Custom training job stuck in QUEUED state for hours

By Sai Kiran Pandrala · Last verified: 2026-05-31 · Source: Google Cloud docs, Google Cloud Community, community Q&A

If you hit Custom training job stuck in QUEUED state for hours on Vertex AI Training in production, the steps below are the path most teams take in 2026. None of them require opening a support case unless your environment has a paid-tier dependency that Google Cloud owns.

What custom training job stuck in queued state for hours actually involves on Vertex AI Training

This task on Vertex AI Training is one of the more searched operational topics on AWS in the last 12 months. The procedure below is the path that works in a current AWS account with default IAM and standard VPC config.

The rest of this page is the structured fix path. Start with diagnose, then remediation, then the automation options so you do not have to do this by hand the next time it surfaces. Verify and safety sections at the end are the discipline that keeps the fix from regressing in production.

Diagnose first, fix second

Run gcloud auth list and gcloud config list first. About one in five 'why does this not work' tickets are actually 'I am in the wrong account' or 'my session expired and the SDK is using stale credentials or ADC pointed at the wrong project'. The 5-second sanity check costs nothing and saves real time when the answer is that simple.

Look at the Cloud Audit Log event for the failed call, even if you are not enrolled in Cloud Logging Log Router. The basic 90-day event history works for most diagnostic purposes and lives in the console under Cloud Audit Logs > Event history. Filter by event name (the API action) and time range; the event JSON shows the exact user identity, source IP, request parameters, and error code.

Check Cloud Monitoring Logs for the calling service. Lambda, ECS, EKS, Step Functions, API Gateway, and most managed services write detailed traces to Cloud Monitoring Logs under predictable log group names. Use Cloud Monitoring Logs Insights with fields @timestamp, @message | filter @message like /ERROR/ | sort @timestamp desc | limit 50 to surface the most recent failures.

Solution-focused remediation path

When the fix involves a destructive operation (delete VPC endpoint, swap Cloud KMS key, rotate root credential), do it during a maintenance window with at least one teammate watching. Several Vertex AI Training operations have implicit dependencies that only show up when traffic starts flowing again. Document the rollback path before you start, not during the incident.

If networking is suspect, use Network Intelligence Connectivity Tests. It is the only tool that simulates the full ENI-to-ENI path including firewall rules, hierarchical firewall policies, routes, and VPC Service Controls perimeters in one call. Manual trace is slower and misses transitive issues. The analyzer charges $0.10 per analysis - cheaper than a 30-minute call with your network team.

For IAM and STS issues, the timing matters. STS sessions can take up to 60 seconds to propagate after creation. The first call right after assume-role can fail with a permission error even when the policy is correct. Add a small retry with backoff before treating the first failure as definitive.

Automate this fix so you do not do it twice

Add a Workflows or Cloud Tasks Automation runbook

For multi-step fixes that include a manual approval, use Workflows runbook. Document the fix as a runbook with workflows.executions.approve steps where a human signs off and workflows.steps.callApi steps where the runbook calls the Google Cloud API. Approvers are notified by SNS; the runbook execution shows up in Cloud Audit Logs with the approver's identity attached. This makes audit trails easy and stops production fixes from being one-person operations.

Automate the fix with Python and boto3

For anything you do more than twice, write a small Python script. The boto3 pattern below uses paginators (so it does not blow up on accounts with thousands of resources), explicit region binding, and a dry-run flag that defaults to True. Keep the script under 100 lines; if it grows beyond that, you are building a tool and should put it behind a Lambda with proper logging.

import boto3, sys
DRY_RUN = '--apply' not in sys.argv
client = boto3.client('vertex', region_name='us-east-1')
paginator = client.get_paginator('describe_...')
for page in paginator.paginate(): for item in page.get('Items', []): if item.get('Status') == 'FAILED': if DRY_RUN: print(f'[dry-run] would fix {item["Id"]}') else: client.modify_...(ResourceId=item['Id']) print(f'fixed {item["Id"]}')

Wire the fix into Eventarc for self-healing

If the failure mode is recurring, automate the remediation instead of the diagnosis. Eventarc Scheduler or rules that watch Cloud Logging events for the specific error code can invoke a Lambda that runs the same fix you would run by hand. The Lambda must be idempotent (re-running it on already-healthy resources must be a no-op) and must emit a Cloud Monitoring metric so you can track how often the auto-fix fires. A spike in auto-fix invocations is itself a signal worth alerting on.

# Eventarc rule pattern (JSON)
{ "source": ["aws.vertex"], "detail-type": ["Google Cloud API Call via Cloud Audit Logs"], "detail": { "errorCode": ["AccessDenied", "ThrottlingException"] }
}

Common pitfalls and what to watch for

The pitfall most teams hit on Vertex AI Training is moving too fast and skipping the read-only validation step. Before any write, list the current state and save it. Google Cloud APIs are eventually consistent for many resource types, so the validation snapshot is your only reliable reference if you need to undo. Save the output of the describe call to S3, not to your laptop.

Second pitfall: confusing IAM permission errors with networking errors. AccessDenied can be IAM (policy missing), networking (VPC endpoint policy blocking the call), or KMS (key policy missing). The error string looks identical for all three. Distinguish by looking at the Cloud Audit Log event's errorCode and the encoded authorization message; do not assume IAM is the culprit just because the message says AccessDenied.

Verify the fix worked

• Reproduce the original symptom path. If it still surfaces in any account or region or IAM role or service account, you have not fixed it.
• Watch for 24 to 48 hours. Cloud Monitoring metrics and Cloud Asset Inventory can mask issues with cached health for 6 to 12 hours, especially Cloud CDN and Cloud DNS.
• Run a smoke test under realistic load. Happy-path tests miss race conditions and IAM session-cache issues.
• Capture the new state in a runbook so the next person on call does not have to rediscover this. Push it to Confluence or your team wiki, not into Slack.
• If the fix involved a permission change, run IAM Access Analyzer one more time to confirm you did not open a separate hole while closing this one.

Safety, rollback, blast radius

• Test in a non-production account if your environment has Resource Manager and Organization Policy or Cloud Resource Manager (organizations, folders, projects). The cost of one sandbox account is cheaper than one rollback meeting.
• Export the existing config before changing it. Most Vertex AI Training resources support describe + export to JSON via CLI - capture that to source control before you start.
• Know your rollback path. Some Vertex AI Training operations are one-way (region migration, account-level feature opt-in, Cloud KMS key deletion past pending window). Confirm reversibility on the Google Cloud doc before you commit.
• Be aware of cross-service impact. IAM role or service account changes ripple to every service trusting that role. Cloud KMS key changes break every workload depending on that key. VPC endpoint changes affect every VPC consumer of that endpoint.
• Maintenance window discipline: if the change touches DNS, certificate rotation, or anything that emits TLS handshakes, line up a window with stakeholder notification, not a heroic mid-day swap.

FAQ

Related fixes

Related guides worth a look while you sort this one out:

• Register model from custom training job output
• How to package source distribution for custom training job
• Feature ingestion job stuck backfill schedule
• Batch prediction job stuck in PENDING for over 1 hour
• Set CMEK on training job intermediate artifacts
• Training job timeout how to extend beyond default 7-day cap