poeely/Deathstar_RCA

DeathStarBench hotel reservation RCA (hotel_t_500_rca_stable_full)

Noisy-neighbor root-cause traces from the DeathStarBench hotel reservation application, running as OpenFaaS functions on Kubernetes. Each fault is a stress-ng Job pinned to the same node as a chosen victim. The archive records the ground-truth intervention, Jaeger traces, and host, node, pod, and per-container performance counters for every window.

The run label is hotel_t_500_rca_stable_full. It started at 2026-09-26T09:17:02Z and ended at 2026-09-27T04:44:50Z (summary.json).

Workload

Piece	Value
Application	DeathStarBench hotel reservation (openfaas-fn)
Load generator	Locust, 1000 users, one process
Request mix	search_hotel, recommend, reserve, user_login
Cluster	libvirt guests on hypervisor dtn-01 (AMD EPYC 7301, 2 sockets, 64 threads, 8 NUMA nodes)
Kubernetes nodes	dtn.controller-0 (8 vCPU) and workers dtn.node-0, dtn.node-1, dtn.node-2 (4 vCPU, 8 GiB each)
Victim node	dtn.node-1, tainted infrabench.io/victim=true:NoSchedule
Stress image	alexeiled/stress-ng:latest

Locust stays at 1000 users for the whole run. The aggregate client log is in locust/. In this run, POST /reservation failed for nearly every request (2,953,659 failures out of 2,953,661). Search, recommend, and user calls completed with a much lower failure rate. Trace latency for reservation spans reflects that.

How a trial is injected

Stress is a co-located noisy neighbor. The orchestrator moves the current victim onto dtn.node-1 and leaves the other hotel workloads on dtn.node-0 and dtn.node-2. It then starts a Kubernetes Job whose pod is pinned to the victim's node (spec.nodeName plus podAffinity on a unique hotel-rca-victim label). The Job runs stress-ng with the arguments below. A satellite sidecar on the stress pod and on the service pods records cgroup performance counters.

A stress window lasts 180 seconds. After the Job is deleted, a settle window of 90 seconds is recorded so PromQL rate() smear and leftover stress-ng processes stay labeled settle. Metrics are scraped every 15 seconds with a 30 second rate window.

Before each new victim placement, a healthy pin window of 180 seconds is collected on that layout with no stressor. The run also has an opening training window, a healthy baseline, and a healthy end window, each 180 seconds. Warmup before collection is 120 seconds.

Stressors

Each placement is crossed with all ten stressors. Hardware stressors target the CPU, cache, memory, disk, or network. Software stressors target process creation, scheduling, pipes, sockets, or timers.

Name	Class	stress-ng arguments	CPU / memory limit	Intended effect
cpu	hardware	--cpu 4 --cpu-load 100 --cpu-method all	4 CPU, 256Mi	CPU saturation, run-queue growth, IPC drop
cache	hardware	--cache 2 --cache-level 3	2 CPU, 512Mi	LLC misses, memory-bound stalls, IPC drop
vm	hardware	--vm 2 --vm-bytes 256M --vm-method all	1 CPU, 768Mi	Memory pressure, LLC misses, DRAM bandwidth
io	hardware	--hdd 2 --hdd-bytes 256M	2 CPU, 512Mi	Disk saturation, filesystem latency
net	hardware	--udp 4	1 CPU, 256Mi	UDP flood, softirq and socket pressure
fork	software	--fork 4	1 CPU, 256Mi	Process creation, scheduler delay
context	software	--switch 4	1 CPU, 256Mi	Context switches, scheduler delay
pipe	software	--pipe 4	1 CPU, 256Mi	Pipe IPC, syscall and context-switch rate
sock	software	--sock 4	1 CPU, 256Mi	Socket IPC, socket and syscall pressure
timer	software	--timer 4	1 CPU, 256Mi	Timer interrupts, context switches

Who is stressed

There are 14 single victims and 6 related pairs. Each is stressed once with every stressor (140 single trials and 60 pair trials).

Single victims: frontend, srv-search, srv-geo, srv-profile, srv-rate, srv-recommendation, srv-reservation, srv-user, mongodb-profile, mongodb-rate, mongodb-reservation, memcached-profile, memcached-rate, memcached-reserve.

Related pairs (the same stressor on both members): memcached-profile+srv-profile, memcached-rate+srv-rate, memcached-reserve+srv-reservation, mongodb-profile+srv-profile, mongodb-rate+srv-rate, mongodb-reservation+srv-reservation.

Other hotel pods (review, attractions, geo mongo, recommendation mongo, and so on) stay up and are scraped, and they are not injection targets in this matrix.

What is in the archive

hotel_t_500_rca_stable_full/Archive.tar.gz (3.58 GB) unpacks to the run directory itself:

tar -xzf Archive.tar.gz

summary.json records 203 trials and 423 progress steps. On disk that is 423 window directories:

Kind	Directories	Role
stress	200	stress-ng is running (180s)
settle	200	Quiet interval after the Job is deleted (90s)
healthy pin	20	No stress, one per victim layout (180s)
healthy bookend	2	0002_healthy-baseline and 0203_healthy-end
training	1	0001_training-traces, healthy traces before the matrix

One settle directory is incomplete: windows/28s_settle-after-net-memcached-profile-srv-profile/ has metrics, fault.json, and logs, and it has no window.json or *_traces.csv. The other 422 directories have window.json.

Directory names look like 0005_cpu-frontend (stress), 05s_settle-after-cpu-frontend (settle), and 0003h_healthy-pin-frontend (healthy pin). Pair names join the two services, for example 0025_cpu-memcached-profile-srv-profile.

Files in a window

File	Contents
window.json	Window label, start and end, and the fault object (kind, trial id, load, placement, interventions)
fault.json	Ground-truth record for this window: kind is stress, settle, or healthy
intervention.json	Stress windows only. Victim pod and node, stress pod and node, stress-ng args, limits, image, and a replay object
stress-job.json	Stress windows only. The Job manifest that was applied
placement.json	OpenFaaS pods by node after the victim was pinned
*_node_metrics.csv	Per Kubernetes node: cpu_utilization_pct, cpu_usage_cores, memory_utilization_pct, memory_usage_bytes
*_pod_metrics.csv	Per pod: cpu_cores, memory_bytes, fs_reads_bytes_per_s, fs_writes_bytes_per_s, including the satellite containers
*_satellite_metrics.csv	Cgroup perf series from the satellite sidecars. Semantic columns include cpu_utilization, IPC, LLC_miss_rate, LLC_MPKI, memory_bound, DRAM_read_BW, context_switches, scheduler_delay, run_queue_length, cache miss rates, and related counters. The column name encodes function, pod, role, metric, stressor, and target
*_host_imc_metrics.csv	Host DRAM traffic from the AMD data-fabric IMC exporter (IMC_total per NUMA node, 64 bytes per count)
*_traces.csv	One row per Jaeger trace (about 18,000 in a 180s window). Columns are span names plus startTime and total duration
*_traces.json	The same traces as raw Jaeger JSON
logs/	Container logs for the hotel pods and Jaeger during the window

index in the metric CSVs is the scrape step (15 seconds).

Files next to windows/

Path	Contents
summary.json	Label, start, end, trial count
experiment.yaml	Matrix that was executed: targets, stressors, durations, Locust settings
experiment.expanded.yaml	Same matrix with the generated trial list expanded
parameters.json, parameters.csv	One-row hardware and schedule record (guest sizes, NUMA, Locust, stressor list, expected window counts)
topology.json	Cluster id, hypervisor, and Kubernetes node capacity
cluster-settings.yaml	Host NUMA, PMU, and satellite settings snapshot
faults.jsonl	One JSON object per completed window, same ground truth as fault.json
gate.json	Preflight result (ok: true for this run)
preflight.log, preflight.xml	Ansible preflight output
training-set.csv	Span-duration table from 0001_training-traces
training/*.csv	One span-duration CSV per victim layout, copied from that layout's healthy-pin trace table (20 files, 18,000 traces each)
locust/	Locust stats, stats_history, failures, and exceptions CSVs
locust_orchestrator.log	Locust process log for the run
data/	Empty

Reading the ground truth

Use fault.kind and fault.trial_id in window.json, or the matching line in faults.jsonl.

• kind: "stress" and active: true means stress-ng was running. interventions[] names the stressor, target_service, arguments, victim pod, and stress pod. combination_scope is single or related_pair.
• kind: "settle" is the 90 seconds after that stress Job. The trial id is settle_after_<stressor>__<target>.
• kind: "healthy" has an empty interventions list. combination_scope is healthy or pin_healthy.

To replay a stress trial, use intervention.json replay: the same cluster, victim_node, stress_node, image, args, and limits. The applied manifest is stress-job.json.