Skip to content

CONCEPT Cited by 1 source

Exponential TTL

Definition

An exponential TTL (or more generally an age-based TTL ladder) is a cache-expiry strategy where the TTL assigned to a cached value increases monotonically with the age of the underlying data — fresh data gets a short TTL, old data gets a long TTL.

The "exponential" label comes from the canonical Netflix schedule — TTL doubles for each additional minute of data age — but the general idea applies with any monotonically-increasing schedule (linear, piecewise, logarithmic).

Why it exists

In time-series systems subject to late-arriving data, the trustworthiness of a value varies with its age:

  • Recent data (last few minutes) may still change as delayed events trickle in. Confidence is low. Cache too long → wrong chart.
  • Older data (30+ minutes old) is effectively final. Confidence is high. Cache it aggressively; re-querying wastes backend load.

A uniform TTL forces a single trade-off for data with very different confidence levels. An exponential / age-based ladder resolves the forced either/or of the cache TTL staleness dilemma — different buckets get different TTLs based on how much staleness is actually acceptable for that bucket.

Canonical instance: Netflix Druid interval cache

Netflix's Druid cache assigns per-bucket TTLs by age (Source: sources/2026-04-06-netflix-stop-answering-the-same-question-twice-interval-aware-caching-for-druid):

Bucket age TTL
< 2 min 5 s (minimum)
2 min 10 s
3 min 20 s
4 min 40 s
5 min 80 s
n min roughly 5 · 2^(n−1) s
(cap) 1 hour (maximum)
  • Fresh data cycles rapidly → late-arriving corrections picked up quickly
  • Old data lingers → confidence grows with time, cache hit rate climbs for the bulk of the query window

For a 3-hour rolling-window query at 1-minute granularity, this ladder ensures the vast majority of the query is served from cache — only the newest few minutes hit Druid.

Why "exponential" is a good fit

  • The confidence-in-bucket-value function is roughly exponential in age for most event pipelines (late events are Pareto-distributed in arrival delay).
  • Doubling gives the freshest bucket a short TTL + reaches the cap in a small number of steps, which bounds both the staleness risk and the TTL-ladder table size.

Netflix's schedule caps at 1 hour rather than growing indefinitely; the cap is a practical bound rather than a correctness bound.

Contrast with uniform TTL and event-driven invalidation

Strategy When it fits
Uniform TTL Stable data, homogeneous confidence
Age-based exponential TTL Time-series data with late-arriving events; confidence grows with age
Invalidation-based Source can push invalidation messages (CDC, pub/sub)
Write-through Writes are the authoritative update path

Time-series + late-arriving-data is the exact shape where uniform TTL forces a bad choice (long TTL = stale; short TTL = load) and invalidation is impractical (no per-value change events; late arrivals retroactively adjust many buckets).

Seen in

Last updated · 319 distilled / 1,201 read