Archivedistributed-systems
Leader Leases Turn Consensus Into a Cache
A time-bounded lease lets a leader serve reads without a round trip — as long as clocks behave.
- Published
- Reading time
- 1 min read
A leader lease is a promise from a quorum: "we will not elect a new leader before time T". The leader can then serve linearizable reads locally until T, avoiding a quorum round trip per read.
The safety condition
lease_end = start + duration
duration < min_clock_drift_budget
The lease must expire before any new leader could be elected, measured with bounded clock skew. If clocks drift more than assumed, two leaders can serve reads.
Variants
- Leader lease: the leader holds it; followers reject writes to others.
- Follower lease: followers promise not to vote for a duration, which lets the leader skip heartbeats.
- Read index: no lease; the leader confirms leadership with a quorum before serving. Slower, safer.
When to use it
Read-heavy systems with stable clocks and short leases (hundreds of milliseconds to a few seconds). Not for geo-distributed clusters with unbounded NTP error, and not for systems where a stale read is unacceptable.
Takeaway: a lease is a clock-dependent optimization. Bound your clock error, keep leases short, and have a read-index fallback.