Introduction

Point of leader election is that (1) only one leader election; and (2) everyone knows who the leader is

Setup

• Any process can call for an election, but you cannot call another election until the one you previously called is finished
• Rounds have ballot_ids , highest ballot_id is elected

⭕ Ring-Based

• Nodes arranged in virtual ring
• Any node an initiate election, pass your id along the ring
  • Multiple nodes can initiate election at any point, but each node suppresses lower id
• Each node forwards max(rcv_id, my_id)
• When the highest id sees itself, it knows that its id went around the whole ring, so it sends it again to tell everyone that it is the leader
• Then everyone elects that id as its leader

Worst-case: Initiator is one after the highest id → 3(N-1) messages and completion time

Best-case: Initiator is the highest id → 2N messages and completion time

⚠️ Problem: if elected leader dies but their elected message lives on, we are in trouble!!

✅ no real solution other than using a failure detector and hoping it catches the failure and initiates a new election

👺 Bully Algorithm

Similar to ring but smarter with re-election upon old leader dying

# once leader dies 

p = process that detected the failure 

if p.id == max(ids_in_my_membership_list)
	**elect self as leader and tell everyone** 
	everyone elects p 

else
	initiate new election -> send ONLY to processes with ids > p.id # avoid wasted effort to send to all peers 
	
	if no messages within timeout: 
		**elect self as leader and tell everyone** 
		
	else if answer from higher id rcved:
		wait for election message from that higher id processs

Is this safe?

• failures during election are caught because even if leader dies but its message lives on, every peer waits for heartbeats from that leader periodically and if its doesn’t receive a message within timeout → it assumes that the leader is dead and restarts an election
  • How to set timeout? based on worst case time to complete election
    • In these 5 steps (on the right), each takes 1 message transmission time
    • Steps 2 and 3 and be parallelized
    • So total = 4 message transmission time = 2 RTT
• as long as failures stop at some point, a leader will be elected, this algo is safe!

Worst-case: lowest id detects failure → N^2 messages