Reserved capacity under-utilized

Provisioned throughput (Azure OpenAI PTUs, dedicated deployments with other providers) is capacity bought up front for a fixed monthly price. It saves at peak, but if real load is below what you committed, you pay for air. Unlike pay-as-you-go, nobody here sends “your bill went up” — the money is charged the same every month, and under-utilization isn’t visible anywhere except the dashboard you don’t visit.

The solution is a daily “committed vs used” reconciliation with a push on sustained under-load.

Daily committed vs used check

We compute the average utilization over the day and alert if it stays below the threshold for several days in a row — which means part of the PTUs can be cut.

import os, json, time, requests

NOTIFLY_URL   = os.environ["NOTIFLY_URL"]
NOTIFLY_TOKEN = os.environ["NOTIFLY_TOKEN"]

COMMITTED_PTU   = 100      # how many units are bought
COST_PER_PTU    = 260.0    # $/mo per unit (example)
MIN_UTIL        = 0.60     # below 60% average utilization = under-utilization
DAYS_TO_CONFIRM = 3        # this many days in a row confirm the conclusion

STATE = os.path.expanduser("~/.reserved-capacity.json")

def notify(title, msg, prio):
    requests.post(f"{NOTIFLY_URL}/message",
                  params={"token": NOTIFLY_TOKEN},
                  json={"title": title, "message": msg, "priority": prio},
                  timeout=5)

def load_state():
    if os.path.exists(STATE):
        return json.load(open(STATE))
    return {"low_days": 0, "last_day": ""}

def daily_check(used_ptu_avg):
    """used_ptu_avg — the average actual utilization over the day (from provider metrics)."""
    day = time.strftime("%Y-%m-%d")
    st = load_state()
    if st["last_day"] == day:          # already counted today — no duplicates
        return
    util = used_ptu_avg / COMMITTED_PTU

    if util < MIN_UTIL:
        st["low_days"] += 1
    else:
        st["low_days"] = 0
    st["last_day"] = day
    json.dump(st, open(STATE, "w"))

    if st["low_days"] >= DAYS_TO_CONFIRM:
        wasted = (COMMITTED_PTU - used_ptu_avg) * COST_PER_PTU
        notify(
            f"📉 Reserved-мощность простаивает: {util:.0%}",
            f"{DAYS_TO_CONFIRM} дня подряд утилизация < {MIN_UTIL:.0%}. "
            f"Впустую ≈${wasted:,.0f}/мес — сократите committed PTU.",
            priority=6,
        )

# usage: plug in the average utilization from Azure Monitor / provider metrics
daily_check(used_ptu_avg=48)

A threshold over several days in a row (DAYS_TO_CONFIRM) cuts off one-off quiet weekends — we react only to a sustained trend. priority=6: this is about money, not an incident, no need to wake you up at night.

Where to get used_ptu_avg

Actual utilization is exposed by provider metrics — for example, Azure Monitor’s Provisioned-managed Utilization metric. Pull it with a REST request in the same script or export it into your TSDB and read it from there. It’s convenient to hang the script itself on a timer-trigger in Yandex Cloud once a day — the skeleton is in the recipe Custom cloud function for integrity checks.

The mirror case — reserved overload

The reverse problem: utilization hits 100% and requests start spilling over to expensive pay-as-you-go. Add a second branch with util >= 0.95 and priority=8 — that’s already about money AND user latency.

What to put in the alert text

• committed vs average actual utilization (in units and in %);
• how many days in a row the under-load has held;
• an estimate of money wasted per month;
• a link to the quota management page, to cut the committed amount.