Why we stopped sampling session events
Sampling kept our bills low and quietly broke funnels for small games. Here is what we changed.
On this page3 sections

Until August, Cadence kept 1 in 10 session_start events for games with more than 50,000 daily players. It cut storage by a third. It also made funnels wrong in a way that was hard to spot.
What went wrong
A funnel needs every step from the same player. With sampling, a player’s session_start could be dropped while their level_complete was kept. The funnel then counted a finish without a start, and step two came out higher than step one.
Big games hid the error in the noise. A studio with 8,000 daily players saw a level that “converted” at 104%.
What we do now
- Every event is stored. No sampling on ingest.
- Raw events older than 90 days are rolled up into daily counts per player.
- Funnels and retention read raw events for the last 90 days, and roll-ups before that.
The cost
Storage went up 28%. We moved the raw tables to compressed columnar storage, which brought it back to 9% over the old bill. We’re keeping that, and prices don’t change.