tsdb-diagnosis

name	tsdb-diagnosis
description	Diagnose training job incidents and check cluster health using the per-job Prometheus TSDB. Use when the user asks to diagnose a failure root cause, check GPU/network health, query Prometheus metrics, investigate a hang, or when the triage skill recommends deeper TSDB analysis.

TSDB Diagnosis

Query the per-job Prometheus TSDB to diagnose incident root causes and assess cluster health. Works on any RAY=1 job — finished or running.

Relationship to other skills:

• job-log-triage identifies what failed (from logs). This skill identifies why using time-series metrics.
• performance-analysis identifies why training is slow (from xplane/HLO traces). This skill identifies why metrics are anomalous at the system level.
• The triage skill's next-step templates often say "Query Prometheus TSDB" — this skill automates that.

Prerequisites

This skill requires a Prometheus TSDB, which is only available for RAY=1 jobs. If the job was launched with RAY=0, no TSDB exists — report this and suggest re-running with RAY=1.

Remote Execution via Ray Jobs API

For RAY=1 live jobs, the Ray Jobs API (http://localhost:8265/api/jobs/ on the head node) is a universal remote execution mechanism. It runs arbitrary Python inside the job's Docker containers — no Slurm CLI required.

Security note (two-part fix): Two separate Ray attack surfaces are closed by default in this repo:

1. Dashboard / job-submission HTTP (port 8265) — Ray's JobSubmissionClient HTTP API is unauthenticated; on a host with a public IP, binding to 0.0.0.0 lets an internet scanner submit code that runs as root inside the training container. Fix: bind dashboard, Prometheus, TensorBoard to 127.0.0.1 on the head node; legitimate diagnostics SSH into the head and curl http://localhost:<port>/... from there.
2. GCS / coordination ports (RAY_PORT for Ray, JAX_COORDINATOR_PORT for JAX) — these are unauthenticated cross-node services; on a public IP, internet scanners can join the cluster as raylets within seconds of ray start --head and execute code in the actor pool. Fix: --node-ip-address=<cluster-private-IP> on ray start (head + workers) and the JAX coordinator pinned to the cluster's RFC1918 subnet via detect_cluster_ip() (in utils/detect_ip.sh). Workers reach the head over the private subnet only; internet scanners can no longer connect.

Both endpoints — dashboard and GCS / JAX coordinator — are off the public internet.

Why this matters: In many environments, the diagnosis machine has no Slurm CLI. The Ray Jobs API is still the only way to reach inside the running containers — it just needs an SSH hop now. The agent uses host_cmd.py (or direct ssh) to reach the head node.

Safety boundary — what you can and cannot do:

Safe (decoupled from training)	Dangerous (touches training)
Read files, copy files, inspect sysfs	Kill any process in the training process tree
ps, dmesg, uname, system inspection	pkill python3 or broad pattern kills
Kill/restart the metrics exporter (watchdog auto-restarts)	Send signals to JAX/XLA processes
Overwrite plugin scripts (picked up on next poll cycle)	Modify training config or data files
Overwrite metrics_exporter.sh + kill exporter	Write to GPU sysfs control files
Query Prometheus, read logs	Anything that could cause a training process to exit

Critical rule: Never kill a training process. In a distributed job, killing one trainer on one node causes all other nodes to hang waiting for the dead peer, eventually crashing the entire job (heartbeat timeout or NCCL timeout across all N nodes). The observability stack (metrics exporter, plugins, Prometheus) is fully decoupled — you can tear it down and rebuild it without any training impact.

Submitting a job (run from the head node — ssh in first):

ssh <head_host> "curl -s -X POST 'http://localhost:8265/api/jobs/' \
  -H 'Content-Type: application/json' \
  -d '{\"entrypoint\": \"python3 -c \\\"<python_code>\\\"\", \"runtime_env\": {}}'"

Polling for results (also from the head node):

ssh <head_host> "curl -s 'http://localhost:8265/api/jobs/<job_id>'"
ssh <head_host> "curl -s 'http://localhost:8265/api/jobs/<job_id>/logs'"

If you've already opened the standard ssh -L 8265:localhost:8265 -J <login> <head> tunnel for the dashboard UI, you can also curl http://localhost:8265/... directly from your laptop without re-SSHing for each call.

Node targeting: By default, Ray jobs run on the head node. To inspect a specific worker node, the Ray job's Python code must explicitly discover and connect to that node's resources, or you can use node-local information (like hostname) to verify which node the job landed on. For operations that must run on a specific node (e.g., reading that node's dmesg), you may need to submit multiple jobs and check which node each lands on, or use Ray's scheduling hints.

This API is referenced throughout the skill — in "Metrics Exporter Operations & Live Hotfix" for deploying fixes, in Playbook 4 for GPU driver fault diagnosis, and in the troubleshooting section for process inspection.

Workflow

0. Triage first. Before querying the TSDB, run the job-log-triage skill on the job (or confirm triage has already been done earlier in the conversation). Triage establishes critical context that shapes the entire diagnosis:

   • Job state — running, completed, failed, cancelled, or crashed (determines how to connect to Prometheus in step 2)
   • Fresh start vs checkpoint restore — affects expected baseline (restored jobs may leak RCCL resources; see "Checkpointing interference")
   • Failure class — hang, heartbeat-timeout, OOM, etc. (determines which playbook to run in step 5)
   • Step range and timing — which steps are comparable, where anomalies occurred
   • Config parameters — PASSTHROUGH_ARGS, NNODES, checkpoint_period, etc.

   Without triage, you risk misinterpreting metrics (e.g., querying post-hang idle metrics as if they were training metrics, or missing that a checkpoint restore caused resource leaks). For proactive health checks on a live job where no failure has occurred, triage is still useful to confirm the job is running and extract the head node hostname/port.

1. Locate the TSDB. Resolve the job directory (same rules as triage: given a log file, job dir, Slurm ID, or k8s job ID). For k8s jobs, the primary log (outputs/<id>-<name>.log) contains rank 0's output; per-rank logs are at outputs/<id>-<name>/rank-N.log. Verify <job_dir>/prometheus/ exists and contains data — it should have subdirectories with ULID names (e.g., 01KHV6MFN61MJKZ3ZSYYRYGDGX) and/or a wal/ directory. An empty or near-empty prometheus/ directory means Prometheus failed to start or never scraped — check the observability stack logs (see "Troubleshooting Missing or Incomplete TSDB Data"), report what went wrong, and skip to log-only diagnosis.

2. Connect to Prometheus. First determine whether the job is still running, then choose the appropriate access method.

   Determine job state (from triage, or verify directly): The triage report from step 0 tells you whether the job is running or finished. If you need to verify independently:

   • squeue -j <id> -h -o %T 2>/dev/null — if it returns RUNNING, the job is live → Case A.
   • If squeue is unavailable or the job isn't found, check the log: a block means the job is finished → . No summary and no advancing steps → likely crashed → also (the live Prometheus is gone).

Querying Prometheus

All queries use the Prometheus HTTP API via curl. Parse responses with inline Python.

Instant query (single point in time)

curl -s 'http://localhost:<port>/api/v1/query?query=<promql>&time=<unix_ts>'

Always specify &time=<unix_ts> for read-only instances (Case B/C). Without it, Prometheus defaults to "now" — which is past the end of the data for finished jobs, returning empty results. Use a timestamp from within the job's time range (discovered in step 3).

Range query (time series)

curl -s 'http://localhost:<port>/api/v1/query_range?query=<promql>&start=<unix_ts>&end=<unix_ts>&step=<interval>'

Use step=30s for high resolution (short windows), step=60s for medium, step=300s for long durations.

URL encoding for rate() and increase()

When using rate() or increase() in curl URLs, the square brackets must be URL-encoded:

# WRONG — brackets break the URL:
curl -s 'http://localhost:9190/api/v1/query?query=rate(hw_tcp_retransmits_total[5m])'

# CORRECT — URL-encode the brackets:
curl -s 'http://localhost:9190/api/v1/query?query=rate(hw_tcp_retransmits_total%5B5m%5D)'

[ = %5B, ] = %5D. This applies to all PromQL queries containing range selectors.

Standard query + tabular output pattern

Use this pattern for all queries — it extracts the metric labels and values into a readable table:

curl -s 'http://localhost:<port>/api/v1/query?query=<promql>&time=<unix_ts>' | python3 -c "
import json, sys
data = json.load(sys.stdin)
for r in data.get('data', {}).get('result', []):
    labels = r['metric']
    host = labels.get('host', labels.get('instance', '?'))
    gpu = labels.get('gpu', '')
    val = r['value'][1]
    print(f'  {host:<20} gpu={gpu:<4} {val}')
"

For range queries, use this to extract per-host min/max/avg:

curl -s 'http://localhost:<port>/api/v1/query_range?query=<promql>&start=<start>&end=<end>&step=30s' | python3 -c "
import json, sys
data = json.load(sys.stdin)
for r in data.get('data', {}).get('result', []):
    labels = r['metric']
    host = labels.get('host', labels.get('instance', '?'))
    gpu = labels.get('gpu', '')
    vals = [float(v[1]) for v in r['values'] if v[1] != 'NaN']
    if vals:
        print(f'  {host:<20} gpu={gpu:<4} min={min(vals):.1f}  max={max(vals):.1f}  avg={sum(vals)/len(vals):.1f}')
"

Filtering by host

Narrow queries to specific hosts (e.g., the accused nodes from triage):

hw_gpu_power_watts{host=~"node001|node002"}
ray_node_gpus_utilization{host="node003"}

Rate queries for counters

Counters (names ending in _total) must use rate() or increase():

rate(hw_tcp_retransmits_total[5m])
increase(hw_oom_kills_total[1h])

Ray Worker Logs

Ray worker logs are the second source of truth alongside TSDB metrics. They contain NCCL/RCCL initialization messages, XLA compilation events, Python tracebacks, and framework-level warnings that the TSDB cannot capture. Use them to confirm hypotheses formed from metric analysis.

Location: <job_dir>/ray_logs/<hostname>/worker-*-<pid>-<pid>.out (stdout) and .err (stderr). Each host has a subdirectory; each Ray worker has a pair of .out/.err files.

Common patterns to search for:

• init.cc:2095 NCCL WARN MSCCL++ — NCCL communicator initialization. Count the number of "waves" (clusters of these messages separated by time gaps) to detect unexpected communicator creation (e.g., extra wave from single-replica restore broadcast).
• Compilation of or Compiled — XLA compilation events. Unexpected recompilations mid-training indicate shape changes or cache misses.
• NCCL WARN or RCCL WARN — collective communication warnings.
• Python tracebacks — stack traces from exceptions (may be caught and logged without crashing).
• SingleReplicaArrayHandler or broadcast_one_replica_to_all — checkpoint restore broadcast messages (relevant to RCCL leak diagnosis).

Practical tips:

• Worker logs can be very large. Use grep to search, don't read entire files.
• The head node (task 0) typically has the most informative logs — start there.
• When comparing two jobs, grep for the same pattern in both and diff the output (e.g., count NCCL init waves in each).

Troubleshooting Missing or Incomplete TSDB Data

When the TSDB is empty, has gaps, or is missing metric families, the problem is in the observability stack itself. The stack has three components, each with its own persistent log:

Component	Log location	What it does
Prometheus	<job_dir>/prometheus/prometheus.log	Scrapes targets, stores time series
Metrics exporter	<job_dir>/metrics_exporter/<hostname>.log	Runs plugins (gpu_metrics_plugin.sh, host_metrics_plugin.sh, tb_metrics_plugin.sh), serves metrics on port 9400
Ray runtime	<job_dir>/ray_logs/<hostname>/ (gcs_server.out, raylet.out, dashboard.log, etc.)	Cluster formation, built-in metrics exporter on port 55080

All three are persisted in the job directory and survive job termination. The watchdog (_run_with_watchdog) restarts Prometheus and the metrics exporter on crash, logging restart events with timestamps.

Diagnosis by symptom:

Symptom	Which log to check	What to look for
TSDB directory empty or missing	Job log (main stdout)	[Prometheus] messages — did it start? Did install_prometheus fail? Port bind failures?
TSDB has data but with time gaps	prometheus/prometheus.log	Crash messages followed by watchdog restarts (Watchdog: exited with code ... restart #N). Each restart replays the WAL — gap duration ≈ crash-to-restart time (~5s)
No hw_* metrics (but ray_* and tb_* present)	metrics_exporter/<hostname>.log	Plugin errors — gpu_metrics_plugin.sh or host_metrics_plugin.sh failing. Exit code 99 means permanently skipped
No tb_* metrics	metrics_exporter/<hostname>.log	tb_metrics_plugin.sh failing — usually because TensorBoard event files haven't been written yet (normal during compilation), or the event file path is wrong
tb_* metrics have gaps (missing steps)	prometheus/prometheus.log + metrics_exporter/<hostname>.log	The tb_metrics_plugin.sh bridge is best-effort — gaps can occur when the GPU metrics plugin hangs (blocking the exporter), the exporter crashes and restarts, or Prometheus rejects samples whose wall_time is older than its minValidTime after TSDB compaction. Check Prometheus log for "samples that are too old" warnings and exporter log for plugin timeouts. When tb_* data has gaps, use the raw TensorBoard event file as ground truth (see "Recovering from tb_* gaps" below)
No ray_* metrics	ray_logs/<hostname>/ (GCS, raylet, dashboard)	Ray didn't start or its metrics exporter on port 55080 failed. Check gcs_server.out and raylet.out for errors
Some hosts missing entirely	prometheus/prometheus.log	Scrape errors for those hosts — or . Means Prometheus couldn't reach the exporter on those nodes (node down, network issue, or exporter crashed before Prometheus could scrape)

Checking the logs:

# Prometheus crash/restart history
grep -i 'watchdog\|error\|fatal\|panic\|restart' <job_dir>/prometheus/prometheus.log

# Metrics exporter plugin failures (all hosts at once)
grep -i 'error\|skip\|fail\|exit' <job_dir>/metrics_exporter/*.log

# Ray runtime issues on head node
head_host=$(ls <job_dir>/ray_logs/ | head -1)
grep -i 'error\|fail\|exception' <job_dir>/ray_logs/$head_host/gcs_server.out

Important: These logs explain why the TSDB is broken — they don't replace the TSDB for diagnosing the training job. Once you've identified and understood the observability gap, report it alongside whatever partial TSDB data is available, and note which time ranges or metric families should not be trusted.

Recovering from tb_* gaps

The tb_metrics_plugin.sh bridge is best-effort. It can lose data when:

• A plugin (e.g., gpu_metrics_plugin.sh) hangs in a D-state, blocking the exporter cycle
• The exporter crashes and the watchdog restarts it, but Prometheus has advanced its minValidTime past the pending steps
• The plugin drops stale steps per Rule 4 (freshness filtering) to avoid Prometheus rejection

The raw TensorBoard event file is always the ground truth. It is written directly by the training process and is never subject to Prometheus's timeline constraints. When tb_* metrics have gaps in the TSDB:

1. Locate the event file: <job_dir>/tensorboard/events.out.tfevents.* (or check the --tensorboard_dir flag in the job config).
2. Read it programmatically (Python):

   from tensorboard.backend.event_processing.event_accumulator import EventAccumulator
   ea = EventAccumulator("<path_to_event_dir>")
   ea.Reload()
   for tag in ea.Tags()["scalars"]:
       for e in ea.Scalars(tag):
           print(f"step={e.step}  wall_time={e.wall_time}  tag={tag}  value={e.value}")
   

3. Cross-reference with TSDB: Compare the steps present in the event file against what Prometheus has. The event file will have every step; the TSDB may be missing ranges.

Trade-off: Raw event files require programmatic access and cannot be queried with PromQL or visualized in Grafana dashboards. For one-off investigations this is fine. For systematic dashboarding, the Prometheus tb_* metrics remain the primary source — gaps are expected to be rare with the current exporter hardening (plugin timeouts, freshness filtering, mid-training restart fallback).

Metrics Exporter Operations & Live Hotfix

The metrics exporter runs inside a Docker container on each node, managed by a watchdog that restarts it on crash. During a live job, you may need to deploy fixes to plugins or to the exporter itself without restarting the job. The two cases are fundamentally different.

Architecture recap:

watchdog (_run_with_watchdog)
  └─ metrics_exporter.sh  (long-running; poll loop every ~10s)
       ├─ gpu_metrics_plugin.sh   (re-sourced each cycle)
       ├─ host_metrics_plugin.sh  (re-sourced each cycle)
       ├─ tb_metrics_plugin.sh    (re-sourced each cycle)
       └─ HTTP server (serves cached metrics file on port 9400)

Case 1 — Plugin hotfix (gpu_metrics_plugin.sh, host_metrics_plugin.sh, tb_metrics_plugin.sh):

Plugins are re-read from disk on every poll cycle (~10s). To deploy a fix:

1. Overwrite the plugin file at the artifact path inside the container.
2. The next poll cycle picks up the new code automatically. No process restart needed.

Network note: The Ray dashboard and the per-job Prometheus query API both bind to 127.0.0.1 only on the head node (user-facing HTTP hardening). All curl http://localhost:8265/... and curl http://localhost:9190/... examples below assume you've already SSH'd to the head node (e.g., ssh <head_host> or host_cmd.py "ssh <head_host> ..."). The localhost in the URL refers to the head node's loopback, not your laptop. Inter-node Ray (worker → head GCS) and the JAX coordinator bind to the cluster-private IP (RFC1918) via detect_cluster_ip() — workers reach the head over the private subnet only, never the public internet. Prometheus scraping (head → worker exporters on :9400) is intentionally cross-node-reachable on the private subnet so Prometheus can pull metrics from each node.

File delivery caveat: The plugin files live at <artifact_dir>/utils/ inside the container. The artifact directory is bind-mounted from the host NFS, but NFS attribute caching can delay visibility of changes by 30-60s or indefinitely. To guarantee immediate delivery, use the Ray Jobs API to copy the file inside the container:

# Run on the head node (ssh there first).  Use a shell command (not python -c)
# so we avoid nested-quote escaping inside the JSON entrypoint string.
curl -s -X POST 'http://localhost:8265/api/jobs/' \
  -H 'Content-Type: application/json' \
  -d '{
    "entrypoint": "cp -p /path/on/nfs/plugin.sh /artifact_dir/utils/plugin.sh && echo done",
    "runtime_env": {}
  }'

The Ray job runs inside the container, bypassing NFS attribute cache issues. If you do need Python (e.g., for a complex transformation), put the script on shared storage and call python3 /path/on/nfs/script.py rather than embedding python3 -c "..." in the JSON.

Case 2 — Exporter hotfix (metrics_exporter.sh):

The exporter is a long-running process — overwriting the file alone does nothing until the process restarts. To deploy:

1. Deliver the updated metrics_exporter.sh to <artifact_dir>/utils/ using the Ray Jobs API (same as above).
2. Kill the running exporter process. The watchdog detects the exit and restarts it with the updated code.

# Run on the head node (ssh there first).
# Kill the exporter (via Ray Jobs API, targeting a specific node):
curl -s -X POST 'http://localhost:8265/api/jobs/' \
  -H 'Content-Type: application/json' \
  -d '{
    "entrypoint": "python3 -c \"import subprocess, os; subprocess.run([\\\"pkill\\\", \\\"-f\\\", \\\"metrics_exporter.sh\\\"])\"",
    "runtime_env": {}
  }'

Pitfalls:

• Port conflict on restart: When the exporter is killed, its HTTP server child process may linger briefly, holding port 9400. The watchdog restarts the exporter immediately, and the new instance gets OSError: [Errno 98] Address already in use. The watchdog will retry on the next crash cycle. If this persists, kill the orphaned HTTP server process explicitly before restarting.
• Self-termination of Ray jobs: When using pkill -f <pattern> inside a Ray job, the Ray job's own entrypoint command string may match the pattern, causing the job to kill itself. Use narrow patterns or filter out the current PID: pkill -f "metrics_exporter" --signal TERM with a pattern that won't match the Ray entrypoint string.
• Multi-node deployment: Ray jobs run on the head node by default. To target a specific worker node, you need to submit a job that explicitly connects to that node's Ray worker, or submit separate jobs per node. For bulk deployment across all nodes, iterate over the host list.
• Watchdog restart logging: Each restart is logged with a timestamp and restart count in <job_dir>/metrics_exporter/<hostname>.log (e.g., [Metrics Exporter] Watchdog: exited with code 137, restart #3). Monitor this log to confirm the restart succeeded.

Case 3 — Remote process inspection and management:

For any ad-hoc operation inside a container (checking process state, reading dmesg, killing specific processes), use the Ray Jobs API (see "Remote Execution via Ray Jobs API" section above). Common diagnostic commands:

# All commands below run on the head node (ssh there first).

# Check for D-state processes on the cluster
curl -s -X POST 'http://localhost:8265/api/jobs/' \
  -H 'Content-Type: application/json' \
  -d '{
    "entrypoint": "python3 -c \"import subprocess; r=subprocess.run([\\\"ps\\\",\\\"axo\\\",\\\"pid,stat,wchan:30,cmd\\\"], capture_output=True, text=True); d=[l for l in r.stdout.splitlines() if l.split()[1:2] and l.split()[1][0]==\\\"D\\\"]; print(f\\\"D-state: {len(d)}\\\"); [print(l.strip()) for l in d[:20]]\"",
    "runtime_env": {}
  }'

# Read GPU-related dmesg errors
curl -s -X POST 'http://localhost:8265/api/jobs/' \
  -H 'Content-Type: application/json' \
  -d '{
    "entrypoint": "python3 -c \"import subprocess; r=subprocess.run([\\\"dmesg\\\",\\\"--level=err,warn\\\",\\\"-T\\\"], capture_output=True, text=True); lines=[l for l in r.stdout.splitlines() if \\\"amdgpu\\\" in l.lower() or \\\"drm\\\" in l.lower() or \\\"cper\\\" in l.lower()]; print(f\\\"GPU dmesg lines: {len(lines)}\\\"); [print(l) for l in lines[-30:]]\"",
    "runtime_env": {}
  }'

# Get kernel and driver version
curl -s -X POST 'http://localhost:8265/api/jobs/' \
  -H 'Content-Type: application/json' \
  -d '{
    "entrypoint": "python3 -c \"import subprocess; print(subprocess.run([\\\"uname\\\",\\\"-r\\\"], capture_output=True, text=True).stdout.strip()); print(open(\\\"/sys/module/amdgpu/version\\\").read().strip())\"",
    "runtime_env": {}
  }'

D-state Process Accumulation from Plugin Timeouts

When a plugin (typically gpu_metrics_plugin.sh) enters kernel D-state (uninterruptible sleep), the exporter's timeout mechanism kills the plugin's bash wrapper process with SIGKILL. However, the actual child process stuck in D-state cannot be killed by any signal — it remains as an unkillable zombie consuming a PID and kernel task_struct memory.

Over time, these accumulate. At ~6 per hour (one per timeout cycle that hits D-state), a node can accumulate hundreds. While Linux's default PID max (4,194,304) provides ample headroom, a large accumulation indicates a persistent kernel-level fault.

Detection:

# Intermittent hw_scrape_duration_seconds alternating between ~12s (timeout)
# and <3s (success) indicates the plugin hangs intermittently.
# Full timeout (all cycles >=12s) means the plugin hangs every time.
hw_scrape_duration_seconds{host="<suspect_host>"}

Confirmation via Ray Jobs API (see "Remote Execution via Ray Jobs API" and "Case 3 — Remote process inspection" for full curl examples):

# Submit a Ray job to count D-state processes and show their kernel wait channel:
# amdgpu_cper_ring_write = GPU driver CPER ring bug
# amdgpu_ras_* = RAS sysfs read hang
# Use: ps axo pid,stat,wchan:30,cmd | filter for D-state
# Also submit: uname -r and cat /sys/module/amdgpu/version for the admin report

Resolution: D-state processes cannot be cleared without a node reboot. They are stuck in a kernel code path that will never return. Schedule a reboot at the next job boundary or maintenance window.

Diagnostic Playbooks

Each playbook is triggered by a specific scenario (from triage output or user request). Run the listed queries, apply the interpretation rules, and report findings.

Correlation is not causation

When an incident occurs, many metrics move at once. A single underlying event — a network blip, a thermal excursion, an I/O stall — ripples through the system and causes correlated changes across GPU power, clocks, utilization, throughput, step time, and more. Most of what you see in the TSDB are symptoms, not root causes. Do not report a correlated symptom as the diagnosis.

The diagnostic principle: find the earliest anomaly. The root cause is the metric that deviates first. Everything that follows is a consequence. When you spot any anomaly:

1. Note the timestamp of the anomaly.
2. Query all other metric families at the same timestamp and in the minutes preceding it — both training metrics (tb_*) and system metrics (hw_*, ray_*).
3. Find which metric deviated first. That is the root cause candidate. Everything that moved after it is a correlated effect.
4. Verify the causal direction. The root cause should logically explain the downstream symptoms.

Root causes can be in either domain — training or system. The TSDB contains both training-level metrics (loss, grad norms, LR, MoE load balance loss, throughput) and system-level metrics (GPU power/clocks, network, I/O) on the same timeline. Do not assume the root cause is always a system event. Examples:

• System → training: A network retransmit burst stalls a collective → GPU power drops → step time spikes → throughput drops. System event is the root cause.
• Training → system: An MoE model's routing changes dramatically → tb_learning_moe_lb_loss spikes → some experts are overloaded while others are idle → GPU utilization becomes uneven → TGS drops. The mathematical event (routing shift) is the root cause; the GPU utilization change is a passive symptom.
• Training → system: A grad norm explosion → NaN values → recompilation or fallback → throughput collapse. The training instability is the root cause.
• Training → system: Learning rate warmup ends → larger weight updates → checkpoint sizes grow → I/O pressure during saves → periodic step time spikes. The LR schedule change is the root cause.

Always query both domains. When system metrics (power, clocks, utilization) change, check whether a training metric (tb_learning_*, tb_perf_*) shifted first — the system may be passively responding to a change in the model's computational behavior.

Always report the full chain: root cause → intermediate effects → observed symptoms. This is the core value of having all metrics in a single time-aligned TSDB.

Trace suspicious findings to source code

Metrics tell you what is happening; source code tells you why. When the TSDB reveals an anomaly that can't be explained by external factors (network, hardware, thermal), trace the causal chain into the code to find the mechanism. This is especially valuable when the anomaly is persistent (not transient) and uniform across all nodes (not a single-node hardware issue) — these patterns point to a software-level root cause.

When to investigate code:

• A metric delta between two jobs with identical configs that isn't explained by hardware, network, or thermal differences (e.g., hw_procs_running is uniformly higher in one job).
• A persistent anomaly that starts at a specific event (checkpoint restore, profiler activation, config change) and never recovers.
• The TSDB points to a specific subsystem (checkpointing, data pipeline, collective communication) but the metric alone doesn't explain the mechanism.

How to trace:

1. Start from the triggering event. The triage report and metric timeline tell you when the anomaly starts. Identify what code executed at that point (e.g., checkpoint restore at step N, profiler hook at step M).
2. Follow the call chain. Read the relevant MaxText entry point (e.g., checkpointing.py:load_state_if_possible for restore issues), then trace into the libraries it calls (Orbax, JAX, XLA). Use semantic search and grep to find the code paths.
3. Look for resource creation without cleanup. The most common source of persistent anomalies is resources (communicators, threads, file handles, caches) that are created during a transient operation but never released. Check whether the code path creates anything that outlives the function call.
4. Check for alternative code paths. Libraries often have multiple implementations of the same operation (e.g., dispatcher-based vs legacy path in Orbax). If one path leaks resources, the other may not — this informs the fix direction.
5. Verify with logs. After forming a hypothesis from the code, confirm it in the job's logs (ray worker .out files, stderr). Look for initialization messages, warnings, or resource creation events that match the code path you identified.

Accessible code locations:

• MaxText source: /workspace/maxtext/src/MaxText/ (checkpointing, training loop, configs)
• Orbax (checkpoint library): /opt/venv/lib/python3.12/site-packages/orbax/checkpoint/
• JAX: /opt/venv/lib/python3.12/site-packages/jax/
• XLA C++ headers (for understanding communicator/backend behavior): search under /opt/venv/ or use python3 -c "import jaxlib; print(jaxlib.__path__)" to locate

Example from real diagnosis: TSDB showed hw_procs_running was ~17 higher per host in a restored job (with enable_single_replica_ckpt_restoring=true) vs a fresh-start baseline. Code tracing revealed: MaxText checkpointing.py → Orbax SingleReplicaArrayHandler.deserialize() → multislice.broadcast_one_replica_to_all() → _merge_globalized_replicas() which uses jax.jit(jnp.sum) → XLA all-reduce collective → new RCCL communicators with unique GpuCliqueKey → cached permanently in C++ clique cache → background polling threads increase hw_procs_running. Without code tracing, the TSDB alone could only say "more processes are running" — the code trace identified the exact mechanism and pointed to a fix.

Checkpointing interference

Checkpointing is one of the most disruptive periodic events in training. During a checkpoint save, the system behavior changes dramatically — and many metrics that look anomalous are actually normal checkpoint effects. Always check whether an anomaly coincides with a checkpoint step before diagnosing it as a problem.

What happens during a checkpoint save:

• GPU utilization drops (GPUs idle while waiting for D2H transfer and I/O to complete)
• GPU power drops (idle GPUs draw less power)
• Host memory spikes (model parameters copied from GPU to host RAM, especially on DP replica #0 nodes)
• I/O pressure rises (hw_io_pressure_full_pct, hw_mem_dirty_bytes) as parameters are written to storage
• Step time spikes (the checkpoint step takes much longer than a normal training step)
• hw_procs_blocked increases (processes waiting on I/O)
• Network traffic may drop (no collectives during the checkpoint window)

How to identify checkpoint steps: Parse checkpoint_period from the log (from Config param checkpoint_period: N). Checkpoint saves occur at steps that are multiples of this period. If async_checkpointing=true is in PASSTHROUGH_ARGS, the save happens in the background and may overlap with subsequent training steps, spreading the I/O impact over multiple steps.

Rules:

• When analyzing steady-state performance, exclude steps around checkpoint boundaries (the checkpoint step and 1-2 steps after for async checkpointing).
• When diagnosing an anomaly, always check whether the timestamp falls on a checkpoint step before investigating further. A step time spike at step 200 with checkpoint_period=200 is expected, not an incident.
• When comparing metrics across jobs with different checkpoint_period values, normalize by excluding checkpoint steps from both.
• Memory spikes during checkpoint saves are not OOM precursors unless they push the host to its limit. DP replica #0 nodes use ~2x model size in host RAM during saves — this is expected.

Expected NFS congestion during checkpoint writes:

Checkpoint-writing nodes (one per FSDP replica) write large model parameters to shared storage (VAST/NFS) simultaneously. This routinely causes:

• TCP retransmit rates of 500–3000/s on checkpoint-writing nodes — this is NFS retransmitting during heavy writes and is normal, not alarming. Do not flag TCP retransmits during checkpoint windows as a network issue unless they also appear outside checkpoint windows.
• I/O pressure (hw_io_pressure_full_pct) up to 80% — transient, lasts the duration of the checkpoint write (typically 5–8 minutes).
• 10–20+ blocked processes (hw_procs_blocked) per checkpoint-writing node — processes waiting on NFS I/O.
• Non-checkpoint-writing nodes (remaining FSDP replicas) show near-zero I/O pressure and TCP retransmits during the same window.

Diagnosing checkpoint filesystem errors (checkpoint-fs-error from triage):

When a checkpoint write fails with an NFS-related error (e.g., manifest.ocdbt.__lock ENOENT), compare the I/O metrics at the failed checkpoint against previous successful checkpoints to assess whether NFS conditions were worse:

# For each checkpoint step, query peak I/O pressure and TCP retransmit rate:
for ts in <ckpt_2600_ts> <ckpt_2800_ts>; do
  end=$((ts+1800))
  curl -s "http://localhost:<port>/api/v1/query_range?query=hw_io_pressure_full_pct&start=$ts&end=$end&step=30s"
  curl -s "http://localhost:<port>/api/v1/query_range?query=rate(hw_tcp_retransmits_total%5B5m%5D)&start=$ts&end=$end&step=30s"
done

Compare peak values per node across checkpoints. A significant increase in TCP retransmits or I/O pressure at the failing checkpoint (vs. previous successes) points to NFS congestion as the trigger. Also check hw_procs_blocked — the head node (task 0) is especially vulnerable because it runs additional I/O-intensive services (Prometheus TSDB, Ray GCS, dashboard) alongside the checkpoint writer.

Single-replica checkpoint restore and RCCL communicator leaks:

When enable_single_replica_ckpt_restoring=true and the job restores from a checkpoint, the restore code path leaks RCCL communicators and their background polling threads. This manifests as a persistent increase in hw_procs_running (~17 extra runnable threads per host on a 24-node / 3-replica MoE run) that appears immediately after restore and never recovers. The leaked threads create constant CPU contention — competing with data pipeline threads, RCCL coordination, and the Python runtime — causing a persistent throughput (TGS) drop of ~1%.

Root cause (code-level): During single-replica restore, Orbax's SingleReplicaArrayHandler broadcasts the restored parameters from one replica to all others. The non-dispatcher code path in orbax/.../multislice.py (_merge_globalized_replicas) uses jax.jit(lambda: jnp.sum(axis=0), out_shardings=...) to perform this broadcast. The jnp.sum across the replica axis compiles to an XLA all-reduce collective, which causes RCCL to initialize new communicators via XLA's AcquireGpuClique(). Because the restore uses a different mesh/sharding configuration (single-replica subset mesh) than training, the resulting GpuCliqueKey is unique — so RCCL creates brand-new communicators instead of reusing the training ones. These communicators and their background polling threads are cached permanently in XLA's C++ GPU clique cache (GpuCliqueKey → LockableGpuClique).

Why cleanup attempts fail:

• jax.clear_caches() only clears Python-level JIT compilation caches. It does not touch the C++ communicator cache. Confirmed ineffective by testing.
• jax.clear_backends() would clear communicators but tears down the entire JAX runtime — unusable mid-training.
• There is no Python API to destroy individual RCCL communicators. The C++ clique cache has no eviction mechanism or public release API.
• MaxText's _restore_original_array_handler() re-registers the original ArrayHandler but cannot affect the C++ communicator state.

Log-level confirmation: The ray worker .out logs show an extra NCCL communicator initialization wave during restore. Look for init.cc:2095 NCCL WARN MSCCL++ messages — a fresh-start job has 2 waves (startup + training), while a restored job has 3 waves (startup + restore broadcast + training). The extra wave corresponds to the _merge_globalized_replicas all-reduce creating new communicators.

Verified fix: Two patches on the yihuang/fix-rccl-thread-leak-single-replica-restore branch address this:

1. Replace Orbax's JIT-based broadcast with a direct RCCL broadcast via ctypes that explicitly destroys communicators after use (ncclCommInitRank / ncclBroadcast / ncclCommDestroy + gc.collect()). This eliminates the leaked threads and the ~1% TGS drop.
2. Replace Orbax's jax.jit(create_zeros) on non-primary hosts with numpy.zeros + jax.device_put, eliminating extra XLA compilations during restore that degrade steady-state performance.

How to detect: When comparing a restored job against a fresh-start baseline with identical config:

• Check hw_procs_running — a uniform delta across all hosts (not a spike, not on specific nodes) after restore is the signature.
• The delta is present from the first training step after restore and does not recover.
• Training-level metrics (loss, grad norms, LR) will be identical between the two jobs — this is purely system-level contention.
• A fresh start with the same config but no actual restore (even if enable_single_replica_ckpt_restoring=true) does not trigger the leak — the restore code path must actually execute.
• Check ray worker .out logs for the extra NCCL init wave (3 waves vs 2) as a definitive confirmation.
• If both jobs ran with _env_ENABLE_XLA_DUMP=1, compare the number of compiled modules in xla_dump/ — a restored job may have extra jit_create_zeros entries from Orbax's non-dispatcher path. These extra XLA compilations are a secondary performance penalty beyond the leaked threads.

---

Playbook 1: RCCL/NCCL Hang

Trigger: Triage reports hang or cancelled with underlying hang. User says "why did the job hang" or "confirm the RCCL hang."

Goal: Confirm the hang mechanism (RCCL busy-wait vs. dead node vs. network partition) and identify the trigger.

Queries (at the hang time window — from last successful training step to job end):

#	PromQL	What it shows
1	ray_node_gpus_utilization	Per-host GPU utilization
2	hw_gpu_power_watts	Per-GPU power draw
3	rate(hw_tcp_retransmits_total[5m])	TCP retransmit rate per host
4	rate(hw_rdma_tx_retx_pkts_total[5m])	RDMA retransmit rate per device
5	rate(hw_rdma_tx_ack_timeout_total[5m])	RDMA ACK timeouts per device
6	hw_rdma_port_state	RDMA port state (1=ACTIVE)
7	rate(hw_rdma_rx_cnp_pkts_total[5m])	Congestion notification packets

Interpretation:

GPU util	Power	Network	Diagnosis
High (variable per GPU)	Idle-level (~260–300W MI355X)	Clean	Confirmed RCCL busy-wait hang. GPUs in the RCCL polling loop report high utilization, but power is at idle/standby. Important: utilization may not be uniform — in partial deadlocks, only GPUs that entered the collective show high utilization (e.g., 2 of 8 GPUs at 100%, rest at 0%). Always check power as the ground truth.
~100% all nodes	Idle-level	Retransmit spike before hang	Network-triggered RCCL hang. A transient network event caused a collective to stall, then all nodes entered busy-wait.
~100% most nodes	Mixed (some idle, some active)	Clean	Possible straggler. One or more nodes may have fallen behind, causing others to wait. Check per-node power to identify the slow node.
0% on some nodes	0W on some nodes	N/A	Node death. Some nodes died — remaining nodes hung waiting for them. Check hw_dmesg_gpu_errors_total and hw_gpu_ras_* on the dead nodes.
~100% all nodes	Active-level (~900W MI355X)	Clean	Not a hang — training was still running. Re-check the triage classification.

Key thresholds (MI355X):

• Active training power: ~900W per GPU
• RCCL busy-wait power: ~300W per GPU (only ~42W above idle — nearly indistinguishable from standby)
• Idle/standby power: ~260W per GPU

---

Playbook 2: Heartbeat False-Positive

Trigger: Triage reports heartbeat-timeout. User wants TSDB confirmation that the accused tasks were healthy.

Goal: Prove or disprove that the accused tasks were alive and training when the heartbeat mechanism killed them.

Setup: From the log, extract:

• Crash time (the heartbeat error timestamp)
• Heartbeat timeout value (jax_distributed_heartbeat_timeout_seconds)
• Accused task IDs and their host mappings (from NNODES, NODE_LIST, task-to-host mapping in log)
• Compute heartbeat-stop time: crash_time - heartbeat_timeout

Queries (at the heartbeat-stop time — this is when the heartbeats actually failed):

#	PromQL	What it shows
1	ray_node_gpus_utilization{host=~"<accused_hosts>"}	Were accused tasks training?
2	ray_node_mem_used{host=~"<accused_hosts>"}	Memory on accused hosts
3	hw_io_pressure_full_pct{host=~"<accused_hosts>"}	I/O pressure (checkpoint stall?)
4	rate(hw_tcp_retransmits_total{host=~"<accused_hosts>"}[5m])	Network health
5	hw_tcp_listen_drops_total{host=~"<coordinator_host>"}	gRPC coordinator overloaded?
6	hw_mem_dirty_bytes{host=~"<accused_hosts>"}	Dirty page pressure

Interpretation:

GPU util on accused	Network	I/O pressure	Diagnosis
High (active training)	Clean	Low	Confirmed false positive. Tasks were healthy. The heartbeat mechanism failed (known gRPC bug — see docs/jax-heartbeat-false-positive-postmortem.md).
High	Retransmit spike	Low	Likely false positive triggered by transient network issue blocking the heartbeat gRPC. Tasks were alive but heartbeat RPCs couldn't reach the coordinator.
High	Clean	High	Likely false positive triggered by checkpoint I/O pressure. Heavy writeback blocked the heartbeat thread.
Low/zero	Any	Any	Possibly a true positive. Tasks may have actually died. Check for OOM, GPU errors, or crashes on those hosts.

Report template for confirmed false positive:

The TSDB confirms this is a heartbeat false-positive kill. At the heartbeat-stop time (), all accused hosts showed GPU utilization at %, TCP retransmit rate near zero, and no I/O pressure. The tasks were alive and actively training when the heartbeat mechanism declared them dead. Root cause: shared gRPC channel bug documented in docs/jax-heartbeat-false-positive-postmortem.md.

---

Playbook 3: Host OOM

Trigger: Triage reports oom-host. User wants to understand the memory trajectory.

Goal: Trace the memory growth pattern, identify what caused the OOM, and determine if it's reproducible.

Queries (over the full job duration, or last 30 minutes before the OOM):

#	PromQL	What it shows
1	ray_node_mem_used	Per-host memory over time
2	increase(hw_oom_kills_total[1h])	OOM kill events
3	hw_mem_dirty_bytes	Dirty page accumulation
4	hw_mem_writeback_bytes	Writeback pressure
5	hw_io_pressure_full_pct	I/O stall percentage
6	hw_procs_blocked	Processes blocked on I/O
7	hw_gpu_vram_used_bytes	GPU VRAM (to correlate D2H transfers)

Interpretation:

• Gradual ramp → memory leak (Python objects, data pipeline buffers, or XLA compilation cache growing unbounded).
• Sudden spike → checkpoint save (D2H copies all parameters to host RAM), data loading burst, or XLA recompilation allocating new buffers.
• Periodic spikes that recover → checkpoint saves. If the peak exceeds available RAM on one cycle, it's a checkpoint OOM.
• Only DP replica #0 nodes OOM → checkpoint save pattern. MaxText saves checkpoints from DP replica 0 only, which requires ~2x model size in host RAM (one copy in GPU VRAM, one on host for the save).
• All nodes OOM simultaneously → likely a data pipeline issue or XLA recompilation event.

---

Playbook 4: Node Failure / Hardware Issue

Trigger: Triage reports node-fail, signal-kill, or unknown-death. User wants to identify hardware problems.

Goal: Identify hardware errors (ECC, PCIe, XGMI) or driver issues that caused the failure.

Queries (over full job duration — hardware errors can accumulate before causing a crash):

#	PromQL	What it shows
1	increase(hw_gpu_ras_umc_ue_total[1h])	HBM uncorrectable ECC errors (fatal)
2	increase(hw_gpu_ras_umc_ce_total[1h])	HBM correctable ECC errors (accumulating = concern)
3	increase(hw_gpu_ras_xgmi_ue_total[1h])	XGMI/WAFL link uncorrectable errors
4	increase(hw_gpu_ras_xgmi_ce_total[1h])	XGMI/WAFL link correctable errors
5	increase(hw_gpu_ras_gfx_ue_total[1h])	Compute engine errors
6	increase(hw_gpu_pcie_fatal_total[1h])	PCIe fatal errors
7	increase(hw_gpu_pcie_nonfatal_total[1h])	PCIe non-fatal errors
8	increase(hw_dmesg_gpu_errors_total[1h])	GPU/driver errors in kernel log
9	hw_rdma_port_state	RDMA port went down?
10	hw_gpu_temperature_celsius	Thermal excursion before crash?

Interpretation:

• Any uncorrectable error (UE) > 0 → hardware fault. That GPU/link is bad. The node should be drained for maintenance.
• High correctable errors (CE) → degrading hardware. Not immediately fatal but indicates the component is failing.
• PCIe fatal → PCIe link reset or card disconnect. Likely unrecoverable without node restart.
• RDMA port state → 0 → network link went down. Check physical connectivity and switch.
• dmesg GPU errors increasing → driver-level GPU fault detected by the kernel.
• Temperature spike before crash → thermal throttling or shutdown. Check cooling.
• Compare across nodes — healthy nodes should have zero RAS/PCIe errors. Any node with non-zero values is the problem.

GPU driver fault / D-state hang (non-fatal but degrading):

Not all GPU hardware issues crash the job. A common pattern on AMD Instinct nodes is a kernel bug in the RAS sysfs reporting path that doesn't affect GPU compute but makes the monitoring interface unreliable. Diagnosis:

1. Detect: hw_dmesg_gpu_errors_total > 0 on one node, 0 on all others. hw_scrape_duration_seconds alternates between ~12s and <3s on the affected node (timeout vs success). hw_gpu_power_watts has intermittent gaps.

2. Confirm via dmesg (Ray Jobs API): Submit a Ray job to read dmesg --level=err,warn -T on the affected node. Look for amdgpu kernel oops, especially in the CPER (Common Platform Error Record) or RAS sysfs paths:

   • amdgpu_cper_ring_get_ent_sz / amdgpu_cper_ring_write — CPER ring buffer corruption
   • amdgpu_ras_aca_sysfs_read / aca_sysfs_read — RAS sysfs read hang
   • amdgpu_cper_generate_ce_records — correctable error record generation crash

3. Check D-state process accumulation (Ray Jobs API): Submit a Ray job to run ps axo pid,stat,wchan:30,cmd and count processes in D-state. Hundreds of D-state processes stuck in amdgpu_cper_ring_write confirms the driver bug is active and processes are accumulating (unkillable without reboot).

4. Assess impact:

   • Training: Unaffected. GPU compute uses /dev/kfd (KFD command submission), which is a completely separate code path from the RAS sysfs interface. The kernel bug is in the error-reporting path, not the compute path.
   • Monitoring: GPU hardware metrics (hw_gpu_power_watts, hw_gpu_temperature_celsius, hw_gpu_vram_used_bytes, RAS counters) are intermittently unavailable. Host and training metrics are unaffected.
   • D-state accumulation: ~6 unkillable processes per hour. Not immediately dangerous (Linux PID max = 4,194,304) but indicates the node needs a reboot.

5. Recovery: Not possible without a node reboot. D-state processes are stuck in kernel space and cannot be signaled. The CPER ring corruption persists in kernel memory. Schedule a reboot at the next job boundary.

6. Report for cluster admin: Include the node hostname, kernel version (uname -r), amdgpu module version (/sys/module/amdgpu/version), exact dmesg call stack, hw_dmesg_gpu_errors_total count, D-state process count, onset time (from dmesg timestamps and timeline), and confirmation that all other nodes are clean.

---

Playbook 5: GPU Health Check

Trigger: User says "check GPU health", "are the GPUs OK", "check thermals", or proactive monitoring of a running job.

Goal: Assess GPU health across the cluster — temperatures, power, clocks, VRAM, error counters.

Queries (range over full job or recent window):

#	PromQL	What it shows	Alert threshold
1	hw_gpu_temperature_celsius	Junction temperature	>90C = concern, >100C = throttling
2	hw_gpu_power_watts	Power draw	Variance >50W across GPUs = investigate
3	hw_gpu_clock_mhz{type="sclk"}	Core clock	Sudden drop = investigate
4	hw_gpu_vram_used_bytes / hw_gpu_vram_total_bytes	VRAM utilization	>95% = risk of GPU OOM
5	hw_gpu_ras_umc_ce_total	HBM correctable errors	Any >0 = accumulating hardware issue
6	hw_gpu_ras_umc_ue_total	HBM uncorrectable errors	Any >0 = bad GPU, drain node
7	hw_gpu_ras_xgmi_ce_total	XGMI correctable errors	Any >0 = inter-GPU link degrading
8	hw_gpu_pcie_correctable_total	PCIe correctable errors	Steady increase = link issue
9	ray_node_gpus_utilization	GPU utilization	<80% during training = underutilization

Report format for health check:

GPU Health Summary (<N> nodes, <G> GPUs)

Temperature:  min <X>C  max <X>C  avg <X>C  (threshold: 90C)
Power:        min <X>W  max <X>W  avg <X>W  spread: <X>W
Core clock:   min <X>MHz  max <X>MHz  (nominal: <X>MHz)
VRAM:         <X>% used  (<X> GB / <X> GB)
Utilization:  min <X>%  max <X>%  avg <X>%

RAS Errors:   <N> correctable, <N> uncorrectable
PCIe Errors:  <N> correctable, <N> non-fatal, <N> fatal

Anomalous GPUs: <list of host:gpu with any non-zero errors or outlier metrics>

---

Playbook 6: Network Health Check

Trigger: User says "check network", "is the network OK", "RDMA health", or investigating intermittent NCCL timeouts.

Goal: Assess network health — TCP retransmits, RDMA errors, congestion, port state.

Queries (range over full job or recent window):

#	PromQL	What it shows	Alert threshold
1	rate(hw_tcp_retransmits_total[5m])	TCP retransmit rate	>10/s sustained = concern
2	rate(hw_rdma_tx_retx_pkts_total[5m])	RDMA retransmit rate	Any sustained > 0 = concern
3	rate(hw_rdma_tx_ack_timeout_total[5m])	RDMA ACK timeouts	Any > 0 = link or switch issue
4	hw_rdma_port_state	Port state (1=ACTIVE)	Any 0 = port down
5	rate(hw_rdma_rx_cnp_pkts_total[5m])	Congestion notifications	Sustained = network congestion
6	rate(hw_rdma_req_tx_retry_excd_err_total[5m])	Retry exhaustion	Any > 0 = packets lost
7	rate(hw_rdma_req_rx_cqe_err_total[5m])	CQE errors	Any > 0 = RDMA errors
8	rate(hw_net_rx_errors_total[5m])	NIC RX errors	Any > 0 = NIC issue
9	rate(hw_net_tx_errors_total[5m])	NIC TX errors	Any > 0 = NIC issue
10	rate(hw_tcp_abort_on_timeout_total[5m])	TCP connections aborted	Any > 0 = severe

Interpretation:

• TCP retransmits only, no RDMA errors → IP/TCP path issue (NFS, coordinator gRPC), not the NCCL/RCCL data path. Critical: high TCP retransmits alone do NOT degrade training TGS. TCP carries NFS (checkpoint I/O, data loading) and gRPC (coordinator heartbeats), but RCCL collectives use RDMA. A job can have thousands of TCP retransmits/sec and still achieve full TGS if RDMA is clean.
• RDMA retransmits + ACK timeouts on specific devices → bad cable, bad port, or switch issue on that link. Identify the device and port labels. This is the #1 cause of unexplained TGS degradation. Even a few nodes with sustained RDMA retransmits can cause 20-30% TGS degradation cluster-wide because all-to-all collectives are synchronous — the slowest RDMA link bounds every node in every step.
• Congestion notifications (CNP) across many nodes → switch-level congestion. May need ECN tuning or traffic engineering.
• Retry exhaustion → packets permanently lost. Indicates a hard link failure, not transient congestion. Nodes with retry exhaustion should be top priority for exclusion.
• Port state 0 → RDMA link is down. Physical layer issue. Check cable and switch port.
• Correlate with training events — if retransmits spike at the same time as a step time increase or hang, the network event caused the training issue.

RDMA degradation signature in training metrics:

• RDMA-induced TGS drops appear as a constant step-time increase (not variable spikes), because every step's collective communication is uniformly slowed by the degraded link. This distinguishes RDMA issues from transient events (which cause isolated spikes) and checkpoint saves (which are periodic).
• Phase correlation: Map RDMA retransmit bursts on specific nodes to TGS degradation phases. If TGS drops when RDMA retransmits spike on node X, recovers when node X stops retransmitting, and drops again when node Y starts retransmitting, this proves a causal link. Use rate(hw_rdma_tx_retx_pkts_total[5m]) per host alongside tb_perf_per_device_tokens_per_sec over the same time window.
• Recovery test: When RDMA retransmits cease on the offending nodes, TGS should recover to baseline within 1-2 steps. If TGS does not recover after retransmits stop, there is an additional or different root cause.

Node exclusion prioritization (when spare nodes are limited):

Priority	Condition	Rationale
1 (highest)	RDMA retry exhaustion (hw_rdma_req_tx_retry_excd_err_total > 0)	Permanent packet loss — hard link failure
2	Sustained RDMA retransmits + ACK timeouts	Active link degradation slowing every collective
3	High CNP (congestion notifications)	Network congestion — may be switch-level, not node-level
4 (lowest)	TCP retransmits only	No RDMA impact — does not affect training TGS

Nodes with only TCP retransmit issues should not be excluded unless they also show RDMA problems. If spare capacity is limited, focus exclusions on priority 1-2 nodes.

---

Playbook 7: Training Stability

Trigger: User says "check training", "is training stable", "loss looks weird", "throughput dropping", or proactive monitoring.

Goal: Assess training health — loss convergence, gradient norms, throughput, step time consistency.

Important: Filter out synthetic anti-staleness fills. Only use data points where tb_metrics_plugin_staleness_fill == 0. The tb_* bridge is best-effort and may have gaps — if critical steps are missing, fall back to the raw TensorBoard event file (see "Recovering from tb_* gaps" in the Troubleshooting section).

Queries (range over full job or recent window):

#	PromQL	What it shows
1	tb_learning_loss and tb_metrics_plugin_staleness_fill == 0	Training loss (real data only)
2	tb_learning_grad_norm and tb_metrics_plugin_staleness_fill == 0	Gradient norm
3	tb_learning_raw_grad_norm and tb_metrics_plugin_staleness_fill == 0	Pre-clipping gradient norm
4	tb_perf_step_time_seconds and tb_metrics_plugin_staleness_fill == 0	Step time
5	tb_perf_per_device_tokens_per_sec and tb_metrics_plugin_staleness_fill == 0	Throughput
6	tb_learning_current_learning_rate and tb_metrics_plugin_staleness_fill == 0	Learning rate
7	tb_step	Current step (for progress tracking)
8	tb_learning_moe_lb_loss and tb_metrics_plugin_staleness_fill == 0	MoE load balance loss (if MoE model)

Interpretation — training metrics:

• Loss divergence (sudden increase or NaN) → learning rate too high, data issue, or numerical instability. Check if grad norm spiked at the same time.
• Gradient norm spikes → unstable training. If raw_grad_norm >> grad_norm, gradient clipping is active and may be too aggressive.
• MoE load balance loss spike → routing changed dramatically, causing expert load imbalance. This is a training-level root cause that will show up as uneven GPU utilization and TGS drop.
• Learning rate anomaly → verify the LR schedule matches expectations. A flat LR when warmup should be active (or vice versa) indicates a config issue.

Interpretation — step time and throughput:

• Step time periodic spikes → likely checkpoint saves. Correlate with checkpoint_period from the job config. Spikes of 10-30s every N steps are normal.
• Throughput regression → >10% drop from steady-state average warrants investigation. Run the contention checklist below.
• Step time gradual increase → resource contention worsening over time. Run the contention checklist below.

Contention checklist. When throughput drops or step time increases, systematically check all contention sources at the same timestamp. Any of these can silently degrade performance:

Contention source	Metrics to check	What to look for
CPU	hw_procs_running, hw_procs_blocked, hw_context_switches_total, ray_node_cpu_utilization, hw_gpu_user_processes	High runnable count (oversubscribed), high blocked count (I/O wait), excessive context switching. For multi-job comparison: a constant but higher process count in one job creates a constant baseline CPU tax — extra processes compete with data pipeline threads, NCCL coordination, and Python runtime. This won't show as a spike but as a persistent throughput gap. Compare hw_procs_running across jobs, not just within a single job. Known cause: enable_single_replica_ckpt_restoring=true leaks RCCL communicator polling threads on restore (~17 threads/host on a 3-replica run) because Orbax's _merge_globalized_replicas creates all-reduce collectives with unique GpuCliqueKeys cached permanently in XLA's C++ clique cache; jax.clear_caches() does not help (see "Checkpointing interference" section).
Network (TCP)	rate(hw_tcp_retransmits_total[5m]), rate(hw_tcp_estab_resets_total[5m]), rate(hw_tcp_abort_on_timeout_total[5m])	Retransmit bursts slow NFS and gRPC; resets and aborts indicate connection failures
Network (RDMA)	rate(hw_rdma_tx_retx_pkts_total[5m]), rate(hw_rdma_tx_ack_timeout_total[5m]), rate(hw_rdma_rx_cnp_pkts_total[5m])	RDMA retransmits slow NCCL/RCCL collectives; CNP indicates switch congestion
Storage I/O	hw_io_pressure_full_pct, hw_io_pressure_some_pct, hw_mem_dirty_bytes, hw_mem_writeback_bytes	I/O pressure stalls data loading and checkpointing; dirty page buildup indicates NFS/storage backlog
Memory	ray_node_mem_used, hw_oom_kills_total, hw_procs_blocked	Memory pressure causes swapping (blocked procs); approaching limit risks OOM

Check all rows — contention sources often compound (e.g., I/O pressure from checkpointing + network congestion from RDMA traffic = amplified step time spike).

Cross-domain correlation queries (the power of a unified TSDB):

# Did a temperature spike cause a throughput drop?
# Query both in the same time window and compare timestamps
tb_perf_per_device_tokens_per_sec{host="node001"}
hw_gpu_temperature_celsius{host="node001"}

# Did network issues cause step time spikes?
tb_perf_step_time_seconds{host="node001"}
rate(hw_tcp_retransmits_total{host="node001"}[5m])

# Did I/O pressure from checkpointing cause a loss spike?
tb_learning_loss{host="node001"}
hw_io_pressure_full_pct{host="node001"}

Multi-Job Comparison

When comparing metrics across two or more jobs (e.g., "why is job B slower than job A?"), follow these rules:

0. Triage all jobs first

Run the job-log-triage skill on each job being compared (per workflow step 0). For multi-job comparison, triage is especially critical because you need to know:

• Which jobs are running vs finished (determines Case A vs B for each)
• Which jobs are fresh starts vs checkpoint restores (restored jobs may have resource leaks — see "Checkpointing interference")
• The step range and checkpoint period for each job (needed to identify overlapping steps and exclude checkpoint boundaries)
• Any failures or hangs that affect which steps are comparable (e.g., don't compare post-hang idle metrics against active training)

1. Isolate config differences first

Before looking at metrics, compare the jobs' configurations. Parse PASSTHROUGH_ARGS from the header of each log file (first ~30 lines) and diff them. Common config parameters that affect performance:

• per_device_batch_size, max_target_length, steps
• remat_policy, quantization, attention
• ici_*_parallelism, dcn_*_parallelism (parallelism strategy)
• enable_checkpointing, checkpoint_period, async_checkpointing
• load_balance_loss_weight (MoE)
• _env_* flags (XLA flags, NCCL tuning)
• Number of nodes (NNODES), GPUs per node

If configs differ, the performance difference may be expected — the metric comparison should account for the config change, not treat it as an anomaly.

2. Align by training step, not wall clock