docs: Add detailed notes on Kafka broker leadership selection process

josix · Dec 20, 2024 · 045c215 · 045c215
1 parent cbe5e18
commit 045c215
Showing 1 changed file with 53 additions and 0 deletions.
diff --git a/random-thoughts/2024-12-21.md b/random-thoughts/2024-12-21.md
@@ -929,3 +929,56 @@ tags:
     - The client code doesn't need to know or decide which mode is being used - it simply makes metadata requests to brokers, and the brokers handle getting the metadata from the appropriate source based on how they were configured.
 
     - This is part of Kafka's evolution to remove the ZooKeeper dependency (KIP-500), making the system simpler to deploy and maintain.
+
+- How brokers decide which broker is the leader for a partition:
+  - Leadership Selection Process:
+    - Initial Leader Selection:
+      - When a topic is created, the Controller assigns leadership
+      - Uses round-robin distribution by default
+      - Considers rack awareness if configured
+      - Ensures even distribution across brokers
+
+    - Leader Election Process:
+
+      ```mermaid
+      sequenceDiagram
+          participant Controller
+          participant ISR as In-Sync Replicas
+          participant Broker
+          
+          Note over Controller: Detect leader failure or<br/>need for new leader
+          
+          Controller->>ISR: Get list of eligible replicas
+          
+          alt Has eligible ISR
+              Controller->>Controller: Select first replica from ISR
+              Controller->>Broker: Notify new leader
+              Broker-->>Controller: Accept leadership
+          else No eligible ISR
+              alt unclean.leader.election.enable=true
+                  Controller->>Controller: Select from all replicas
+              else unclean.leader.election.enable=false
+                  Controller->>Controller: Keep partition offline
+              end
+          end
+          
+          Controller->>Broker: Broadcast metadata update
+      ```
+
+    - Leadership Eligibility:
+      - Must be in ISR (In-Sync Replicas) list
+      - Must be alive and responsive
+      - Must have caught up with previous leader
+      - Configured minimum ISR size must be met
+
+    - Failure Handling:
+      - Controller monitors broker health
+      - Detects leader failures via ZooKeeper/KRaft
+      - Initiates new leader election if needed
+      - Updates cluster metadata after changes
+
+    - Configuration Factors:
+      - unclean.leader.election.enable
+      - min.insync.replicas
+      - replica.lag.time.max.ms
+      - leader.imbalance.check.interval.seconds