1 min lesson
Tiering and the replay window
Rebuild the parts of "Tiering and the replay window", then say why each one matters.
Step 1 of 2
Tiering and the replay windowthe question that decides retention
- Hot
- Recent, frequently queried - fast storage, full fidelity, fully indexed
- Cold / archive
- Old, rarely read - cheap tier, accept slower retrieval
- Replay window
- How far back you might need to reprocess raw from scratch
- Past the window
- Drop or pre-aggregate raw - keep the rollup, shed the row-level cost
The replay window is the honest answer to "why are we still paying to store 18-month-old raw events?"
Cost is the differentiator most data candidates forget. Whatever the design question - ingestion, modeling, observability - close with a cost angle unprompted: "and here's what this costs and the one lever I'd pull first." For unit economics, frame it as cost per million events or per query, not the raw monthly bill - the ratio tells you whether spend grew from more usage (fine) or more waste (a problem). That instinct is exactly the pragmatic, cost-conscious judgment the JD calls out.
Learn more
Optional practice
Practice: Tiering and the replay window
QReads on a streaming-ingested bronze table have gotten painfully slow and the storage bill is climbing, even though total data volume looks reasonable. What's the most likely root cause and the first fix?