删除追踪

您可以使用 mlflow.client.MlflowClient.delete_traces() 方法根据特定条件删除跟踪。此方法允许您按 实验 ID、最大时间戳 或 跟踪 ID 删除跟踪。

提示

删除跟踪是一个不可逆的过程。请确保提供给 delete_traces API 的参数符合预期的删除范围。

删除方法

按时间戳

删除早于特定时间戳的跟踪

import time
from mlflow import MlflowClient

client = MlflowClient()

# Get current timestamp in milliseconds
current_time = int(time.time() * 1000)

# Delete traces older than current time, limit to 100 traces
deleted_count = client.delete_traces(
    experiment_id="1", max_timestamp_millis=current_time, max_traces=100
)

print(f"Deleted {deleted_count} traces")

删除早于特定时间段的跟踪

from datetime import datetime, timedelta

# Calculate timestamp for 7 days ago
seven_days_ago = datetime.now() - timedelta(days=7)
timestamp_ms = int(seven_days_ago.timestamp() * 1000)

deleted_count = client.delete_traces(
    experiment_id="1", max_timestamp_millis=timestamp_ms
)

按跟踪 ID

通过跟踪 ID 删除特定的跟踪

from mlflow import MlflowClient

client = MlflowClient()

# Delete specific traces
trace_ids = ["trace_id_1", "trace_id_2", "trace_id_3"]

deleted_count = client.delete_traces(experiment_id="1", trace_ids=trace_ids)

print(f"Deleted {deleted_count} traces")

批量删除

批量删除跟踪以提高性能

import time
from datetime import datetime, timedelta
from mlflow import MlflowClient


def cleanup_old_traces(experiment_id: str, days_old: int = 30, batch_size: int = 100):
    """Delete traces older than specified days in batches"""
    client = MlflowClient()

    # Calculate cutoff timestamp
    cutoff_date = datetime.now() - timedelta(days=days_old)
    cutoff_timestamp = int(cutoff_date.timestamp() * 1000)

    total_deleted = 0

    while True:
        deleted_count = client.delete_traces(
            experiment_id=experiment_id,
            max_timestamp_millis=cutoff_timestamp,
            max_traces=batch_size,
        )

        total_deleted += deleted_count
        print(f"Deleted {deleted_count} traces (total: {total_deleted})")

        if deleted_count < batch_size:
            break

        time.sleep(0.1)  # Brief pause between batches

    return total_deleted


# Usage
cleanup_old_traces(experiment_id="1", days_old=7)

高级用例

选择性清理

根据特定条件删除跟踪

import mlflow
from mlflow import MlflowClient


def delete_error_traces(experiment_id: str):
    """Delete only traces that resulted in errors"""
    client = MlflowClient()

    # Search for error traces
    traces = mlflow.search_traces(
        experiment_ids=[experiment_id],
        filter_string="status = 'ERROR'",
        max_results=1000,
    )

    if traces:
        trace_ids = [trace.info.trace_id for trace in traces]
        deleted_count = client.delete_traces(
            experiment_id=experiment_id, trace_ids=trace_ids
        )
        print(f"Deleted {deleted_count} error traces")
        return deleted_count

    return 0


# Usage
delete_error_traces("1")

试运行模式

在实际删除前测试删除条件

import mlflow
from mlflow import MlflowClient


def delete_with_dry_run(experiment_id: str, max_timestamp: int, dry_run: bool = True):
    """Delete traces with optional dry-run mode"""
    client = MlflowClient()

    if dry_run:
        # Search to see what would be deleted
        traces = mlflow.search_traces(
            experiment_ids=[experiment_id], filter_string=f"timestamp < {max_timestamp}"
        )
        print(f"DRY RUN: Would delete {len(traces)} traces")
        return len(traces)
    else:
        deleted_count = client.delete_traces(
            experiment_id=experiment_id,
            max_timestamp_millis=max_timestamp,
            max_traces=1000,
        )
        print(f"ACTUAL: Deleted {deleted_count} traces")
        return deleted_count


# Always test with dry run first
count = delete_with_dry_run("1", 1234567890000, dry_run=True)
if count < 100:  # Only proceed if reasonable number
    delete_with_dry_run("1", 1234567890000, dry_run=False)

错误处理

优雅地处理删除错误

from mlflow import MlflowClient
from mlflow.exceptions import MlflowException


def safe_delete_traces(experiment_id: str, **delete_params):
    """Delete traces with error handling"""
    client = MlflowClient()

    try:
        deleted_count = client.delete_traces(
            experiment_id=experiment_id, **delete_params
        )
        print(f"Successfully deleted {deleted_count} traces")
        return deleted_count

    except MlflowException as e:
        if "experiment not found" in str(e).lower():
            print("Error: Experiment not found")
        elif "permission" in str(e).lower():
            print("Error: Permission denied")
        else:
            print(f"MLflow error: {e}")
        return 0

    except Exception as e:
        print(f"Unexpected error: {e}")
        return 0


# Usage
safe_delete_traces("1", max_timestamp_millis=1234567890000, max_traces=50)

最佳实践

始终先测试：使用搜索查询或试运行模式来验证在继续操作前将要删除哪些跟踪。

批量删除：对于大量的跟踪，分批删除以避免超时和性能问题。

设置合理的限制：使用 max_traces 参数来防止一次意外删除太多跟踪。

监控和记录：跟踪删除活动以用于审计目的，尤其是在生产环境中。

优雅地处理错误：实施适当的错误处理来管理网络问题、权限问题或无效参数。

参数参考

参数	类型	描述
experiment_id	str	必需。 包含要删除跟踪的实验的 ID
max_timestamp_millis	int	删除在此时间戳（以毫秒为单位）之前创建的跟踪
trace_ids	List[str]	删除具有这些特定跟踪 ID 的跟踪
max_traces	int	在此操作中要删除的最大跟踪数

注意

您必须指定 max_timestamp_millis 或 trace_ids，但不能同时指定两者。max_traces 参数不能与 trace_ids 一起使用。