Azure Compute Batch Java

Azure Batch SDK for Java

用于在 Azure 中运行大规模并行和高性能计算（HPC）批处理作业的客户端库。

安装

<dependency>
    <groupId>com.azure</groupId>
    <artifactId>azure-compute-batch</artifactId>
    <version>1.0.0-beta.5</version>
</dependency>

前提条件

• Azure Batch 账户
• 已配置计算节点的池
• Azure 订阅

环境变量

AZURE_BATCH_ENDPOINT=https://<account>.<region>.batch.azure.com
AZURE_BATCH_ACCOUNT=<account-name>
AZURE_BATCH_ACCESS_KEY=<account-key>

创建客户端

使用 Microsoft Entra ID（推荐）

import com.azure.compute.batch.BatchClient;
import com.azure.compute.batch.BatchClientBuilder;
import com.azure.identity.DefaultAzureCredentialBuilder;

BatchClient batchClient = new BatchClientBuilder()
    .credential(new DefaultAzureCredentialBuilder().build())
    .endpoint(System.getenv("AZURE_BATCH_ENDPOINT"))
    .buildClient();

异步客户端

import com.azure.compute.batch.BatchAsyncClient;

BatchAsyncClient batchAsyncClient = new BatchClientBuilder()
    .credential(new DefaultAzureCredentialBuilder().build())
    .endpoint(System.getenv("AZURE_BATCH_ENDPOINT"))
    .buildAsyncClient();

使用共享密钥凭据

import com.azure.core.credential.AzureNamedKeyCredential;

String accountName = System.getenv("AZURE_BATCH_ACCOUNT");
String accountKey = System.getenv("AZURE_BATCH_ACCESS_KEY");
AzureNamedKeyCredential sharedKeyCreds = new AzureNamedKeyCredential(accountName, accountKey);

BatchClient batchClient = new BatchClientBuilder()
    .credential(sharedKeyCreds)
    .endpoint(System.getenv("AZURE_BATCH_ENDPOINT"))
    .buildClient();

核心概念

池操作

创建池

import com.azure.compute.batch.models.*;

batchClient.createPool(new BatchPoolCreateParameters("myPoolId", "STANDARD_DC2s_V2")
    .setVirtualMachineConfiguration(
        new VirtualMachineConfiguration(
            new BatchVmImageReference()
                .setPublisher("Canonical")
                .setOffer("UbuntuServer")
                .setSku("22_04-lts")
                .setVersion("latest"),
            "batch.node.ubuntu 22.04"))
    .setTargetDedicatedNodes(2)
    .setTargetLowPriorityNodes(0), null);

获取池

BatchPool pool = batchClient.getPool("myPoolId");
System.out.println("Pool state: " + pool.getState());
System.out.println("Current dedicated nodes: " + pool.getCurrentDedicatedNodes());

列出池

import com.azure.core.http.rest.PagedIterable;

PagedIterable<BatchPool> pools = batchClient.listPools();
for (BatchPool pool : pools) {
    System.out.println("Pool: " + pool.getId() + ", State: " + pool.getState());
}

调整池大小

import com.azure.core.util.polling.SyncPoller;

BatchPoolResizeParameters resizeParams = new BatchPoolResizeParameters()
    .setTargetDedicatedNodes(4)
    .setTargetLowPriorityNodes(2);

SyncPoller<BatchPool, BatchPool> poller = batchClient.beginResizePool("myPoolId", resizeParams);
poller.waitForCompletion();
BatchPool resizedPool = poller.getFinalResult();

启用自动缩放

BatchPoolEnableAutoScaleParameters autoScaleParams = new BatchPoolEnableAutoScaleParameters()
    .setAutoScaleEvaluationInterval(Duration.ofMinutes(5))
    .setAutoScaleFormula("$TargetDedicatedNodes = min(10, $PendingTasks.GetSample(TimeInterval_Minute * 5));");

batchClient.enablePoolAutoScale("myPoolId", autoScaleParams);

删除池

SyncPoller<BatchPool, Void> deletePoller = batchClient.beginDeletePool("myPoolId");
deletePoller.waitForCompletion();

作业操作

创建作业

batchClient.createJob(
    new BatchJobCreateParameters("myJobId", new BatchPoolInfo().setPoolId("myPoolId"))
        .setPriority(100)
        .setConstraints(new BatchJobConstraints()
            .setMaxWallClockTime(Duration.ofHours(24))
            .setMaxTaskRetryCount(3)),
    null);

获取作业

BatchJob job = batchClient.getJob("myJobId", null, null);
System.out.println("Job state: " + job.getState());

列出作业

PagedIterable<BatchJob> jobs = batchClient.listJobs(new BatchJobsListOptions());
for (BatchJob job : jobs) {
    System.out.println("Job: " + job.getId() + ", State: " + job.getState());
}

获取任务计数

BatchTaskCountsResult counts = batchClient.getJobTaskCounts("myJobId");
System.out.println("Active: " + counts.getTaskCounts().getActive());
System.out.println("Running: " + counts.getTaskCounts().getRunning());
System.out.println("Completed: " + counts.getTaskCounts().getCompleted());

终止作业

BatchJobTerminateParameters terminateParams = new BatchJobTerminateParameters()
    .setTerminationReason("Manual termination");
BatchJobTerminateOptions options = new BatchJobTerminateOptions().setParameters(terminateParams);

SyncPoller<BatchJob, BatchJob> poller = batchClient.beginTerminateJob("myJobId", options, null);
poller.waitForCompletion();

删除作业

SyncPoller<BatchJob, Void> deletePoller = batchClient.beginDeleteJob("myJobId");
deletePoller.waitForCompletion();

任务操作

创建单个任务

BatchTaskCreateParameters task = new BatchTaskCreateParameters("task1", "echo 'Hello World'");
batchClient.createTask("myJobId", task);

创建带退出条件的任务

batchClient.createTask("myJobId", new BatchTaskCreateParameters("task2", "cmd /c exit 3")
    .setExitConditions(new ExitConditions()
        .setExitCodeRanges(Arrays.asList(
            new ExitCodeRangeMapping(2, 4, 
                new ExitOptions().setJobAction(BatchJobActionKind.TERMINATE)))))
    .setUserIdentity(new UserIdentity()
        .setAutoUser(new AutoUserSpecification()
            .setScope(AutoUserScope.TASK)
            .setElevationLevel(ElevationLevel.NON_ADMIN))),
    null);

创建任务集合（最多100个）

List<BatchTaskCreateParameters> taskList = Arrays.asList(
    new BatchTaskCreateParameters("task1", "echo Task 1"),
    new BatchTaskCreateParameters("task2", "echo Task 2"),
    new BatchTaskCreateParameters("task3", "echo Task 3")
);
BatchTaskGroup taskGroup = new BatchTaskGroup(taskList);
BatchCreateTaskCollectionResult result = batchClient.createTaskCollection("myJobId", taskGroup);

创建大量任务（无限制）

List<BatchTaskCreateParameters> tasks = new ArrayList<>();
for (int i = 0; i < 1000; i++) {
    tasks.add(new BatchTaskCreateParameters("task" + i, "echo Task " + i));
}
batchClient.createTasks("myJobId", tasks);

获取任务

BatchTask task = batchClient.getTask("myJobId", "task1");
System.out.println("Task state: " + task.getState());
System.out.println("Exit code: " + task.getExecutionInfo().getExitCode());

列出任务

PagedIterable<BatchTask> tasks = batchClient.listTasks("myJobId");
for (BatchTask task : tasks) {
    System.out.println("Task: " + task.getId() + ", State: " + task.getState());
}

获取任务输出

import com.azure.core.util.BinaryData;
import java.nio.charset.StandardCharsets;

BinaryData stdout = batchClient.getTaskFile("myJobId", "task1", "stdout.txt");
System.out.println(new String(stdout.toBytes(), StandardCharsets.UTF_8));

终止任务

batchClient.terminateTask("myJobId", "task1", null, null);

节点操作

列出节点

PagedIterable<BatchNode> nodes = batchClient.listNodes("myPoolId", new BatchNodesListOptions());
for (BatchNode node : nodes) {
    System.out.println("Node: " + node.getId() + ", State: " + node.getState());
}

重启节点

SyncPoller<BatchNode, BatchNode> rebootPoller = batchClient.beginRebootNode("myPoolId", "nodeId");
rebootPoller.waitForCompletion();

获取远程登录设置

BatchNodeRemoteLoginSettings settings = batchClient.getNodeRemoteLoginSettings("myPoolId", "nodeId");
System.out.println("IP: " + settings.getRemoteLoginIpAddress());
System.out.println("Port: " + settings.getRemoteLoginPort());

作业计划操作

创建作业计划

batchClient.createJobSchedule(new BatchJobScheduleCreateParameters("myScheduleId",
    new BatchJobScheduleConfiguration()
        .setRecurrenceInterval(Duration.ofHours(6))
        .setDoNotRunUntil(OffsetDateTime.now().plusDays(1)),
    new BatchJobSpecification(new BatchPoolInfo().setPoolId("myPoolId"))
        .setPriority(50)),
    null);

获取作业计划

BatchJobSchedule schedule = batchClient.getJobSchedule("myScheduleId");
System.out.println("Schedule state: " + schedule.getState());

错误处理

import com.azure.compute.batch.models.BatchErrorException;
import com.azure.compute.batch.models.BatchError;

try {
    batchClient.getPool("nonexistent-pool");
} catch (BatchErrorException e) {
    BatchError error = e.getValue();
    System.err.println("Error code: " + error.getCode());
    System.err.println("Message: " + error.getMessage().getValue());
    
    if ("PoolNotFound".equals(error.getCode())) {
        System.err.println("The specified pool does not exist.");
    }
}

最佳实践

1. 使用 Entra ID — 认证时优先选择而非共享密钥
2. 使用管理 SDK 管理池 — azure-resourcemanager-batch 支持托管标识
3. 批量创建任务 — 创建多个任务时使用 createTaskCollection 或 createTasks
4. 正确处理长时间运行操作 — 池调整大小、删除操作是长时间运行的
5. 监控任务计数 — 使用 getJobTaskCounts 跟踪进度
6. 设置约束 — 配置 maxWallClockTime 和 maxTaskRetryCount
7. 使用低优先级节点 — 为容错工作负载节省成本
8. 启用自动缩放 — 根据工作负载动态调整池大小

参考链接

适用场景

本技能适用于执行概述中描述的工作流或操作。

限制

• 仅当任务明确匹配上述范围时使用本技能。
• 不要将输出作为环境特定验证、测试或专家审查的替代品。
• 如果缺少必需的输入、权限、安全边界或成功标准，请停止并请求澄清。