Automatically reattempt after internal errors

apps/webapp/app/components/runs/v3/RunInspector.tsx

@@ -40,7 +40,7 @@ import {
 } from "~/utils/pathBuilder";
 import { TraceSpan } from "~/utils/taskEvent";
 import { SpanLink } from "~/v3/eventRepository.server";
-import { isFinalRunStatus } from "~/v3/taskStatus";
+import { isFailedRunStatus, isFinalRunStatus } from "~/v3/taskStatus";
 import { RunTimelineEvent, RunTimelineLine } from "./InspectorTimeline";
 import { RunTag } from "./RunTag";
 import { TaskRunStatusCombo } from "./TaskRunStatus";
@@ -479,6 +479,7 @@ function RunTimeline({ run }: { run: RawRun }) {
   const updatedAt = new Date(run.updatedAt);
 
   const isFinished = isFinalRunStatus(run.status);
+  const isError = isFailedRunStatus(run.status);
 
   return (
     <div className="min-w-fit max-w-80">
@@ -535,7 +536,7 @@ function RunTimeline({ run }: { run: RawRun }) {
               <RunTimelineEvent
                 title="Finished"
                 subtitle={<DateTimeAccurate date={updatedAt} />}
-                state="complete"
+                state={isError ? "error" : "complete"}
               />
             </>
           ) : (

apps/webapp/app/presenters/v3/SpanPresenter.server.ts

@@ -7,7 +7,7 @@ import {
 import { RUNNING_STATUSES } from "~/components/runs/v3/TaskRunStatus";
 import { eventRepository } from "~/v3/eventRepository.server";
 import { machinePresetFromName } from "~/v3/machinePresets.server";
-import { FINAL_ATTEMPT_STATUSES, isFinalRunStatus } from "~/v3/taskStatus";
+import { FINAL_ATTEMPT_STATUSES, isFailedRunStatus, isFinalRunStatus } from "~/v3/taskStatus";
 import { BasePresenter } from "./basePresenter.server";
 import { getMaxDuration } from "~/v3/utils/maxDuration";
 
@@ -294,6 +294,7 @@ export class SpanPresenter extends BasePresenter {
       usageDurationMs: run.usageDurationMs,
       isFinished,
       isRunning: RUNNING_STATUSES.includes(run.status),
+      isError: isFailedRunStatus(run.status),
       payload,
       payloadType: run.payloadType,
       output,

apps/webapp/app/routes/api.v1.runs.$runParam.attempts.ts

@@ -29,7 +29,10 @@ export async function action({ request, params }: ActionFunctionArgs) {
   const service = new CreateTaskRunAttemptService();
 
   try {
-    const { execution } = await service.call(runParam, authenticationResult.environment);
+    const { execution } = await service.call({
+      runId: runParam,
+      authenticatedEnv: authenticationResult.environment,
+    });
 
     return json(execution, { status: 200 });
   } catch (error) {

apps/webapp/app/routes/resources.orgs.$organizationSlug.projects.v3.$projectParam.runs.$runParam.spans.$spanParam/route.tsx

@@ -857,7 +857,7 @@ function RunTimeline({ run }: { run: SpanRun }) {
               <RunTimelineEvent
                 title="Finished"
                 subtitle={<DateTimeAccurate date={run.updatedAt} />}
-                state="complete"
+                state={run.isError ? "error" : "complete"}
               />
             </>
           ) : (

apps/webapp/app/v3/failedTaskRun.server.ts

@@ -1,15 +1,41 @@
-import { sanitizeError, TaskRunFailedExecutionResult } from "@trigger.dev/core/v3";
+import {
+  calculateNextRetryDelay,
+  RetryOptions,
+  TaskRunExecution,
+  TaskRunExecutionRetry,
+  TaskRunFailedExecutionResult,
+} from "@trigger.dev/core/v3";
 import { logger } from "~/services/logger.server";
-import { createExceptionPropertiesFromError, eventRepository } from "./eventRepository.server";
 import { BaseService } from "./services/baseService.server";
-import { FinalizeTaskRunService } from "./services/finalizeTaskRun.server";
-import { FAILABLE_RUN_STATUSES } from "./taskStatus";
+import { isFailableRunStatus, isFinalAttemptStatus } from "./taskStatus";
+import type { Prisma, TaskRun } from "@trigger.dev/database";
+import { CompleteAttemptService } from "./services/completeAttempt.server";
+import { CreateTaskRunAttemptService } from "./services/createTaskRunAttempt.server";
+import { sharedQueueTasks } from "./marqs/sharedQueueConsumer.server";
+import * as semver from "semver";
+
+const includeAttempts = {
+  attempts: {
+    orderBy: {
+      createdAt: "desc",
+    },
+    take: 1,
+  },
+  lockedBy: true, // task
+  lockedToVersion: true, // worker
+} satisfies Prisma.TaskRunInclude;
+
+type TaskRunWithAttempts = Prisma.TaskRunGetPayload<{
+  include: typeof includeAttempts;
+}>;
 
 export class FailedTaskRunService extends BaseService {
   public async call(anyRunId: string, completion: TaskRunFailedExecutionResult) {
+    logger.debug("[FailedTaskRunService] Handling failed task run", { anyRunId, completion });
+
     const isFriendlyId = anyRunId.startsWith("run_");
 
-    const taskRun = await this._prisma.taskRun.findUnique({
+    const taskRun = await this._prisma.taskRun.findFirst({
       where: {
         friendlyId: isFriendlyId ? anyRunId : undefined,
         id: !isFriendlyId ? anyRunId : undefined,
@@ -25,7 +51,7 @@ export class FailedTaskRunService extends BaseService {
       return;
     }
 
-    if (!FAILABLE_RUN_STATUSES.includes(taskRun.status)) {
+    if (!isFailableRunStatus(taskRun.status)) {
       logger.error("[FailedTaskRunService] Task run is not in a failable state", {
         taskRun,
         completion,
@@ -34,33 +60,217 @@ export class FailedTaskRunService extends BaseService {
       return;
     }
 
-    // No more retries, we need to fail the task run
-    logger.debug("[FailedTaskRunService] Failing task run", { taskRun, completion });
+    const retryHelper = new FailedTaskRunRetryHelper(this._prisma);
+    const retryResult = await retryHelper.call({
+      runId: taskRun.id,
+      completion,
+    });
 
-    const finalizeService = new FinalizeTaskRunService();
-    await finalizeService.call({
-      id: taskRun.id,
-      status: "SYSTEM_FAILURE",
-      completedAt: new Date(),
-      attemptStatus: "FAILED",
-      error: sanitizeError(completion.error),
+    logger.debug("[FailedTaskRunService] Completion result", {
+      runId: taskRun.id,
+      result: retryResult,
     });
+  }
+}
+
+interface TaskRunWithWorker extends TaskRun {
+  lockedBy: { retryConfig: Prisma.JsonValue } | null;
+  lockedToVersion: { sdkVersion: string } | null;
+}
 
-    // Now we need to "complete" the task run event/span
-    await eventRepository.completeEvent(taskRun.spanId, {
-      endTime: new Date(),
-      attributes: {
-        isError: true,
+export class FailedTaskRunRetryHelper extends BaseService {
+  async call({
+    runId,
+    completion,
+    isCrash,
+  }: {
+    runId: string;
+    completion: TaskRunFailedExecutionResult;
+    isCrash?: boolean;
+  }) {
+    const taskRun = await this._prisma.taskRun.findFirst({
+      where: {
+        id: runId,
       },
-      events: [
-        {
-          name: "exception",
-          time: new Date(),
-          properties: {
-            exception: createExceptionPropertiesFromError(completion.error),
-          },
-        },
-      ],
+      include: includeAttempts,
     });
+
+    if (!taskRun) {
+      logger.error("[FailedTaskRunRetryHelper] Task run not found", {
+        runId,
+        completion,
+      });
+
+      return "NO_TASK_RUN";
+    }
+
+    const retriableExecution = await this.#getRetriableAttemptExecution(taskRun, completion);
+
+    if (!retriableExecution) {
+      return "NO_EXECUTION";
+    }
+
+    logger.debug("[FailedTaskRunRetryHelper] Completing attempt", { taskRun, completion });
+
+    const executionRetry =
+      completion.retry ??
+      (await FailedTaskRunRetryHelper.getExecutionRetry({
+        run: taskRun,
+        execution: retriableExecution,
+      }));
+
+    const completeAttempt = new CompleteAttemptService(this._prisma);
+    const completeResult = await completeAttempt.call({
+      completion: {
+        ...completion,
+        retry: executionRetry,
+      },
+      execution: retriableExecution,
+      isSystemFailure: !isCrash,
+      isCrash,
+    });
+
+    return completeResult;
+  }
+
+  async #getRetriableAttemptExecution(
+    run: TaskRunWithAttempts,
+    completion: TaskRunFailedExecutionResult
+  ): Promise<TaskRunExecution | undefined> {
+    let attempt = run.attempts[0];
+
+    // We need to create an attempt if:
+    // - None exists yet
+    // - The last attempt has a final status, e.g. we failed between attempts
+    if (!attempt || isFinalAttemptStatus(attempt.status)) {
+      logger.error("[FailedTaskRunRetryHelper] No attempts found", {
+        run,
+        completion,
+      });
+
+      const createAttempt = new CreateTaskRunAttemptService(this._prisma);
+
+      try {
+        const { execution } = await createAttempt.call({
+          runId: run.id,
+          // This ensures we correctly respect `maxAttempts = 1` when failing before the first attempt was created
+          startAtZero: true,
+        });
+        return execution;
+      } catch (error) {
+        logger.error("[FailedTaskRunRetryHelper] Failed to create attempt", {
+          run,
+          completion,
+          error,
+        });
+
+        return;
+      }
+    }
+
+    // We already have an attempt with non-final status, let's use it
+    try {
+      const executionPayload = await sharedQueueTasks.getExecutionPayloadFromAttempt({
+        id: attempt.id,
+        skipStatusChecks: true,
+      });
+
+      return executionPayload?.execution;
+    } catch (error) {
+      logger.error("[FailedTaskRunRetryHelper] Failed to get execution payload", {
+        run,
+        completion,
+        error,
+      });
+
+      return;
+    }
   }
+
+  static async getExecutionRetry({
+    run,
+    execution,
+  }: {
+    run: TaskRunWithWorker;
+    execution: TaskRunExecution;
+  }): Promise<TaskRunExecutionRetry | undefined> {
+    try {
+      const retryConfig = run.lockedBy?.retryConfig;
+
+      if (!retryConfig) {
+        if (!run.lockedToVersion) {
+          logger.error("[FailedTaskRunRetryHelper] Run not locked to version", {
+            run,
+            execution,
+          });
+
+          return;
+        }
+
+        const sdkVersion = run.lockedToVersion.sdkVersion ?? "0.0.0";
+        const isValid = semver.valid(sdkVersion);
+
+        if (!isValid) {
+          logger.error("[FailedTaskRunRetryHelper] Invalid SDK version", {
+            run,
+            execution,
+          });
+
+          return;
+        }
+
+        // With older SDK versions, tasks only have a retry config stored in the DB if it's explicitly defined on the task itself
+        // It won't get populated with retry.default in trigger.config.ts
+        if (semver.lt(sdkVersion, FailedTaskRunRetryHelper.DEFAULT_RETRY_CONFIG_SINCE_VERSION)) {
+          logger.warn(
+            "[FailedTaskRunRetryHelper] SDK version not recent enough to determine retry config",
+            {
+              run,
+              execution,
+            }
+          );
+
+          return;
+        }
+      }
+
+      const parsedRetryConfig = RetryOptions.safeParse(retryConfig);
+
+      if (!parsedRetryConfig.success) {
+        logger.error("[FailedTaskRunRetryHelper] Invalid retry config", {
+          run,
+          execution,
+        });
+
+        return;
+      }
+
+      const delay = calculateNextRetryDelay(parsedRetryConfig.data, execution.attempt.number);
+
+      if (!delay) {
+        logger.debug("[FailedTaskRunRetryHelper] No more retries", {
+          run,
+          execution,
+        });
+
+        return;
+      }
+
+      return {
+        timestamp: Date.now() + delay,
+        delay,
+      };
+    } catch (error) {
+      logger.error("[FailedTaskRunRetryHelper] Failed to get execution retry", {
+        run,
+        execution,
+        error,
+      });
+
+      return;
+    }
+  }
+
+  // TODO: update this to the correct version
+  static DEFAULT_RETRY_CONFIG_SINCE_VERSION = "3.0.12";
 }

apps/webapp/app/v3/handleSocketIo.server.ts

@@ -193,7 +193,11 @@ function createCoordinatorNamespace(io: Server) {
           }
 
           const service = new CreateTaskRunAttemptService();
-          const { attempt } = await service.call(message.runId, environment, false);
+          const { attempt } = await service.call({
+            runId: message.runId,
+            authenticatedEnv: environment,
+            setToExecuting: false,
+          });
 
           const payload = await sharedQueueTasks.getExecutionPayloadFromAttempt({
             id: attempt.id,

apps/webapp/app/v3/marqs/sharedQueueConsumer.server.ts

@@ -509,7 +509,10 @@ export class SharedQueueConsumer {
           if (!deployment.worker.supportsLazyAttempts) {
             try {
               const service = new CreateTaskRunAttemptService();
-              await service.call(lockedTaskRun.friendlyId, undefined, false);
+              await service.call({
+                runId: lockedTaskRun.id,
+                setToExecuting: false,
+              });
             } catch (error) {
               logger.error("Failed to create task run attempt for outdate worker", {
                 error,