Skip to content

Truthound

Drift and Anomaly Config

seadonggyun4/Truthound

Upstream Source

This page is part of Truthound Orchestration 3.x.

Source repository: seadonggyun4/truthound-orchestration Upstream docs path: docs/common/drift-config.md Edit upstream page: Edit in orchestration

Drift and Anomaly Config¶

This page describes the configuration objects used by advanced drift, anomaly, and streaming workflows.

Who This Is For¶

operators enabling advanced quality monitoring beyond deterministic checks
maintainers wiring platform-specific drift or anomaly tasks
teams standardizing reusable config objects across orchestration hosts

`DriftConfig`¶

DriftConfig controls how a baseline is compared with a current dataset.

from common.base import DriftConfig

config = DriftConfig(
    method="ks",
    columns=("revenue", "count"),
    threshold=0.05,
    min_severity="medium",
    timeout_seconds=60,
    extra={"n_permutations": 1000},
)

Builder Pattern¶

config = DriftConfig()
config = config.with_method("psi")
config = config.with_columns(("revenue", "user_count"))
config = config.with_threshold(0.1)

`AnomalyConfig`¶

AnomalyConfig controls detector choice and sensitivity for anomaly-oriented workflows.

from common.base import AnomalyConfig

config = AnomalyConfig(
    detector="isolation_forest",
    columns=("amount",),
    contamination=0.05,
    threshold=None,
    timeout_seconds=120,
    extra={"n_estimators": 200},
)

Builder Pattern¶

config = AnomalyConfig()
config = config.with_detector("lof")
config = config.with_columns(("amount", "frequency"))
config = config.with_contamination(0.03)

`StreamConfig`¶

StreamConfig is used when validation happens incrementally or in chunks.

from common.base import StreamConfig

config = StreamConfig(
    batch_size=5000,
    max_batches=100,
    timeout_per_batch_seconds=30.0,
    fail_fast=True,
)

Builder Pattern¶

config = StreamConfig()
config = config.with_batch_size(10000)
config = config.with_fail_fast(True)

Platform-Specific Config Surfaces¶

Individual adapters expose platform-native config types on top of the shared concepts:

Platform	Drift Config	Anomaly Config
Dagster	`DriftOpConfig`	`AnomalyOpConfig`
Prefect	`DriftTaskConfig`	`AnomalyTaskConfig`
Mage	`DriftBlockConfig`	`AnomalyBlockConfig`
Kestra	`DriftScriptConfig`	`AnomalyScriptConfig`

Operational Guidance¶

keep thresholds explicit and documented
treat anomaly configs as monitoring policy, not only as model settings
prefer shared config builders when multiple pipelines should behave consistently