-4e64-816b-c593b0b6c45c",
            api_url="http://localhost:8000"
        )
    })

config.save("collector_config.json")

Then, run the following command:

evidently collector --config-path collector_config.json

Setup via API

First, run the collector service:

evidently collector

Then, use the CollectorClient to add new collector config:

config = CollectorConfig(
        trigger=IntervalTrigger(interval=60 * 60),
        report_config=ReportConfig.from_report(report),
        reference_path="reference_data.parquet",
        project_id="834ec9a0-ee58-4e64-816b-c593b0b6c45c",
        api_url="<http://localhost:8000>"
    )

Update reference via API

To specify the path to the reference dataset:

reference: pd.DataFrame = ...
client = CollectorClient("<http://localhost:8001>")
client.set_reference("main", reference)

Send events via API

To send events from your ML service:

client = CollectorClient("http://localhost:8001")

events: pd.DataFrame = ...
client.send_data("main", events)

Send events via curl

To send data with curl:

curl 

POST '.../<collector config id>/data'

headers {'evidently-secret': '...', 'Content-Type': 'application/json'}

body '{"column1": {"0": 7.0, "1": 5.0}, "column2": {"0": "a", "1": "b"}}'

Example:

curl -d '{"column1": {"0": 7.0, "1": 5.0}, "column2": {"0": "a", "1": "b"}}' -H 'Content-Type: application/json' http://0.0.0.0:8001/default/data

This is how it looks in the Terminal.

Sending data:

The data is received by the collector service: