GUIDES · OPERATE
Capacity planning for ingest spikes.
Ingest is elastic, but elasticity has a slope. Pre-warm ahead of the spikes you can predict, learn to read the backpressure signals for the ones you cannot, and know the thresholds that say scale now.
Plan for known spikes
A product launch, a migration, a holiday sale — these are predictable. Pre-warm the cell so capacity is in place before the traffic, rather than chasing it on the way up.
SHELLpre-warm $ td fabric prewarm --factor 3x --until "2026-07-04T23:00Z" -> current sustained: 4.6M events/s -> target headroom: 13.8M events/s ok warm by 06:00 . auto-releases after the window
Read backpressure
For the spikes you cannot predict, the fabric tells you it is under load before it sheds anything. Watch the lag and queue-depth signals.
| Signal | Means |
|---|---|
| Ingest lag p99 rising | Events are queuing — still safe, but trending hot. |
| Queue depth climbing | Producers are outpacing consumers; scale is imminent. |
| Shed rate > 0 | Hard limit hit — low-priority streams are being deferred. Act now. |
When to scale
The autoscaler handles ordinary swings. Reach for a manual pre-warm or a word with us when a planned event will multiply sustained throughput, or when lag stays elevated after the autoscaler has already responded.
Ingest is prioritized, not dropped. Under extreme load the fabric defers low-priority streams before high-severity ones — detection never starves because a noisy debug source spiked. Shed events backfill from the source once headroom returns.
Where to go next
- Healthchecking the fabric — the standing checks.
- Data fabric — how ingest scales.