Investigating Model Execution Failures In Dev February 7‐9, 2026 - conrad-blucher-institute/semaphore GitHub Wiki

Semaphore Development Environment Analysis

Created from this google doc: dev_analysis_02-09-2026

February 7-9, 2026

Analysis Period: February 7, 2026 12:00 AM - February 9, 2026 5:00 PM

Executive Summary

The development environment shows a 97.85% success rate across 6,777 model executions, with 146 failures requiring attention. The analysis reveals three critical issues:

  • MAGNOLIA TRANSFORM CRISIS: magnolia_transform_48 has a 0% success rate (65/65 failures)
  • LIGHTHOUSE DATA SOURCE: New failure source causing 44% of all failures
  • SEMAPHORE INTERNAL PREDICTIONS: Missing data affecting transform models

Overall Statistics

Total Executions 6,777
Successful 6,631
Failed 146
Success Rate 97.85%
Failure Rate 2.15%

Failure Analysis

Failure Reasons Breakdown

missing_data 98 (67%)
stale_data 48 (33%)

Failures by Data Source

LIGHTHOUSE 64 (44%)
SEMAPHORE 54 (37%)
NOAATANDC 27 (18%)
NDFD_JSON 1 (<1%)

Models with LIGHTHOUSE or NOAATANDC failures.

Model Name Runs Success % Fails Data Source(s) Failure Reason(s)
magnolia_transform_48 65 0.0% 65 LIGHTHOUSE (47, 72%), SEMAPHORE (18, 28%) stale_data (47, 72%), missing_data (18, 28%)
magnolia_transform_12 65 50.77% 32 SEMAPHORE (18, 56%), LIGHTHOUSE (14, 44%) missing_data (32, 100%)
96hr_VirginiaKey_wl 65 93.85% 4 NOAATANDC (3, 75%), NDFD_JSON (1, 25%) missing_data (3, 75%), stale_data (1, 25%)
72hr_VirginiaKey_wl 65 95.38% 3 NOAATANDC (3, 100%) missing_data (3, 100%)
ar_inundation_apr_24h 65 96.92% 2 NOAATANDC (2, 100%) missing_data (2, 100%)
pi_mlp_surge_48h 65 96.92% 2 NOAATANDC (2, 100%) missing_data (2, 100%)
ar_inundation_apr_48h 65 96.92% 2 NOAATANDC (2, 100%) missing_data (2, 100%)
ar_inundation_apr_12h 65 96.92% 2 NOAATANDC (2, 100%) missing_data (2, 100%)
rp_S2S_surge_72h 64 96.88% 2 NOAATANDC (2, 100%) missing_data (2, 100%)
rp_S2S_surge_48h 65 96.92% 2 NOAATANDC (2, 100%) missing_data (2, 100%)
pi_mlp_surge_72h 65 96.92% 2 NOAATANDC (2, 100%) missing_data (2, 100%)
48hr_VirginiaKey_wl 64 98.44% 1 NOAATANDC (1, 100%) missing_data (1, 100%)
magnolia_12 64 98.44% 1 LIGHTHOUSE (1, 100%) missing_data (1, 100%)
nj_mlp_surge_24h 44 97.73% 1 NOAATANDC (1, 100%) missing_data (1, 100%)
nj_mlp_surge_12h 44 97.73% 1 NOAATANDC (1, 100%) missing_data (1, 100%)
magnolia_24 64 98.44% 1 LIGHTHOUSE (1, 100%) missing_data (1, 100%)
magnolia_48 64 98.44% 1 LIGHTHOUSE (1, 100%) missing_data (1, 100%)
pi_mlp_surge_24h 65 98.46% 1 NOAATANDC (1, 100%) missing_data (1, 100%)
pi_mlp_surge_12h 65 98.46% 1 NOAATANDC (1, 100%) missing_data (1, 100%)
nj_mlp_surge_72h 44 97.73% 1 NOAATANDC (1, 100%) missing_data (1, 100%)
nj_mlp_surge_48h 44 97.73% 1 NOAATANDC (1, 100%) missing_data (1, 100%)

Critical Issues & Recommended Tickets

ISSUE #1: Magnolia Transform Models

Impact: 115 failures (78.8% of all failures)

magnolia_transform_48

  • Status: 0% success rate (0/65 successful)
  • Primary Issue: Stale LIGHTHOUSE data (47) + Missing SEMAPHORE data (18)

magnolia_transform_12

  • Status: 50.8% success rate (33/65 successful)
  • Primary Issue: Missing SEMAPHORE (18) + LIGHTHOUSE (14)

magnolia_transform_24

  • Status: 72.3% success rate (47/65 successful)
  • Primary Issue: Missing SEMAPHORE data (18)

Recommended Tickets

TICKET 1: Investigate LIGHTHOUSE Data Source

  • Impact: 64 failures (44% of total)
  • What is going on with thissssss 🎶
  • Three potential fixes
  • 1. Rethink Absolute value… nah -> add comment about britalness
  • 2. Change lighthouse to set generated time to now time for predicted values

TICKET 2: Fix Magnolia Transform LIGHTHOUSE Staleness

  • Impact: magnolia_transform_48 has 0% success rate
  • Action: Review and adjust staleness thresholds for LIGHTHOUSE inputs

TICKET 3: Investigate Magnolia Transform Missing SEMAPHORE Data

  • Impact: 54 failures across all transform models
  • I fear this is due to Lighthouse issues for the base magnolia models, ticket below. Whenever the base models have one failure we have to wait until there's a full uninterrupted 12 hour run before we can run the transforms again.

TICKET 4: Monitor LIGHTHOUSE Integration for Base Magnolia Models

  • Impact: 3 failures (1 per base magnolia model)

TICKET 5: Investigate NOAATANDC 01:00 UTC Data Gap

  • Impact: 27 failures, primarily at 01:00 UTC

Key Insights

  • NDFD issue from previous report is FIXED - only 1 failure (was 58)
  • Overall system reliability improved +1.21% compared to previous analysis
  • LIGHTHOUSE is a new data source requiring immediate attention
  • Failure type shifted from stale_data (94%) to missing_data (67%)
  • 84 models maintain perfect 100% success rate

Summary

The development environment shows improved overall performance with excellent resolution of the NDFD staleness issue. However, LIGHTHOUSE integration and magnolia_transform models require immediate attention. The shift from stale_data to missing_data suggests threshold fixes revealed underlying data availability problems.