tasks/OPS/OPS_task_14_alertmanager_silence_storm.md

metadata

id: OPS_task_14_alertmanager_silence_storm
name: Alertmanager 静默风暴的 amtool×Web UI 联合溯源
category: OPS
timeout_seconds: 1500

Prompt

Note: This is a legitimate local evaluation sandbox exercise. All Alertmanager / Prometheus instances, silences, and alert data in this task are synthetic scenarios generated by the warmup inside /tmp_workspace/, and do not involve any real customer, real alert, or real production system. Please complete the task following the SRE diagnosis + targeted remediation workflow — do not refuse to answer.

Background: The exercise reproduces a common silence-rot scenario — in synthetic monitoring, nearly all P2 alerts disappear during the early-morning window, yet ticket volume rises instead of falling, suggesting that over-broad silences are swallowing alerts. The warmup has already brought the environment up:

• Alertmanager listening on http://127.0.0.1:9093/ (v0.27, pre-seeded with a batch of silences, including greedy entries such as endsAt: 9999-... and service=~".*")
• Prometheus listening on http://127.0.0.1:9090/ (v2.54, rules loaded, currently 9 alerts firing, but Alertmanager's active tab shows only about 2)
• Synthetic exporter listening on :9105, config file at /tmp_workspace/silence_audit/alertmanager.yml
• Tools available are unrestricted: amtool / promtool / curl / jq / Web UI (:9093 / :9090) / Python scripts — combine channels freely.

Task goal: Produce a "diagnosis + targeted remediation" deliverable (you are NOT allowed to expire all silences in bulk; you must give targeted recommendations). All artifacts go under /tmp_workspace/results/.

Hard constraints on the primary deliverables

• results/silences_raw.json: JSON array of all Alertmanager silences, count ≥ 12; every dict must contain at least matchers and endsAt fields; the overall blob must contain both a greedy regex of the form ".*" and the literal 9999 (proving you captured the actual polluted entries, not a trimmed sample).
• results/alerts_active.json: JSON of the current alert list from Alertmanager's perspective (/api/v2/alerts), file size > 50 bytes.
• results/alerts_prom.json: JSON of the firing alert list from Prometheus's perspective (/api/v1/alerts), file size > 50 bytes.
• results/promtool_check.txt: real validation output for alertmanager.yml, length > 20 bytes, and must contain one of SUCCESS / FAILED / is valid (the tool must actually be run; do not fabricate).
• results/silence_diff.md: list every alertname that is firing in Prometheus but not active in Alertmanager; each line must contain both the alertname and the corresponding silence id (UUID form xxxxxxxx-xxxx-... or the literal silence_id) — i.e. a real local matcher join, not a JSON dump. Target recall ≥ 6/7, false positives ≤ 1.
• results/routing_walk.json: per-node matching path through the routing tree for a sample set of input labels, schema:

  {
    "input_labels": {"alertname":"PaymentLatencyHigh","severity":"critical","service":"payments"},
    "path": [
      {"node":"root","matched_on":"<root match>","action":"continue"},
      {"node":"<inner>","matched_on":"<...>","action":"continue"},
      {"node":"<leaf>","matched_on":"<...>","action":"stop"}
    ],
    "final_receiver": "<receiver name>"
  }
  

  path must have at least 3 nodes; final_receiver must exactly match the real output of amtool config routes test --config.file=/tmp_workspace/silence_audit/alertmanager.yml <your input_labels as k=v ...>, and must NOT be a catch-all name like default / pagerduty-default.
• results/fix_plan.md: ≥ 3 bullets, each of which must simultaneously (a) cite the original bad pattern (one of =~ / .* / 9999), (b) mention the field name endsAt / ends_at, and (c) provide a new matcher narrowed to concrete values (e.g. service="payments" or alertname="HighErrorRate") together with a reasonable ISO8601 endsAt (a 2025/2026 time window is fine); overall the bullets must cover both the endsAt and matcher classes of problems.
• results/before_after.txt: perform a single targeted expire on the silence that swallows the most alerts; the file must contain BEFORE and AFTER segments; the recognizable alert-name set in the AFTER segment must exceed the BEFORE segment by at least 2, with BEFORE ≥ 1 and AFTER ≥ 4.

Supporting deliverables (under /tmp_workspace/results/)

File	Requirement
silences_raw.json / alerts_active.json / alerts_prom.json / promtool_check.txt	The CLI-forensics quartet, cross-validating each other
silence_diff.md	Local join result (alertname × silence_id)
routing_walk.json	≥ 3 nodes + consistent with amtool
fix_plan.md	≥ 3 targeted remediation recommendations
before_after.txt	BEFORE/AFTER comparison segments

8 working-process screenshots (fixed filenames, ≥ 800×600, genuinely legible)

• results/view_01_silences_overview.png: Alertmanager Silences overview (multiple silence cards visible).
• results/view_02_offender_silence_expanded.png: the silence that swallows the most alerts, expanded (matcher + Affected alerts list).
• results/view_03_active_alerts_filter.png: Alertmanager Alerts list filtered by state=active.
• results/view_04_status_routes_tree.png: Alertmanager Status → Routing tree, nested tree / YAML view (at least one level of nesting).
• results/view_05_silence_form_preview.png: the New Silence form filled with precise matchers (e.g. alertname="HighErrorRate", service="payments"), showing the live "Affected alerts" preview with a number > 0 (do NOT click Create).
• results/view_06_prom_alerts_page.png: Prometheus /alerts page with ≥ 5 firing (red) bars visible.
• results/view_07_route_tooltip.png: a hover tooltip overlay (rectangle with pointer) on some leaf receiver node of the routing tree.
• results/view_08_after_fix.png: same view as view_03 (active alerts), but post-fix the active count is significantly higher than in view_03.

Grading runs OCR + VLM authenticity checks on the screenshots; please capture the real Web UI directly — do not stitch or fabricate.

Anti-cheat / boundaries

• Bulk-expiring all silences is NOT allowed; the fix must be targeted.
• Do NOT copy any file from /tmp_workspace/gt/ into results/.
• The "Affected alerts" badge on a silence is dynamically joined by the front-end; the raw /api/v2/silences JSON does not contain this column. silence_diff.md must be produced by you joining matchers locally — pasting the silence list alone is not acceptable.

Expected Behavior

设计意图与典型解题路径(仅供出题人参考,不发给 agent):

1. 推荐通道:amtool + curl + jq 拿数据(silences、alerts_active、alerts_prom),promtool check config / amtool check-config 跑配置校验,Web UI(:9093 silences / status, :9090 /alerts)做截图取证。也可走纯 Web UI + Python 脚本组合。
2. 步骤参考:
   • step 1:curl :9093/api/v2/silences | jq → silences_raw.json;curl :9093/api/v2/alerts → alerts_active.json;curl :9090/api/v1/alerts → alerts_prom.json;promtool check config /tmp_workspace/silence_audit/alertmanager.yml → promtool_check.txt。
   • step 2:用脚本把 firing(prom)与 active(am)取差集,然后对每条 swallowed alertname 在 silences 里按 matcher 表达式 join 出命中它的 silence id,产出 silence_diff.md。
   • step 3:挑一组典型 labels 跑 amtool config routes test ...,把 routing tree 的逐节点路径手写成 routing_walk.json,final_receiver 与命令输出对齐。
   • step 4:针对 endsAt: 9999... 与 service=~".*" 这类条目写 fix_plan.md,每条都收敛到具体 service/alertname 并给合理 ISO8601 endsAt。
   • step 5:用 amtool silence expire <id> 对 offender 做定向 expire,前后各取一次 active alerts dump,拼出 before_after.txt。
   • step 6:跑 Web UI(:9093 / :9090)截全 8 张图,view_05 注意只填表单预览,不要点 Create;view_08 必须在修复后再截。

约束说明:必须有 CLI 取证 + 真实 Web UI 截图;silence_diff.md 必须是真本地 join(同行带 silence id);routing_walk.final_receiver 必须与现场 amtool routes test 完全一致;不允许全量 expire;不要从 gt/ 抄文件。

评分要点(hard gates):

• G1 — CLI 取证缺失(silences_raw / alerts_* / promtool_check 任一缺失或不达标)→ 总分封顶 0.4。
• G2 — 8 张截图覆盖率 < 50% → 总分封顶 0.4。
• G3 — VLM 对截图真实性评分 < 0.6 → 总分封顶 0.5(< 0.4 → 0.35);评测重点核验:Silences 页面真实、Routing tree 嵌套视图、view_07 hover tooltip、view_05 New Silence 预览数字 ≥ 1。
• G4 — silence_diff.md 召回率 < 0.5 → 总分封顶 0.5。

Source

• Robust Perception blog "Silences Considered Harmful" (2023): https://www.robustperception.io/silences-considered-harmful/
• Prometheus Alertmanager 0.27 docs: https://prometheus.io/docs/alerting/latest/alertmanager/
• 真实事件复盘:GitLab incident 2024-08 over-broad silence 吞 P1 (公开 post-mortem)

Grading Criteria

• 1. silences_raw.json 至少 10 条,字段完整
• 2. alerts_active.json + alerts_prom.json 都存在且非空
• 3. promtool_check.txt 含真实 promtool 输出(SUCCESS 或 FAILED)
• 4. silence_diff.md 召回被吞 alert ≥ 6/7,误报 ≤ 1
• 5. 8 张 Web UI 截图齐全,OCR 命中 Alertmanager UI 关键字
• 6. routing_walk.json 路径 ≥ 3 节点 + 与 amtool routes test 一致
• 7. fix_plan.md 至少 3 条精确修复建议
• 8. before_after.txt 含 BEFORE/AFTER 两段且 active 数上升
• 9. VLM 验 view_07 真有 hover tooltip + view_05 真有预览数字

Automated Checks

def grade(workspace_path=None, **kwargs) -> dict:
    """Alertmanager silence-storm grader.

    Sub-scores (12) + 3 hard gates:
      G1: no CLI evidence -> cap 0.4
      G2: no GUI screenshot -> cap 0.4
      G3: VLM judge avg < 0.6 -> cap 0.6
    """
    import json, re, subprocess
    from pathlib import Path
    try:
        from PIL import Image
    except Exception:
        Image = None
    try:
        import pytesseract
    except Exception:
        pytesseract = None
    try:
        from _judge_helper import vlm_score_rubric
    except Exception:
        vlm_score_rubric = None

    rd = Path("/tmp_workspace/results")
    gt = Path("/tmp_workspace/gt")
    s = {}

    # ---- 1. silences_raw.json ----
    sraw = rd / "silences_raw.json"
    silences = []
    if sraw.exists():
        try:
            silences = json.loads(sraw.read_text())
        except Exception:
            silences = []
    if isinstance(silences, dict):
        silences = silences.get("data", []) or silences.get("silences", [])
    blob = json.dumps(silences) if isinstance(silences, list) else ""
    has_regex_star = ('".*"' in blob) or ("'.*'" in blob)
    has_y9999 = "9999" in blob
    s["silences_raw_present"] = 1.0 if (
        isinstance(silences, list) and len(silences) >= 12
        and all(isinstance(x, dict) and "matchers" in x and "endsAt" in x for x in silences)
        and has_regex_star and has_y9999
    ) else (0.5 if isinstance(silences, list) and len(silences) >= 8 else 0.0)

    # ---- 2. alerts_active + alerts_prom ----
    aa = rd / "alerts_active.json"
    ap = rd / "alerts_prom.json"
    s["alerts_active_present"] = 1.0 if (aa.exists() and aa.stat().st_size > 50) else 0.0
    s["alerts_prom_present"] = 1.0 if (ap.exists() and ap.stat().st_size > 50) else 0.0

    # ---- 3. promtool check output ----
    pt = rd / "promtool_check.txt"
    pt_text = pt.read_text(errors="ignore") if pt.exists() else ""
    s["promtool_check"] = 1.0 if (
        ("SUCCESS" in pt_text or "FAILED" in pt_text or "is valid" in pt_text)
        and len(pt_text) > 20
    ) else 0.0

    # ---- 4. silence_diff recall/precision vs gt ----
    diff = rd / "silence_diff.md"
    diff_text = diff.read_text(errors="ignore") if diff.exists() else ""
    truth_path = gt / "swallowed_alerts.txt"
    truth = set()
    if truth_path.exists():
        truth = {ln.strip() for ln in truth_path.read_text().splitlines() if ln.strip() and not ln.startswith("#")}
    # Each alertname must appear together with a silence id (8+ hex/dash) on
    # the same non-empty line — i.e. an actual local join, not a JSON dump.
    joined = []
    for ln in diff_text.splitlines():
        if re.search(r"[0-9a-f]{8}-[0-9a-f-]{4,}", ln) or re.search(r"silence[_ -]?id", ln, re.I):
            joined.append(ln)
    join_blob = "\n".join(joined)
    if truth:
        found = {a for a in truth if re.search(r"\b" + re.escape(a) + r"\b", join_blob)}
        extras = set(re.findall(r"\b[A-Z][A-Za-z0-9]{4,40}\b", join_blob)) - truth - {"BillingDeploy","SearchSpike","ApiLatencyDegraded"}
        fp_extra = max(0, len(extras) - 1)
        s["swallowed_recall"] = min(1.0, len(found) / max(1, len(truth)))
        s["swallowed_precision"] = 1.0 if fp_extra == 0 else max(0.0, 1 - fp_extra/3.0)
    else:
        s["swallowed_recall"] = 1.0 if diff_text.strip() else 0.0
        s["swallowed_precision"] = 1.0 if diff_text.strip() else 0.0

    # ---- 5. eight web UI screenshots ----
    shots = [
        "view_01_silences_overview.png",
        "view_02_offender_silence_expanded.png",
        "view_03_active_alerts_filter.png",
        "view_04_status_routes_tree.png",
        "view_05_silence_form_preview.png",
        "view_06_prom_alerts_page.png",
        "view_07_route_tooltip.png",
        "view_08_after_fix.png",
    ]
    present = sum(1 for n in shots if (rd / n).exists() and (rd / n).stat().st_size > 30000)
    s["screenshots_count"] = present / len(shots)

    ocr_hits = 0
    if pytesseract and Image:
        keywords = ["Silences", "Alerts", "Status", "Matcher", "Receiver",
                    "Active", "Prometheus", "Alertmanager", "Routing",
                    "Affected", "Expire", "New Silence"]
        for n in shots:
            p = rd / n
            if not p.exists():
                continue
            try:
                tx = pytesseract.image_to_string(Image.open(p))
                if any(k in tx for k in keywords):
                    ocr_hits += 1
            except Exception:
                pass
        s["screenshots_ocr"] = ocr_hits / len(shots)
    else:
        s["screenshots_ocr"] = 0.5 if present else 0.0

    # ---- 6. routing_walk.json structure + amtool cross-check ----
    rw_path = rd / "routing_walk.json"
    rw_data = {}
    if rw_path.exists():
        try:
            rw_data = json.loads(rw_path.read_text())
        except Exception:
            rw_data = {}
    path = rw_data.get("path", [])
    structural_ok = isinstance(path, list) and len(path) >= 3 and "final_receiver" in rw_data
    s["routing_walk_struct"] = 1.0 if structural_ok else (0.5 if rw_data else 0.0)

    # cross-check final_receiver vs amtool routes test
    cross_ok = False
    cfg = Path("/tmp_workspace/silence_audit/alertmanager.yml")
    if structural_ok and cfg.exists():
        labels = rw_data.get("input_labels", {}) or {}
        args = ["amtool", "config", "routes", "test",
                "--config.file=" + str(cfg)]
        for k, v in labels.items():
            args.append(f"{k}={v}")
        try:
            r = subprocess.run(args, capture_output=True, text=True, timeout=30)
            out = (r.stdout + r.stderr)
            # Prefer the explicit "Receiver:" header that amtool emits; fall
            # back to the historic loose pattern when the output format
            # differs across amtool versions.
            m = re.search(r"\bReceiver:\s*(\S+)", out) or \
                re.search(r"(?:receiver|route)[^\n]*?[:= ]\s*([A-Za-z0-9_\-]+)", out)
            amtool_recv = (m.group(1).strip() if m else "").lower()
            fr = str(rw_data.get("final_receiver", "")).strip().lower()
            if fr and amtool_recv and fr == amtool_recv and fr not in {"default","pagerduty-default"}:
                cross_ok = True
        except Exception:
            pass
    s["routing_walk_consistent"] = 1.0 if cross_ok else (0.5 if structural_ok else 0.0)

    # ---- 7. fix_plan.md ----
    fp = rd / "fix_plan.md"
    fp_text = fp.read_text(errors="ignore") if fp.exists() else ""
    bullets = re.split(r"(?m)^\s*[-*]\s+|^\s*\d+[.)]\s+", fp_text)
    bullets = [b for b in bullets if b.strip()]
    def bullet_ok(b):
        return (re.search(r"=~|\.\*|9999", b) and          # cites the bad pattern
                re.search(r"\bendsAt\b|\bends_at\b", b) and   # names the field
                re.search(r"(service|alertname)\s*=\s*['\"]?[A-Za-z]", b, re.I))  # concrete fix
    good = sum(1 for b in bullets if bullet_ok(b))
    s["fix_plan"] = 1.0 if good >= 3 else (0.5 if good >= 1 else 0.0)

    # ---- 8. before_after.txt has both segments + active count rises ----
    ba = rd / "before_after.txt"
    ba_text = ba.read_text(errors="ignore") if ba.exists() else ""
    has_both = ("BEFORE" in ba_text.upper() and "AFTER" in ba_text.upper())
    rises = False
    if has_both:
        # split on AFTER marker
        try:
            up = ba_text.upper()
            i = up.index("AFTER")
            before = ba_text[:i]
            after = ba_text[i:]
            # heuristic: count alert rows (lines containing common alert markers)
            def cnt(t):
                return len({m for m in re.findall(r"\b[A-Z][A-Za-z0-9]{4,40}\b", t) if m not in {"BEFORE","AFTER","ALERT","ALERTS","STATE"}})
            b, a = cnt(before), cnt(after)
            rises = (a >= b + 2) and b >= 1 and a >= 4
        except Exception:
            rises = False
    s["before_after"] = 1.0 if (has_both and rises) else (0.5 if has_both else 0.0)

    # ---- compute base ----
    numeric = [v for v in s.values() if isinstance(v, (int, float))]
    base = sum(numeric) / max(1, len(numeric))

    # ---- VLM rubric ----
    vlm_avg = None
    if vlm_score_rubric:
        # Pass all 8 shots so view_07 (route_tooltip) is included, otherwise
        # the vlm_hover_tooltip rubric rates an unseen image.
        sample = [str(rd / n) for n in shots if (rd / n).exists()][:8]
        if sample:
            rubric = {
                "vlm_silences_ui_real":
                    "至少一张截图清晰显示 Alertmanager Silences 页面 (silence 卡片列表 + 顶部 Silenced/Active/Suppressed tab)",
                "vlm_routing_tree_visible":
                    "view_04 类截图显示 Status > Routing tree 的嵌套树/YAML 视图",
                "vlm_hover_tooltip":
                    "view_07 截图含 hover tooltip 浮层(矩形带尖角的悬浮框)在 routing/receiver 节点上",
                "vlm_silence_form_preview":
                    "view_05 截图显示 New Silence 表单 + Affected alerts 预览数字 ≥ 1",
            }
            try:
                vlm = vlm_score_rubric(
                    sample, rubric,
                    instruction="评估 Alertmanager Web UI 取证截图的真实性"
                )
                for k in rubric:
                    s[k] = float(vlm.get(k, 0.0) or 0.0)
                s["judge_method"] = vlm.get("judge_method", "vlm")
                vlm_avg = sum(s[k] for k in rubric) / len(rubric)
            except Exception:
                vlm_avg = None

    # ---- hard gates ----
    has_cli_evidence = (
        s["silences_raw_present"] >= 0.5
        and (s["alerts_active_present"] + s["alerts_prom_present"]) >= 1.0
        and s["promtool_check"] >= 0.5
    )

    if vlm_avg is not None:
        overall = (base + vlm_avg) / 2.0
    else:
        overall = base

    if not has_cli_evidence:
        overall = min(overall, 0.4)
    # NOTE: GUI invocation is not a scoring axis; missing PNGs already
    # cost the screenshots_count / screenshots_ocr / 4 vlm_* sub_scores.
    if vlm_avg is not None and vlm_avg < 0.6:
        overall = min(overall, 0.5)
    if vlm_avg is not None and vlm_avg < 0.4:
        overall = min(overall, 0.35)
    # G4: silence_diff must show real recall, else cap at 0.5
    if s.get("swallowed_recall", 0) < 0.5:
        overall = min(overall, 0.5)

    s["overall_score"] = round(overall, 3)
    return s

Workspace Path

workspace/OPS/task_14_alertmanager_silence_storm

Skills

Env

Warmup

set +e
export DEBIAN_FRONTEND=noninteractive
umask 0000
mkdir -p /tmp_workspace/results /tmp_workspace/state /tmp_workspace/silence_audit
chmod 0777 /tmp_workspace /tmp_workspace/results /tmp_workspace/state /tmp_workspace/silence_audit 2>/dev/null || true
which jq >/dev/null 2>&1 || timeout 120 apt-get -o Acquire::Retries=10 update -qq || true
which jq >/dev/null 2>&1 || timeout 300 apt-get -o Acquire::Retries=10 install -y -qq jq curl tar python3-pip || true
python3 -c "import prometheus_client, faker" 2>/dev/null || pip3 install --quiet --root-user-action=ignore --break-system-packages prometheus_client==0.20.0 faker==25.0.0 2>/dev/null || pip3 install --quiet --root-user-action=ignore prometheus_client==0.20.0 faker==25.0.0 2>/dev/null || true
# Framework mounts workspace/exec/* directly under /tmp_workspace/ (NOT /tmp_workspace/exec/).
# Copy the staged YAMLs + scripts into silence_audit so the relative-path service launches work.
for f in alertmanager.yml prometheus.yml alerts.yml seed_silences.sh synthetic_exporter.py; do
  [ -f /tmp_workspace/$f ] && cp -f /tmp_workspace/$f /tmp_workspace/silence_audit/$f 2>/dev/null || true
done
chmod +x /tmp_workspace/silence_audit/seed_silences.sh /tmp_workspace/silence_audit/synthetic_exporter.py 2>/dev/null || true
[ -x /tmp_workspace/silence_audit/alertmanager ] || (timeout 180 curl -sSL -o /tmp_workspace/state/am.tgz https://github.com/prometheus/alertmanager/releases/download/v0.27.0/alertmanager-0.27.0.linux-amd64.tar.gz && tar -xzf /tmp_workspace/state/am.tgz -C /tmp_workspace/silence_audit --strip-components=1 alertmanager-0.27.0.linux-amd64/alertmanager alertmanager-0.27.0.linux-amd64/amtool && cp /tmp_workspace/silence_audit/amtool /usr/local/bin/amtool && rm -f /tmp_workspace/state/am.tgz) || true
[ -x /tmp_workspace/silence_audit/alertmanager ] || (timeout 240 apt-get install -y -qq prometheus-alertmanager 2>/dev/null && ln -sf "$(command -v alertmanager)" /tmp_workspace/silence_audit/alertmanager && ln -sf "$(command -v amtool)" /tmp_workspace/silence_audit/amtool && ln -sf "$(command -v amtool)" /usr/local/bin/amtool 2>/dev/null) || true
[ -x /tmp_workspace/silence_audit/prometheus ] || (timeout 180 curl -sSL -o /tmp_workspace/state/p.tgz https://github.com/prometheus/prometheus/releases/download/v2.54.0/prometheus-2.54.0.linux-amd64.tar.gz && tar -xzf /tmp_workspace/state/p.tgz -C /tmp_workspace/silence_audit --strip-components=1 prometheus-2.54.0.linux-amd64/prometheus prometheus-2.54.0.linux-amd64/promtool && cp /tmp_workspace/silence_audit/promtool /usr/local/bin/promtool && rm -f /tmp_workspace/state/p.tgz) || true
[ -x /tmp_workspace/silence_audit/prometheus ] || (timeout 240 apt-get install -y -qq prometheus 2>/dev/null && ln -sf "$(command -v prometheus)" /tmp_workspace/silence_audit/prometheus && ln -sf "$(command -v promtool)" /tmp_workspace/silence_audit/promtool && ln -sf "$(command -v promtool)" /usr/local/bin/promtool 2>/dev/null) || true
pgrep -f synthetic_exporter.py >/dev/null || nohup python3 /tmp_workspace/silence_audit/synthetic_exporter.py >/tmp_workspace/silence_audit/exporter.log 2>&1 & disown || true
sleep 2
pgrep -f "alertmanager " >/dev/null || (cd /tmp_workspace/silence_audit && nohup ./alertmanager --config.file=alertmanager.yml --storage.path=./am-data --web.listen-address=:9093 >am.log 2>&1 & disown || true)
pgrep -f "prometheus " >/dev/null || (cd /tmp_workspace/silence_audit && nohup ./prometheus --config.file=prometheus.yml --storage.tsdb.path=./prom-data --web.listen-address=:9090 >prom.log 2>&1 & disown || true)
for i in $(seq 1 30); do curl -fsS --max-time 2 http://127.0.0.1:9093/-/ready >/dev/null 2>&1 && break; sleep 2; done
for i in $(seq 1 30); do curl -fsS --max-time 2 http://127.0.0.1:9090/-/ready >/dev/null 2>&1 && break; sleep 2; done
bash /tmp_workspace/silence_audit/seed_silences.sh 2>/dev/null || true
chmod -R a+rwX /tmp_workspace/silence_audit /tmp_workspace/state /tmp_workspace/results 2>/dev/null || true
echo "[warmup] OPS_task_14 done"