Debug Buttercup

Debug Buttercup

When to Use

• Pods in the crs namespace are in CrashLoopBackOff, OOMKilled, or restarting
• Multiple services restart simultaneously (cascade failure)
• Redis is unresponsive or showing AOF warnings
• Queues are growing but tasks are not progressing
• Nodes show DiskPressure, MemoryPressure, or PID pressure
• Build-bot cannot reach the Docker daemon (DinD failures)
• Scheduler is stuck and not advancing task state
• Health check probes are failing unexpectedly
• Deployed Helm values don't match actual pod configuration

When NOT to Use

• Deploying or upgrading Buttercup (use Helm and deployment guides)
• Debugging issues outside the crs Kubernetes namespace
• Performance tuning that doesn't involve a failure symptom

Namespace and Services

All pods run in namespace crs. Key services:

Triage Workflow

Always start with triage. Run these three commands first:

# 1. Pod status - look for restarts, CrashLoopBackOff, OOMKilled
kubectl get pods -n crs -o wide

# 2. Events - the timeline of what went wrong
kubectl get events -n crs --sort-by='.lastTimestamp'

# 3. Warnings only - filter the noise
kubectl get events -n crs --field-selector type=Warning --sort-by='.lastTimestamp'

Then narrow down:

# Why did a specific pod restart? Check Last State Reason (OOMKilled, Error, Completed)
kubectl describe pod -n crs <pod-name> | grep -A8 'Last State:'

# Check actual resource limits vs intended
kubectl get pod -n crs <pod-name> -o jsonpath='{.spec.containers[0].resources}'

# Crashed container's logs (--previous = the container that died)
kubectl logs -n crs <pod-name> --previous --tail=200

# Current logs
kubectl logs -n crs <pod-name> --tail=200

Historical vs Ongoing Issues

High restart counts don't necessarily mean an issue is ongoing -- restarts accumulate over a pod's lifetime. Always distinguish:

• --tail shows the end of the log buffer, which may contain old messages. Use --since=300s to confirm issues are actively happening now.
• --timestamps on log output helps correlate events across services.
• Check Last State timestamps in describe pod to see when the most recent crash actually occurred.

Cascade Detection

When many pods restart around the same time, check for a shared-dependency failure before investigating individual pods. The most common cascade: Redis goes down -> every service gets ConnectionError/ConnectionRefusedError -> mass restarts. Look for the same error across multiple --previous logs -- if they all say redis.exceptions.ConnectionError, debug Redis, not the individual services.

Log Analysis

# All replicas of a service at once
kubectl logs -n crs -l app=fuzzer-bot --tail=100 --prefix

# Stream live
kubectl logs -n crs -l app.kubernetes.io/name=redis -f

# Collect all logs to disk (existing script)
bash deployment/collect-logs.sh

Resource Pressure

# Per-pod CPU/memory
kubectl top pods -n crs

# Node-level
kubectl top nodes

# Node conditions (disk pressure, memory pressure, PID pressure)
kubectl describe node <node> | grep -A5 Conditions

# Disk usage inside a pod
kubectl exec -n crs <pod> -- df -h

# What's eating disk
kubectl exec -n crs <pod> -- sh -c 'du -sh /corpus/* 2>/dev/null'
kubectl exec -n crs <pod> -- sh -c 'du -sh /scratch/* 2>/dev/null'

Redis Debugging

Redis is the backbone. When it goes down, everything cascades.

# Redis pod status
kubectl get pods -n crs -l app.kubernetes.io/name=redis

# Redis logs (AOF warnings, OOM, connection issues)
kubectl logs -n crs -l app.kubernetes.io/name=redis --tail=200

# Connect to Redis CLI
kubectl exec -n crs <redis-pod> -- redis-cli

# Inside redis-cli: key diagnostics
INFO memory          # used_memory_human, maxmemory
INFO persistence     # aof_enabled, aof_last_bgrewrite_status, aof_delayed_fsync
INFO clients         # connected_clients, blocked_clients
INFO stats           # total_connections_received, rejected_connections
CLIENT LIST          # see who's connected
DBSIZE               # total keys

# AOF configuration
CONFIG GET appendonly     # is AOF enabled?
CONFIG GET appendfsync   # fsync policy: everysec, always, or no

# What is /data mounted on? (disk vs tmpfs matters for AOF performance)

kubectl exec -n crs <redis-pod> -- mount | grep /data
kubectl exec -n crs <redis-pod> -- du -sh /data/

Queue Inspection

Buttercup uses Redis streams with consumer groups. Queue names:

# Check stream length (pending messages)
kubectl exec -n crs <redis-pod> -- redis-cli XLEN fuzzer_build_queue

# Check consumer group lag
kubectl exec -n crs <redis-pod> -- redis-cli XINFO GROUPS fuzzer_build_queue

# Check pending messages per consumer
kubectl exec -n crs <redis-pod> -- redis-cli XPENDING fuzzer_build_queue build_bot_consumers - + 10

# Task registry size
kubectl exec -n crs <redis-pod> -- redis-cli HLEN tasks_registry

# Task state counts
kubectl exec -n crs <redis-pod> -- redis-cli SCARD cancelled_tasks
kubectl exec -n crs <redis-pod> -- redis-cli SCARD succeeded_tasks
kubectl exec -n crs <redis-pod> -- redis-cli SCARD errored_tasks

Consumer groups: build_bot_consumers, orchestrator_group, patcher_group, index_group, tracer_bot_group.

Health Checks

Pods write timestamps to /tmp/health_check_alive. The liveness probe checks file freshness.

# Check health file freshness
kubectl exec -n crs <pod> -- stat /tmp/health_check_alive
kubectl exec -n crs <pod> -- cat /tmp/health_check_alive

If a pod is restart-looping, the health check file is likely going stale because the main process is blocked (e.g. waiting on Redis, stuck on I/O).

Telemetry (OpenTelemetry / Signoz)

All services export traces and metrics via OpenTelemetry. If Signoz is deployed (global.signoz.deployed: true), use its UI for distributed tracing across services.

# Check if OTEL is configured
kubectl exec -n crs <pod> -- env | grep OTEL

# Verify Signoz pods are running (if deployed)
kubectl get pods -n platform -l app.kubernetes.io/name=signoz

Traces are especially useful for diagnosing slow task processing, identifying which service in a pipeline is the bottleneck, and correlating events across the scheduler -> build-bot -> fuzzer-bot chain.

Volume and Storage

# PVC status
kubectl get pvc -n crs

# Check if corpus tmpfs is mounted, its size, and backing type
kubectl exec -n crs <pod> -- mount | grep corpus_tmpfs
kubectl exec -n crs <pod> -- df -h /corpus_tmpfs 2>/dev/null

# Check if CORPUS_TMPFS_PATH is set
kubectl exec -n crs <pod> -- env | grep CORPUS

# Full disk layout - what's on real disk vs tmpfs
kubectl exec -n crs <pod> -- df -h

CORPUS_TMPFS_PATH is set when global.volumes.corpusTmpfs.enabled: true. This affects fuzzer-bot, coverage-bot, seed-gen, and merger-bot.

Deployment Config Verification

When behavior doesn't match expectations, verify Helm values actually took effect:

# Check a pod's actual resource limits
kubectl get pod -n crs <pod-name> -o jsonpath='{.spec.containers[0].resources}'

# Check a pod's actual volume definitions
kubectl get pod -n crs <pod-name> -o jsonpath='{.spec.volumes}'

Helm values template typos (e.g. wrong key names) silently fall back to chart defaults. If deployed resources don't match the values template, check for key name mismatches.

Service-Specific Debugging

For detailed per-service symptoms, root causes, and fixes, see references/failure-patterns.md.

Quick reference:

• DinD: kubectl logs -n crs -l app=dind --tail=100 -- look for docker daemon crashes, storage driver errors
• Build-bot: check build queue depth, DinD connectivity, OOM during compilation
• Fuzzer-bot: corpus disk usage, CPU throttling, crash queue backlog
• Patcher: LiteLLM connectivity, LLM timeout, patch queue depth
• Scheduler: the central brain -- kubectl logs -n crs -l app=scheduler --tail=-1 --prefix | grep "WAIT_PATCH_PASS\|ERROR\|SUBMIT"

Diagnostic Script

Run the automated triage snapshot:

bash {baseDir}/scripts/diagnose.sh

Pass --full to also dump recent logs from all pods:

bash {baseDir}/scripts/diagnose.sh --full

This collects pod status, events, resource usage, Redis health, and queue depths in one pass.

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.