At a glance:
- A campaign can hit every impression target while partner fill rates drop, because impression volume and fill rate measure two different things entirely.
- Average clearing latency can hold steady while p99 climbs into the range that costs partners bids, so the metric that looks fine isn’t the one driving the complaint.
- Auto-scaling adds capacity, but new instances land on the same contended shared infrastructure, so it doesn’t restore the headroom that’s already gone.
- Shared compute has enough slack to absorb load at moderate utilization and none once every tenant is running hot at the same time, which is when a premium campaign tends to land.
- A monthly average will say 94% again next month and still be accurate, and it still won’t be the number that decides whether this happens again.
What SSP clearing performance looks like at steady state vs. under a major campaign push
Your weekly performance review starts with the same dashboard it always does: auction clearing latency averaging 28 milliseconds and demand partner fill rates holding at 94%. Everything has been running clean.
Then a media agency calls with good news: a major advertiser is moving a significant Q4 holiday campaign through your exchange. Expected impression volume is about 3X your current daily average and arrives in 48 hours.
Your team spends the next day confirming everything is ready. Load tests at elevated QPS look acceptable, and auto-scaling policies are reviewed and updated.
The campaign launches on schedule.
Three hours after go-live, the first message arrives from a demand partner. Their fill rate has dropped from 94% to 81% on your exchange. They aren’t seeing any connectivity issues on their side, so they want to know what’s happening.
Your dashboards show average clearing latency at 34 milliseconds, still within spec. But p99 clearing latency is at 97 milliseconds and climbing.
Some bids are clearing inside the partners’ response windows, and others are clearing outside. The partners hitting the slow tail are dropping them. The rest are fine.
From your side, the system is handling the load. From the partner’s side, your exchange started behaving inconsistently the moment the premium campaign arrived.
Your team pulls the utilization data and sees that the clearing engine’s compute nodes are at 78%, up from their normal 45%. Auto-scaling launched new instances during the ramp-up, but provisioning took several minutes. The nodes handling clearing traffic are running on shared infrastructure.
During that window, the existing nodes handled the full volume on their own. The new instances picked up the extra load once they came online, but the baseline nodes are still running at 78%, and p99 latency on those nodes reflects the contention.
That’s the part auto-scaling doesn’t solve. It adds compute capacity, but the new instances land on the same shared infrastructure as the nodes already under contention, competing for the same physical CPU cycles as every other tenant running hot. More nodes doesn’t restore the headroom that disappeared at 78%; it just puts more nodes into a pool where that headroom is already gone.
The clearing engine is getting less consistent access to compute than it’s allocated on paper, because a shared environment can’t guarantee that allocation when every tenant is running hot at the same time. At 45% utilization there was enough headroom to absorb that; at 78%, there isn’t.
By the time the fill rate conversation happens at 3pm, the advertiser’s campaign has already hit its impression targets. The partner is calling about something different: the clearing performance they saw while it was running.
The monthly performance report will show your exchange cleared 94% of demand partner bids on average across the month. The partner SLA review will show fill rates dropped 13 percentage points during the highest-value campaign of the quarter, and "we were on shared infrastructure under contention" doesn’t make that conversation any easier.
The performance report for the week before the campaign still says 94%, and that number is accurate. It just doesn’t describe the three hours that actually mattered to the partner on the phone.
None of it was bad luck. The clearing engine ran fine at 45% utilization because there was headroom to run fine on. It ran into trouble at 78% because that headroom was never really there once every tenant on that infrastructure started running hot at the same time. The next campaign that pushes past that same line gets the same call, from the same kind of partner, unless something about the environment changes before it does.
