Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

llm_requests_hanging alert fires continuously for requests with response time below 600s threshold

What happened?

The llm_requests_hanging Slack alert fires continuously reporting "Requests are hanging - 600s+ request time" for multiple models, even though the actual response times (difference between start_time and end_time) are well below the 600-second threshold.

Slack Alert Details

• Alert type: llm_requests_hanging (Digest)
• Level: Medium
• Message: "Requests are hanging - 600s+ request time"
• Affected models: gemini-3-pro-preview, azure/gpt-5.1, and others
• Behavior: Alerts fire continuously with identical start/end timestamps (e.g., Start: 2026-05-13 14:51:55, End: 2026-05-13 14:51:55)

Actual Log Data (max_time_diff_seconds between start and end)

Key finding: The alerted models (gemini-3-pro-preview = 334s, azure/gpt-5.1 = 205s) have max_time_diff_seconds below 600s, yet the alert fires claiming "600s+ request time."

Relevant log output

Alert type: llm_requests_hanging (Digest) Level: Medium
Start: 2026-05-13 14:51:55 End: 2026-05-13 14:51:55
Count: 1
Message: Requests are hanging - 600s+ request time
Request Model: gemini-3-pro-preview API Base: None Key Alias: Team Alias:

Alert type: llm_requests_hanging (Digest) Level: Medium
Start: 2026-05-13 14:51:55 End: 2026-05-13 14:51:55
Count: 2
Message: Requests are hanging - 600s+ request time
Request Model: azure/gpt-5.1 API Base: None Key Alias: Team Alias:

Note: The Start and End timestamps are identical, which suggests the request completed near-instantly or the timestamps are not being set correctly in the alert context.

Expected behavior

1. The llm_requests_hanging alert should only fire for requests that are genuinely still in-flight (no response received yet) and have exceeded 600 seconds without completing.
2. Once a request completes successfully (success or failure callback fires), it should be immediately removed from the hanging request tracker.
3. Alerts should not fire repeatedly/continuously for the same request if the condition has already been evaluated or resolved.
4. The start and end timestamps in the alert should not be identical — identical timestamps suggest the request completed instantly, contradicting the "hanging" classification.

Configuration

general_settings:
  alerting: ["slack"]
  alert_type_config:
    llm_requests_hanging:
      digest: true
      digest_interval: 3600  # 1 hour
    llm_too_slow:
      digest: true
      digest_interval: 3600  # 1 hour
    llm_exceptions:
      digest: true

Additional context

• Running LiteLLM proxy deployed on AWS ECS with Slack alerting configured
• The request_hanging_alerting_threshold is set to 600 seconds (default)
• This causes significant alert fatigue as the team receives continuous false-positive notifications in Slack
• The issue affects all model providers (Azure OpenAI, Google Gemini, AWS Bedrock)
• All requests shown in the log data completed successfully (status = success) with response times well under 600s
• The alerts appear to fire in a digest batch but reference requests that have already completed

Suggested Fix

• In the hanging request checker, ensure only requests without a recorded end_time (i.e., genuinely in-flight) are evaluated against the threshold.
• When async_log_success_event or async_log_failure_event fires, immediately remove the request from the in-flight tracking dictionary.
• Add deduplication logic so the same request ID is not alerted on multiple times.

Steps to Reproduce

Configuration

general_settings:
  alerting: ["slack"]
  alert_type_config:
    llm_requests_hanging:
      digest: true
      digest_interval: 3600  # 1 hour
    llm_too_slow:
      digest: true
      digest_interval: 3600  # 1 hour
    llm_exceptions:
      digest: true

Relevant log output

Alert type: llm_requests_hanging (Digest) Level: Medium
Start: 2026-05-13 14:51:55 End: 2026-05-13 14:51:55
Count: 1
Message: Requests are hanging - 600s+ request time
Request Model: gemini-3-pro-preview API Base: None Key Alias: Team Alias:

Alert type: llm_requests_hanging (Digest) Level: Medium
Start: 2026-05-13 14:51:55 End: 2026-05-13 14:51:55
Count: 2
Message: Requests are hanging - 600s+ request time
Request Model: azure/gpt-5.1 API Base: None Key Alias: Team Alias:

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

1.83.13

Twitter / LinkedIn details

No response