AIOps Capacity Planning Agent
When to Use
- Forecasting when a cluster will run out of CPU/memory/disk before it happens
- Right-sizing HPA min/max replicas based on historical utilization patterns
- Estimating Kafka broker capacity for projected message volume growth
- Analyzing Spark job history to recommend executor configuration
- Generating quarterly capacity planning reports for budget planning
Prometheus-Based Resource Forecasting
import pandas as pd
import numpy as np
from prophet import Prophet
from prometheus_api_client import PrometheusConnect
from datetime import datetime, timedelta
class ResourceForecaster:
"""Forecast resource usage to predict capacity exhaustion."""
def __init__(self, prometheus_url: str):
self.prom = PrometheusConnect(url=prometheus_url)
def fetch_metric_history(
self,
query: str,
days_back: int = 30,
step: str = "1h",
) -> pd.DataFrame:
end = datetime.now()
start = end - timedelta(days=days_back)
results = self.prom.custom_query_range(
query=query,
start_time=start,
end_time=end,
step=step,
)
if not results:
return pd.DataFrame()
df = pd.DataFrame(results[0]["values"], columns=["ds", "y"])
df["ds"] = pd.to_datetime(df["ds"], unit="s")
df["y"] = df["y"].astype(float)
return df
def forecast(
self,
metric_name: str,
query: str,
forecast_days: int = 30,
capacity_limit: float | None = None,
) -> dict:
df = self.fetch_metric_history(query)
if df.empty:
return {"error": f"No data for {metric_name}"}
model = Prophet(
daily_seasonality=True,
weekly_seasonality=True,
yearly_seasonality=False,
changepoint_prior_scale=0.05, # conservative — data infra grows smoothly
)
model.fit(df)
future = model.make_future_dataframe(periods=forecast_days * 24, freq="H")
forecast = model.predict(future)
result = {
"metric": metric_name,
"current_value": round(df["y"].iloc[-1], 2),
"forecast_30d_p50": round(forecast["yhat"].iloc[-1], 2),
"forecast_30d_p95": round(forecast["yhat_upper"].iloc[-1], 2),
}
if capacity_limit:
# Find when p95 forecast exceeds capacity
exceed = forecast[forecast["yhat_upper"] >= capacity_limit]
if not exceed.empty:
days_until_full = (exceed["ds"].iloc[0] - datetime.now()).days
result["days_until_capacity_breach"] = max(0, days_until_full)
result["breach_date"] = exceed["ds"].iloc[0].strftime("%Y-%m-%d")
result["action_required"] = days_until_full < 14
else:
result["days_until_capacity_breach"] = None
result["breach_date"] = None
return result
def run_capacity_forecast(prom_url: str) -> list[dict]:
"""Run full capacity forecast for all critical platform resources."""
forecaster = ResourceForecaster(prom_url)
reports = []
checks = [
{
"name": "airflow_scheduler_cpu",
"query": 'avg(rate(container_cpu_usage_seconds_total{pod=~"airflow-scheduler.*"}[5m]))',
"capacity_limit": 3.8, # 4 cores limit
},
{
"name": "kafka_broker_disk_pct",
"query": 'avg(1 - kubelet_volume_stats_available_bytes{pvc=~"data-production-kafka.*"} / kubelet_volume_stats_capacity_bytes) * 100',
"capacity_limit": 85.0, # alert at 85% full
},
{
"name": "spark_node_memory_gib",
"query": 'sum(container_memory_working_set_bytes{namespace="spark"}) / 1073741824',
"capacity_limit": 200.0, # 200 GiB cluster limit
},
]
for check in checks:
report = forecaster.forecast(
metric_name=check["name"],
query=check["query"],
capacity_limit=check.get("capacity_limit"),
)
reports.append(report)
return reports
Kubernetes HPA Right-Sizing
def recommend_hpa_settings(
deployment_name: str,
namespace: str,
prom_url: str,
target_cpu_utilization: float = 0.70,
p99_multiplier: float = 1.5, # headroom above p99 load
) -> dict:
"""
Analyze historical CPU utilization to recommend HPA min/max replicas
and target utilization threshold.
"""
prom = PrometheusConnect(url=prom_url)
# Fetch 30-day CPU utilization history (% of request)
query = f"""
avg(
rate(container_cpu_usage_seconds_total{{namespace="{namespace}", pod=~"{deployment_name}-.*"}}[5m])
)
/
avg(
kube_pod_container_resource_requests{{namespace="{namespace}", pod=~"{deployment_name}-.*", resource="cpu"}}
) * 100
"""
results = prom.custom_query_range(
query=query,
start_time=datetime.now() - timedelta(days=30),
end_time=datetime.now(),
step="5m",
)
if not results:
return {"error": "No metrics found"}
values = [float(v[1]) for v in results[0]["values"]]
values_arr = np.array(values)
# Current replica count
import subprocess, json
replicas_raw = subprocess.run(
["kubectl", "get", "deployment", deployment_name, "-n", namespace,
"-o", "jsonpath={.spec.replicas}"],
capture_output=True, text=True
).stdout
current_replicas = int(replicas_raw.strip() or "1")
p50_util = np.percentile(values_arr, 50)
p95_util = np.percentile(values_arr, 95)
p99_util = np.percentile(values_arr, 99)
# At target utilization, how many replicas do we need for p99 load?
recommended_max = int(np.ceil(current_replicas * (p99_util * p99_multiplier) / (target_cpu_utilization * 100)))
recommended_min = max(1, int(np.ceil(current_replicas * p50_util / (target_cpu_utilization * 100))))
return {
"deployment": deployment_name,
"namespace": namespace,
"current_replicas": current_replicas,
"cpu_p50_pct": round(p50_util, 1),
"cpu_p95_pct": round(p95_util, 1),
"cpu_p99_pct": round(p99_util, 1),
"recommended_hpa": {
"minReplicas": recommended_min,
"maxReplicas": recommended_max,
"targetCPUUtilizationPercentage": int(target_cpu_utilization * 100),
},
"current_hpa_adequate": current_replicas >= recommended_min,
}
Kafka Capacity Model
def kafka_capacity_model(
current_brokers: int,
current_topics: list[dict], # [{name, partitions, replication_factor, MB_per_sec}]
growth_rate_monthly_pct: float = 20.0, # 20% MoM growth
forecast_months: int = 6,
broker_max_throughput_MBps: float = 150.0, # per broker network capacity
broker_disk_TB: float = 2.0,
retention_days: int = 7,
) -> dict:
"""Forecast when current Kafka cluster needs scaling."""
total_write_MBps = sum(t["MB_per_sec"] * t["replication_factor"] for t in current_topics)
total_storage_TB = sum(
t["MB_per_sec"] * t["replication_factor"] * retention_days * 86400 / 1e6
for t in current_topics
)
cluster_throughput_capacity_MBps = current_brokers * broker_max_throughput_MBps
cluster_disk_TB = current_brokers * broker_disk_TB
# Monthly projections
monthly_projections = []
for month in range(1, forecast_months + 1):
factor = (1 + growth_rate_monthly_pct / 100) ** month
proj_throughput = total_write_MBps * factor
proj_storage = total_storage_TB * factor
throughput_pct = proj_throughput / cluster_throughput_capacity_MBps * 100
storage_pct = proj_storage / cluster_disk_TB * 100
needs_scaling = throughput_pct > 80 or storage_pct > 80
monthly_projections.append({
"month": month,
"throughput_MBps": round(proj_throughput, 1),
"storage_TB": round(proj_storage, 2),
"throughput_utilization_pct": round(throughput_pct, 1),
"storage_utilization_pct": round(storage_pct, 1),
"needs_broker_scaling": needs_scaling,
})
# When does first scaling need to happen?
scaling_month = next(
(p["month"] for p in monthly_projections if p["needs_broker_scaling"]),
None
)
brokers_needed_at_6m = int(np.ceil(
monthly_projections[-1]["throughput_MBps"] / (broker_max_throughput_MBps * 0.8)
))
return {
"current_state": {
"brokers": current_brokers,
"write_throughput_MBps": round(total_write_MBps, 1),
"storage_TB": round(total_storage_TB, 2),
},
"capacity_headroom": {
"throughput_pct_used": round(total_write_MBps / cluster_throughput_capacity_MBps * 100, 1),
"storage_pct_used": round(total_storage_TB / cluster_disk_TB * 100, 1),
},
"scaling_needed_in_month": scaling_month,
"brokers_needed_at_month_6": brokers_needed_at_6m,
"monthly_projections": monthly_projections,
}
Data Lake Storage Forecast
def forecast_s3_storage(
bucket_size_queries: list[dict], # [{prefix, current_TB, monthly_growth_TB}]
forecast_months: int = 12,
storage_budget_usd_month: float = 5000.0,
s3_standard_cost_per_gb: float = 0.023,
s3_ia_cost_per_gb: float = 0.0125,
) -> dict:
"""Forecast S3 storage costs and identify prefixes for tiering."""
projections = []
for month in range(1, forecast_months + 1):
monthly_data = []
for prefix_info in bucket_size_queries:
projected_tb = prefix_info["current_TB"] + prefix_info["monthly_growth_TB"] * month
monthly_data.append({
"prefix": prefix_info["prefix"],
"size_TB": round(projected_tb, 2),
})
total_tb = sum(d["size_TB"] for d in monthly_data)
cost_standard = total_tb * 1000 * s3_standard_cost_per_gb
cost_with_tiering = (
min(total_tb * 0.2, total_tb) * 1000 * s3_standard_cost_per_gb + # 20% hot
max(0, total_tb * 0.8) * 1000 * s3_ia_cost_per_gb # 80% IA
)
projections.append({
"month": month,
"total_TB": round(total_tb, 2),
"cost_standard_usd": round(cost_standard, 0),
"cost_with_tiering_usd": round(cost_with_tiering, 0),
"over_budget": cost_standard > storage_budget_usd_month,
})
budget_breach_month = next(
(p["month"] for p in projections if p["over_budget"]),
None
)
return {
"budget_breach_month": budget_breach_month,
"monthly_savings_from_tiering_usd": round(
projections[-1]["cost_standard_usd"] - projections[-1]["cost_with_tiering_usd"], 0
),
"projections": projections,
}
Capacity Planning Report Generator
def generate_capacity_report(prom_url: str, output_path: str = "capacity_report.md"):
"""Generate quarterly capacity planning report."""
forecasts = run_capacity_forecast(prom_url)
critical = [f for f in forecasts if f.get("days_until_capacity_breach") is not None and f["days_until_capacity_breach"] < 30]
warning = [f for f in forecasts if f.get("days_until_capacity_breach") is not None and 30 <= f["days_until_capacity_breach"] < 60]
lines = [
f"# Capacity Planning Report — {datetime.now().strftime('%Y-%m')}",
"",
f"**Generated**: {datetime.now().isoformat()}",
f"**Critical items** (< 30 days): {len(critical)}",
f"**Warning items** (30–60 days): {len(warning)}",
"",
"## Critical: Action Required",
]
for item in critical:
lines.append(
f"- **{item['metric']}**: breach in {item['days_until_capacity_breach']}d "
f"(~{item['breach_date']}), current={item['current_value']}, "
f"p95 forecast={item.get('forecast_30d_p95', 'N/A')}"
)
lines += ["", "## Warning: Monitor Closely"]
for item in warning:
lines.append(
f"- **{item['metric']}**: breach in {item['days_until_capacity_breach']}d "
f"(~{item['breach_date']})"
)
with open(output_path, "w") as f:
f.write("\n".join(lines))
return output_path
Anti-Patterns
- Planning only to current capacity — always plan to 1.5× current capacity; unexpected spikes (marketing campaigns, product launches) are the norm not the exception.
- Single-metric forecasting — a service can breach disk while CPU is healthy; forecast all dimensions: CPU, memory, disk, network, partitions.
- Not accounting for seasonal peaks — a 20% YoY growth rate ignores Black Friday 10× spikes; segment seasonality in the forecast model.
- Reacting to capacity instead of forecasting — adding brokers after the cluster is at 95% creates operational emergencies; start scaling at 70%.
- Capacity planning without cost modeling — scaling resources has a budget impact; always include cost forecasts alongside resource forecasts.
References
- Kubernetes HPA:
kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/
- Prometheus API client:
github.com/prometheus-api-client-python/prometheus-api-client-python
- Prophet forecasting:
facebook.github.io/prophet/
- Related skills:
[[aiops-infrastructure-anomaly-detection]], [[de-cost-optimization]], [[infra-kafka-platform-review]], [[infra-aws-data-platform-review]]
ivanshamaev