基于Python脚本实现Flink on YARN任务批量触发Savepoint的实践指南

基于Python脚本实现Flink on YARN任务批量触发Savepoint的实践指南

一、背景与价值

在流计算生产环境中，Flink on YARN的部署方式凭借其资源管理优势被广泛采用。Savepoint作为Flink任务状态的一致性快照，承载着故障恢复、版本升级、作业暂停等重要场景的核心保障。当集群中运行数十个Flink作业时，手动逐个触发Savepoint耗时且易出错。本文提出一种基于Python脚本的批量化操作方案，可显著提升运维效率。

二、技术实现原理

2.1 核心流程设计

1. 作业发现机制 ：通过YARN REST API或yarn application命令获取RUNNING状态作业列表
2. 元数据解析：提取Application ID和Job ID（需注意Per-Job模式与Session模式的区别）
3. 并发控制：采用队列机制管理并行触发请求，避免集群资源过载
4. 状态反馈 ：实时捕获flink cancel -s [targetDirectory]命令输出，记录成功/失败状态

2.2 关键实现步骤

#******************************************************************#
##author: david.zhou
##create time: 2025-03-27 16:56:34
#******************************************************************#
#!/usr/bin/env python3
# -*- coding: utf-8 -*-

#"""
#获取并为指定的Flink任务创建savepoint的脚本
#"""

import os
import re
import json
import time
import subprocess
import requests
from typing import List, Dict, Tuple, Optional


# 设置不同版本Flink的路径
FLINK_PATHS = {
    "1.14": "/opt/flink-1.14/bin/flink",
    "1.15": "/opt/flink-1.15/bin/flink",
    "1.16": "/opt/flink-1.16/bin/flink",
    "1.17": "/opt/flink-1.17/bin/flink",
    "1.18": "/opt/flink-1.18/bin/flink"
}
DEFAULT_FLINK_PATH = "flink"  # 默认使用环境变量中的flink命令

# 直接在脚本中定义过滤参数（可以根据需要修改）
# 留空列表表示处理所有任务，添加元素表示只处理匹配的任务
FILTERS = ["aaaa"]
# 如果要处理所有任务，请使用：FILTERS = []
# FILTERS = []


def run_command(command: str) -> Tuple[str, int]:
    #"""执行shell命令并返回输出和状态码"""
    # 检查Python版本，决定使用哪个参数
    import sys
    if sys.version_info >= (3, 7):
        process = subprocess.Popen(
            command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True
        )
    else:
        process = subprocess.Popen(
            command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, universal_newlines=True
        )
    stdout, stderr = process.communicate()
    return stdout + stderr, process.returncode



def get_running_flink_jobs() -> List[Dict]:
    #"""获取所有运行中的Flink任务"""
    print("正在获取所有运行中的Flink任务...")
    output, _ = run_command("yarn application -list -appStates RUNNING")
    
    running_jobs = []
    for line in output.splitlines():
        if "Apache Flink" in line:
            # 提取应用ID
            app_id_match = re.search(r"(application_\d+_\d+)", line)
            if app_id_match:
                app_id = app_id_match.group(1)
                
                # 提取任务名称
                job_name_match = re.search(r"\s+([^\s]+)\s+Apache\s+Flink", line)
                job_name = job_name_match.group(1) if job_name_match else app_id
                
                running_jobs.append({
                    "app_id": app_id,
                    "info": line,
                    "name": job_name
                })
                print(f"- {line}")
    
    return running_jobs


def filter_jobs(jobs: List[Dict], filters: List[str]) -> List[Dict]:
    #"""根据过滤条件筛选任务"""
    if not filters:
        return jobs
    
    print(f"\n应用过滤条件: {filters}")
    filtered_jobs = []
    
    for job in jobs:
        app_id = job["app_id"]
        app_info = job["info"]
        app_name = job["name"]
        matched = False
        
        for filter_str in filters:
            if filter_str == app_id or filter_str in app_info:
                matched = True
                break
        
        if matched:
            filtered_jobs.append(job)
            print(f"匹配任务: {app_id} ({app_name})")
    
    return filtered_jobs


def get_tracking_url(app_id: str) -> Optional[str]:
    #"""获取应用的Tracking URL"""
    print(f"  获取Tracking URL...")
    output, _ = run_command(f"yarn application -status {app_id}")
    
    tracking_url_match = re.search(r"Tracking-URL\s*:\s*(http://[^\s]+)", output, re.IGNORECASE)
    if tracking_url_match:
        tracking_url = tracking_url_match.group(1)
        print(f"  找到Tracking URL: {tracking_url}")
        return tracking_url
    
    return None


def extract_host_from_url(url: str) -> Optional[str]:
    #"""从URL中提取主机名"""
    host_match = re.search(r"http://([^/]+)", url)
    if host_match:
        host = host_match.group(1)
        print(f"  Flink主机: {host}")
        return host
    
    return None


def get_flink_version(host: str) -> str:
    #"""获取Flink版本"""
    print("  获取Flink版本...")
    config_api_url = f"http://{host}/config/"
    print(f"  访问配置API: {config_api_url}")
    
    try:
        response = requests.get(config_api_url, timeout=10)
        config_json = response.text
        print(f"  配置API返回: {config_json}")
        
        # 尝试解析JSON
        try:
            config_data = json.loads(config_json)
            flink_version = config_data.get("flink-version")
            if flink_version:
                print(f"  Flink版本: {flink_version}")
                return flink_version
        except json.JSONDecodeError:
            pass
        
        # 使用正则表达式提取版本
        version_match = re.search(r'"flink-version":"([0-9\.]+)"', config_json)
        if version_match:
            flink_version = version_match.group(1)
            print(f"  Flink版本: {flink_version}")
            return flink_version
    except Exception as e:
        print(f"  获取Flink版本出错: {e}")
    
    # 默认版本
    print("  无法获取Flink版本，假设为1.14版本")
    return "1.14"


def get_flink_job_id(host: str) -> Optional[str]:
    #"""获取Flink JobID"""
    print("  通过REST API获取JobID...")
    jobs_api_url = f"http://{host}/jobs"
    print(f"  访问作业API: {jobs_api_url}")
    
    try:
        response = requests.get(jobs_api_url, timeout=10)
        jobs_json = response.text
        print(f"  作业API返回: {jobs_json}")
        
        # 尝试解析JSON
        try:
            jobs_data = json.loads(jobs_json)
            for job in jobs_data.get("jobs", []):
                if job.get("status") == "RUNNING":
                    job_id = job.get("id")
                    if job_id:
                        print(f"  找到运行中的Flink JobID: {job_id}")
                        return job_id
        except json.JSONDecodeError:
            pass
        
        # 使用正则表达式提取JobID
        job_id_match = re.search(r'"id":"([a-z0-9]+)","status":"RUNNING"', jobs_json)
        if job_id_match:
            job_id = job_id_match.group(1)
            print(f"  找到运行中的Flink JobID: {job_id}")
            return job_id
    except Exception as e:
        print(f"  获取Flink JobID出错: {e}")
    
    print("  无法从JSON响应中提取JobID")
    return None


def get_flink_command(version: str) -> str:
    #"""根据Flink版本选择对应的命令路径"""
    major_version, minor_version = version.split(".")[:2]
    version_key = f"{major_version}.{minor_version}"
    
    print(f"  Flink主版本: {major_version}, 次版本: {minor_version}")
    
    if major_version == "1":
        flink_path = FLINK_PATHS.get(version_key)
        if flink_path and os.path.isfile(flink_path):
            print(f"  使用Flink {version_key}路径: {flink_path}")
            return flink_path
        else:
            print(f"  Flink {version_key}路径不存在，使用默认路径")
            # 对于1.14版本，使用特定的默认路径
            if version_key == "1.14":
                return "/opt/flink-1.14/bin/flink"
    else:
        print(f"  未知的Flink主版本: {major_version}，使用默认路径")
    
    return DEFAULT_FLINK_PATH


def check_savepoint_status(host: str, job_id: str, timeout: int = 3600) -> bool:
    #"""检查savepoint状态，直到完成或超时"""
    print("  Savepoint操作已触发，开始检查执行状态...")
    
    start_time = time.time()
    completed = False
    savepoint_path = None
    
    #超时时间可以自定义，默认 1 小时
    while time.time() - start_time < timeout:
        checkpoints_api_url = f"http://{host}/jobs/{job_id}/checkpoints"
        print(f"  检查savepoint状态: {checkpoints_api_url}")
        
        try:
            response = requests.get(checkpoints_api_url, timeout=10)
            checkpoints_json = response.text
            
            # 尝试解析JSON
            in_progress = False
            try:
                checkpoints_data = json.loads(checkpoints_json)
                
                # 检查历史记录中的savepoint
                history = checkpoints_data.get("history", [])
                for checkpoint in history:
                    if checkpoint.get("is_savepoint") == True:
                        status = checkpoint.get("status")
                        print(f"  历史记录中最新savepoint状态: {status}")
                        
                        if status == "IN_PROGRESS":
                            in_progress = True
                            print("  Savepoint操作正在进行中")
                        elif status == "COMPLETED":
                            completed = True
                            savepoint_path = checkpoint.get("external_path")
                            if savepoint_path:
                                print(f"  Savepoint保存路径: {savepoint_path}")
                            return True
                        break
                
                # 如果历史记录中没有找到，检查最新的savepoint
                latest_savepoints = checkpoints_data.get("latest", {}).get("savepoints", [])
                if latest_savepoints:
                    latest = latest_savepoints[0]
                    status = latest.get("status")
                    is_savepoint = latest.get("is_savepoint")
                    
                    print(f"  最新savepoint状态: {status}, 是否为savepoint: {is_savepoint}")
                    
                    if status == "COMPLETED" and is_savepoint == True:
                        completed = True
                        savepoint_path = latest.get("external_path")
                        if savepoint_path:
                            print(f"  Savepoint保存路径: {savepoint_path}")
                        return True
                    elif status == "IN_PROGRESS" and is_savepoint == True:
                        in_progress = True
                        print("  Savepoint操作正在进行中")
            except json.JSONDecodeError:
                # 使用正则表达式检查
                if '"status":"COMPLETED"' in checkpoints_json and '"is_savepoint":true' in checkpoints_json:
                    completed = True
                    print("  检测到savepoint已完成")
                    
                    path_match = re.search(r'"external_path":"([^"]+)"', checkpoints_json)
                    if path_match:
                        savepoint_path = path_match.group(1)
                        print(f"  Savepoint保存路径: {savepoint_path}")
                    return True
                elif '"status":"IN_PROGRESS"' in checkpoints_json and '"is_savepoint":true' in checkpoints_json:
                    in_progress = True
                    print("  检测到savepoint正在进行中")
            
            if in_progress:
                print("  Savepoint操作正在进行中，等待30秒后再次检查...")
            else:
                print("  未检测到正在进行的savepoint操作，等待30秒后再次检查...")
            
            # 等待30秒后再次检查
            time.sleep(30)
            
        except Exception as e:
            print(f"  检查savepoint状态出错: {e}")
            time.sleep(30)
    
    # 超时
    elapsed = int(time.time() - start_time)
    print(f"  Savepoint操作超时或失败！已等待 {elapsed} 秒，超过最大等待时间 {timeout} 秒")
    return False


def create_savepoint(job: Dict, savepoint_dir: str) -> None:
    #"""为指定任务创建savepoint"""
    app_id = job["app_id"]
    app_name = job["name"]
    
    # 创建包含任务名称的savepoint路径
    task_savepoint_dir = f"{savepoint_dir}/{app_name}"
    
    print(f"\n处理任务: {app_id} ({app_name})")
    
    # 获取Tracking URL
    tracking_url = get_tracking_url(app_id)
    if not tracking_url:
        print("  无法获取Tracking URL")
        return
    
    # 从Tracking URL提取主机
    host = extract_host_from_url(tracking_url)
    if not host:
        print("  无法从Tracking URL提取主机信息")
        return
    
    # 获取Flink版本
    flink_version = get_flink_version(host)
    
    # 获取JobID
    job_id = get_flink_job_id(host)
    if not job_id:
        print("  无法获取Flink JobID")
        return
    
    # 提取主版本号和次版本号
    major_version, minor_version = flink_version.split(".")[:2]
    
    # 针对 Flink 1.18 版本跳过 savepoint 操作
    #if major_version == "1" and minor_version == "18":
    #    print("  检测到 Flink 1.18 版本任务，根据配置跳过 savepoint 操作")
    #    return
    
    # 获取对应版本的Flink命令
    flink_cmd = get_flink_command(flink_version)
    
    # 构建savepoint命令
    savepoint_cmd = f"{flink_cmd} savepoint {job_id} {task_savepoint_dir} -yid {app_id}"
    print(f"  使用Flink SAVEPOINT_CMD 的命令格式")
    print(f"  执行命令: {savepoint_cmd}")
    
    # 执行savepoint命令
    output, status = run_command(savepoint_cmd)
    print(f"  命令输出: {output}")
    print(f"  命令返回状态: {status}")
    
    # 检查savepoint状态
    success = check_savepoint_status(host, job_id)
    
    if success:
        print(f"  Savepoint创建成功！保存在: {task_savepoint_dir}")
    else:
        print(f"  任务: {app_id} ({app_name})：Savepoint操作失败！")


def main():
    #"""主函数"""
    # 获取所有运行中的Flink任务
    running_jobs = get_running_flink_jobs()
    
    if not running_jobs:
        print("未找到正在运行的Flink任务。")
        return
    
    # 过滤任务
    filtered_jobs = filter_jobs(running_jobs, FILTERS)
    
    if not filtered_jobs:
        print("没有匹配的任务，退出操作。")
        return
    
    print(f"\n开始为匹配的 {len(filtered_jobs)} 个任务创建savepoint...")
    
    # 请在此处设置savepoint目录
    savepoint_dir = "hdfs:///flink-savepoints"  # hdfs     

    # 为每个任务创建savepoint
    for job in filtered_jobs:
        create_savepoint(job, savepoint_dir)
    
    print("\n所有匹配任务的savepoint操作已完成！")


if __name__ == "__main__":
    main()

2.3 异常处理策略

• 超时重试机制 ：对于比较大的状态的作业，默认会检查是完成，超时1小时

• 状态验证 ：完成后检查HDFS目录是否存在_metadata文件

• 报警集成：告警可以按需添加即可

三、生产环境最佳实践

3.1 安全增强建议

1. 权限隔离：为Savepoint目录设置专属HDFS账户，避免作业间相互覆盖
2. 存储限额：通过HDFS Quota限制目录容量，防止磁盘写满
3. 生命周期管理：添加定期清理脚本，保留最近3次Savepoint

3.2 性能优化方案

是 否 触发请求 集群负载>80%? 进入等待队列 立即执行 每5分钟重试检查

四、运维监控方案

通过Prometheus+Grafana实现可视化监控：

1. 统计成功率指标
2. 记录单次触发耗时
3. 跟踪HDFS存储用量增长趋势