Lecture 4: Primary-Backup Replication

Fault Tolerance

Ensures system continues operation despite component failures.

Replication

Maintains multiple copies for fault tolerance, handling fail-stop failures (e.g., power loss, network disconnection).
Does not handle software bugs, hardware defects, or correlated failures (e.g., shared defects, natural disasters).
Value of replication depends on the cost of failure.

Replication Approaches

State Transfer: Primary sends full state to backup.
Replicated State Machine: Sends only events that affect state; preferred due to smaller data size.

Key Questions in Replication

What to replicate?
Synchronization level between primary and backup.
Client switchover method when primary fails.
Handling anomalies during switchover.
Creating new replica on failure.

VMware FT (Fault Tolerance)

Replicates full machine state (memory, registers) on separate physical machines.
Clients connect to primary; backup activated if primary stops logging entries.

VMware FT Challenges & Solutions

Non-deterministic events: Logged and synchronized with special CPU support.
Backup lagging primary: Backup waits for event buffer before proceeding.
Direct Memory Access (DMA): Data copied to private memory, emulating NIC interrupts.
Output rule: Primary waits for backup acknowledgment before sending output.
Synchronous waits: Output rule impacts performance, especially over distances.
Split-brain problem: Uses an external test-and-set service as an arbitrator.