Case study · Distributed systems
Forecast-driven ad prefetch for Pluto TV's ad breaks
A statistical forecast picked the ad breaks expected to see the most demand, and the ads for them were fetched and processed ahead of time into Pluto TV's own storage, so those breaks did not depend on a burst of live calls the second they started.
- Product
- pluto.tv ↗
- Client
- Pluto TV
- Industry
- Free ad-supported streaming (FAST)
- Role
- Tech Lead · Solution designer
- Period
- 2020 – 2024
Scroll sideways to see the full diagram →
Context
Pluto TV is a free, ad-supported streaming service with linear channels: viewers on the same channel reach each ad break at the same moment. Its revenue services decide, fetch and prepare the ads for every break, calling third-party ad servers. The supply of ads was limited, and different platforms competed for the same ads.
The challenge
The moment of highest demand was also the moment when getting ads was hardest: every viewer of a channel reached the break in the same second, while other platforms were asking for the same limited ads. Filling those breaks could not depend on a burst of calls to the ad servers at the exact second each break started.
The solution
Get ahead of the break instead of racing it. A statistical forecast, fed with the event being streamed, the day and outside forecasts, picked the points expected to see the most demand. For those points, ads were fetched and processed ahead of time and kept in Pluto TV's own storage, so the break was served from storage. What actually happened was fed back into the statistics, so the forecast adjusted over time.
Demand forecast
Statistical model that picks the points expected to see the most demand, using the event, the day and outside forecasts as inputs.
Prefetch
For the points the forecast flags, asks for the ads and processes them ahead of the break.
Prepared-ads storage
Pluto TV's own storage holding the ads already fetched and processed, ready to serve when the break starts.
Feedback loop
What actually happened feeds the statistics back, so the forecast adjusts over time instead of staying a fixed guess.
Key decisions
- 01
Do the work before the break
When viewers arrive together, the cheapest request is the one already answered. Moving the calls to the ad servers and the processing ahead of time took them out of the second when every viewer of the channel arrived.
Alternatives considered Fetch every ad live when the break starts.
- 02
Prefetch where the forecast says, not everywhere
Ads were limited and other platforms competed for them, so the prefetch targeted the points expected to see the most demand.
- 03
A forecast that learns
Demand depends on the event, the day and conditions outside the platform. Feeding actual results back into the statistics let the forecast adjust over time.
What I did
- Designed the approach: forecast the breaks with the most demand and prefetch their ads ahead of time.
- Designed the architecture: forecast, prefetch, prepared-ads storage and the feedback loop.
- Led the team as Tech Lead.
Outcomes
- For the breaks the forecast flagged, asking the ad servers and processing the ads happened ahead of time, on a schedule the team controlled, instead of in the same second as every viewer of the channel.
- A forecast that adjusted over time with what actually happened, instead of a fixed guess.