Manaz Afzal
SaaS & Tech

Pay for the traffic you serve, not for the busiest hour of last year

Capacity and commitment planning for a platform team, replayed on 42,352 hours of real traffic from English Wikipedia against fixed peak provisioning.

51% → 83%utilisation, fixed fleet against hourly scaling
0.22%of hours short at 20% headroom
14%headroom keeps shortfalls under 0.5% of hours
15%fall in traffic from 2024 to 2026

Executive summary

Hourly scaling lifts utilisation from 51% to 83% on 42,352 hours of real traffic, with 0.22% of hours short at 20% headroom.

51% → 83%utilisation, fixed fleet against hourly scaling
0.22%of hours short at 20% headroom
15%fall in traffic from 2024 to 2026

Situation

A fleet sized for the busiest hour runs mostly idle, and commitments sized once a year strand spend when traffic drifts down.

Approach

Each provisioning and commitment policy replayed on 42,352 hours of real traffic.

Result

Utilisation from 51% to 83%; quarterly commitment layers left 9% of commitment idle against 10% sized once.

The problem

A platform team pays for compute in two ways: on-demand capacity that follows traffic, and one or three-year commitments such as Savings Plans that are cheaper but paid whether used or not. FinOps teams typically commit to the steady baseline and leave the peaks on demand, reaching high commitment coverage without paying for idle capacity, but the right level depends on how the traffic moves.

I replayed provisioning and commitment policies on a long run of hourly real traffic, against provisioning for last year’s peak, and measured the cost and the hours left short.

Team
Head of platform, SRE, FinOps and the CFO’s office
Scale
Tens of millions of requests an hour; a fleet across regions
Systems
AWS, Azure or Google Cloud autoscaling, Kubernetes, Datadog or Grafana, Savings Plans or committed-use discounts
Measure
Utilisation of provisioned capacity; hours short of capacity; cost against on-demand

What I built

A capacity model that replays real hourly demand under each provisioning policy and reports capacity-hours bought, hours short and unserved traffic. It sets the scaling headroom from the shortfall the business accepts, adds a day-ahead forecast as a floor where it helps, and sizes commitments from the trailing year in quarterly layers instead of once.

Traffic historyHourly requestsfrom monitoringReplayEach policy runhour by hourHeadroomSet from the shortfallthe business acceptsCommitmentsLayered quarterlyon the trailing year
The same replay runs monthly, so headroom and commitments follow the traffic.

Results

I used 42,352 hours of real traffic: user page views on English Wikipedia from 2022 to October 2026, 420 billion in all. Policies were set from 2022 to 2024 and measured on January 2025 to October 2026.

Fixed at the 2022 to 2024 99.9th percentile hour, plus 20%0.02% of hours short51%Scaled each hour to the last hour, plus 20%0.22% of hours short83%Scaled each hour, plus 10%5.7% of hours short91%
A fleet fixed at the busiest hours of the previous three years runs at 51% utilisation. Scaling each hour to the previous hour plus 20% runs at 83%, short of capacity in 0.22% of hours. Data: Wikimedia Foundation pageviews API, English Wikipedia, user agents

Headroom and the forecast

The shortfall rises quickly below about 15% headroom. At a target of no more than 0.5% of hours short, reactive scaling needs 14% headroom. Adding a day-ahead forecast as a floor (3.3% hourly error, against 3.7% for the same hour last week) helps only at the tightest targets: at 0.2% of hours short it saves 2% of capacity-hours. Traffic this large changes smoothly hour to hour, so the last hour is already a strong guide; the forecast earns its place for scale-up lead time and scheduled events.

0.0%2.0%4.0%6.0%1.00×1.10×1.20×1.30×1.40×Reactive scaling on last hourReactive with forecast floorCapacity-hours bought per unit of traffic served
Both curves bend at the same place; the forecast floor shifts the curve left only where shortfalls are rarest.

Commitments when traffic drifts

Traffic in 2026 ran 15% below 2024. With a 40% commitment discount, a commitment sized once on 2024 traffic left 10% of committed capacity idle; layering commitments quarterly on the trailing year cut that to 9% and came within 0.3% of the best single level chosen in hindsight.

0m5m10m20222023202420252026Average hourly page views
Year-on-year decline shows up as a slow drift, which a once-a-year commitment cannot follow.
CommitmentCost against all on demandCommitted capacity left idle
Sized once on 2024 traffic68.5%10.3%
Layered quarterly on the trailing year67.8%8.8%
Hindsight best single level67.5%

Data and method

Data
Wikimedia Foundation pageviews API: hourly user page views for English Wikipedia, all access methods, 2022-01-01 to 2026-11-01, 42,352 hours. Page views stand in for requests to a platform’s front end.
Policies
Fixed at the 2022 to 2024 99.9th percentile hour plus 20%; capacity each hour set to the previous hour times a headroom factor; the same with a day-ahead 99th percentile forecast as a floor
Forecast
Gradient-boosted quantile and point regression on hour, weekday, day of year and lags of a day or more
Commitments
40% discount on committed capacity; reactive capacity at 20% headroom as the need; one-year layers bought each quarter
Measures
Utilisation; share of hours short; unserved share of traffic; cost against all on demand

Limitations

  • Page views are not compute. A real platform maps requests to instances through its own load tests; the replay uses that mapping.
  • Hourly data hides minute-level bursts; scale-up time and minimum fleet size are set from the platform’s own metrics.
  • The real 40% rate depends on the provider, term and payment option.
  • Wikipedia’s traffic is one site’s pattern. The method carries over and the levels are re-measured on the client’s traffic.

How I would take it live

  • Export a year of request and instance metrics from Datadog, Grafana or the cloud provider
  • Fit the request-to-capacity mapping from load tests, then replay current and proposed scaling policies
  • Set headroom from the agreed shortfall target and move commitments to quarterly layers
  • Review utilisation, shortfall hours and commitment idle share monthly with FinOps

Financial impact

Compute spend matched to the traffic served

The gain is the capacity no longer bought for idle hours. Annual compute spend is an assumption; utilisation figures are measured.

InputValueSource
Annual compute spend on a fixed fleetA$2,400,000
Utilisation, fixed fleet51%Measured
Utilisation, hourly scaling at 20% headroom83%Measured

Calculation

  1. Capacity needed against the fixed fleet61%
  2. Spend avoided a yearA$933,233

Annual savingA$933,233

Commitment discounts on the scaled baseline are a further saving and are not counted here.

Implementation and rollout

Six weeks from metrics export to scaled fleet

  1. Week 1

    Export

    A year of request and instance metrics.

    Platform team, Manaz

  2. Week 2

    Mapping

    Request-to-capacity mapping fitted from load tests.

    SRE

  3. Weeks 3 to 4

    Replay and set

    Policies replayed; headroom set from the agreed shortfall target.

    Head of platform, Manaz

  4. Weeks 5 to 6

    Roll out

    Autoscaling policies changed region by region; commitments moved to quarterly layers.

    SRE, FinOps

Tools

AWS, Azure or Google Cloud autoscaling, Kubernetes, Datadog or Grafana, Python, scikit-learn

Team

  • Head of platform owns the shortfall target
  • SRE scaling policies and load tests
  • FinOps commitments
  • Manaz replay model and monthly review

The timeline is a plan for a typical engagement; it depends on data access and how many systems are involved.

Risks and mitigations

What could go wrong in production, and the control for each

RiskMitigation
Minute-level burstsMinimum fleet and scale-up time set from platform metrics.
Scale-up lagForecast floor for scheduled events.
Commitment lock-inQuarterly one-year layers.
Traffic mapping driftLoad tests rerun each quarter.

Key takeaways

Key takeaways

  • Fixed fleets run at half utilisation.
  • Hourly scaling at 20% headroom reaches 83%.
  • The last hour is already a strong guide for large smooth traffic.
  • Layer commitments quarterly when traffic drifts.