THIS IS A TEST INSTANCE. ALL YOUR CHANGES WILL BE LOST!!!!
Status
Current state: Under Discussion
Discussion thread: TBD
JIRA: KAFKA-13587
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
If none of the in-sync replicas are alive, the controller allows the user to elect a replica that was not a part of the in-sync replica set using the unclean leader election strategy. Since this new leader replica was not a part of the last known in-sync replica set, there is the possibility for data loss by deleting log records committed by the previous leader(s). In addition, the lost of these records can cause some inconsistency with other parts of the system like the transaction coordinators and the group coordinators. If the controller is able to communicate to the new topic partition leader that it was elected using unclean leader election, the new topic partition leader can coordinate this recovery and mitigate these inconsistencies.
Proposed Changes
This KIP proposes extending the communication between the controller and brokers to include enough information so that topic partition leaders know if they were elected because of an unclean leader election. This feature needs to support both the traditional ZooKeeper controller and the more recent KRaft controller. The messages sent by the controller to the broker during leader election will be extended to include if the leader was elected using unclean leader election. It is important to note that when the controller performs unclean leader election the ISR size is 1 and the "is unclean" field will be true. In such cases the topic partition leader will perform any necessary recovery steps.
The brokers will communicate to the controller when it has finished recovering from unclean leader election. An important invariant that the partition leader must satisfied is that the ISR size is 1 while the "is unclean" field is true. This means that the leader will not allow followers to join the ISR until it has recovered from the unclean leader election.
One of the ways Kafka is going to use this feature is to abort all pending transaction when recovering from unclean leader election. See for more details.
ZooKeeper Controller
The controller will use ZooKeeper to persist the decision that it elected a leader through the unclean leader election strategy. This decision will get propagated to all of the affected brokers through the LeaderAndIsr request. Topic partition leaders will inform the controller that they have recovered from the unclean leader election using the AlterIsr request.
KRaft Controller
The controller will persist the decision that it elected a leader through the unclean leader election strategy in the cluster metadata internal partition using the PartitionRecord or PartitionChangeRecord. These records will get replicated to all of the brokers using the cluster metadata topic partition. Topic partition leaders will inform the controller that they have recovered from unclean leader election using the AlterIsr request.
Changed Public Interfaces
ZooKeeper State
The state of the partition stored in /brokers/topics/<topic>/partitions/<partition>/state as JSON will be extended to add a new property named is_unclean to have the following schema:
{
"version": 1,
"leader": NUMBER,
"leader_epoch": NUMBER,
"controller_epoch": NUMBER,
"isr" ARRAY of NUMBERS,
"is_unclean": BOOLEAN // New property
}
The property is_unclean indicates if the leader was elected using the unclean leader election strategy. The default value for this property, in case it is missing, is false .
LeaderAndIsr RPC
Request Schema
Add a property to indicate to the leader of the topic partition that it was elected using unclean leader election.
{
"apiKey": 4,
"type": "request",
"listeners": ["zkBroker"],
"name": "LeaderAndIsrRequest",
// Version 1 adds IsNew.
//
// Version 2 adds broker epoch and reorganizes the partitions by topic.
//
// Version 3 adds AddingReplicas and RemovingReplicas.
//
// Version 4 is the first flexible version.
//
// Version 5 adds Topic ID and Type to the TopicStates, as described in KIP-516.
//
// Version 6 adds IsUnclean as describe in KIP-704.
"validVersions": "0-6",
"flexibleVersions": "4+",
"fields": [
{ "name": "ControllerId", "type": "int32", "versions": "0+", "entityType": "brokerId",
"about": "The current controller ID." },
{ "name": "ControllerEpoch", "type": "int32", "versions": "0+",
"about": "The current controller epoch." },
{ "name": "BrokerEpoch", "type": "int64", "versions": "2+", "ignorable": true, "default": "-1",
"about": "The current broker epoch." },
{ "name": "Type", "type": "int8", "versions": "5+",
"about": "The type that indicates whether all topics are included in the request"},
{ "name": "UngroupedPartitionStates", "type": "[]LeaderAndIsrPartitionState", "versions": "0-1",
"about": "The state of each partition, in a v0 or v1 message." },
// In v0 or v1 requests, each partition is listed alongside its topic name.
// In v2+ requests, partitions are organized by topic, so that each topic name
// only needs to be listed once.
{ "name": "TopicStates", "type": "[]LeaderAndIsrTopicState", "versions": "2+",
"about": "Each topic.", "fields": [
{ "name": "TopicName", "type": "string", "versions": "2+", "entityType": "topicName",
"about": "The topic name." },
{ "name": "TopicId", "type": "uuid", "versions": "5+", "ignorable": true,
"about": "The unique topic ID." },
{ "name": "PartitionStates", "type": "[]LeaderAndIsrPartitionState", "versions": "2+",
"about": "The state of each partition" }
]},
{ "name": "LiveLeaders", "type": "[]LeaderAndIsrLiveLeader", "versions": "0+",
"about": "The current live leaders.", "fields": [
{ "name": "BrokerId", "type": "int32", "versions": "0+", "entityType": "brokerId",
"about": "The leader's broker ID." },
{ "name": "HostName", "type": "string", "versions": "0+",
"about": "The leader's hostname." },
{ "name": "Port", "type": "int32", "versions": "0+",
"about": "The leader's port." }
]}
],
"commonStructs": [
{ "name": "LeaderAndIsrPartitionState", "versions": "0+", "fields": [
{ "name": "TopicName", "type": "string", "versions": "0-1", "entityType": "topicName", "ignorable": true,
"about": "The topic name. This is only present in v0 or v1." },
{ "name": "PartitionIndex", "type": "int32", "versions": "0+",
"about": "The partition index." },
{ "name": "ControllerEpoch", "type": "int32", "versions": "0+",
"about": "The controller epoch." },
{ "name": "Leader", "type": "int32", "versions": "0+", "entityType": "brokerId",
"about": "The broker ID of the leader." },
{ "name": "LeaderEpoch", "type": "int32", "versions": "0+",
"about": "The leader epoch." },
{ "name": "Isr", "type": "[]int32", "versions": "0+", "entityType": "brokerId",
"about": "The in-sync replica IDs." },
{ "name": "ZkVersion", "type": "int32", "versions": "0+",
"about": "The ZooKeeper version." },
{ "name": "Replicas", "type": "[]int32", "versions": "0+", "entityType": "brokerId",
"about": "The replica IDs." },
{ "name": "AddingReplicas", "type": "[]int32", "versions": "3+", "ignorable": true, "entityType": "brokerId",
"about": "The replica IDs that we are adding this partition to, or null if no replicas are being added." },
{ "name": "RemovingReplicas", "type": "[]int32", "versions": "3+", "ignorable": true, "entityType": "brokerId",
"about": "The replica IDs that we are removing this partition from, or null if no replicas are being removed." },
{ "name": "IsNew", "type": "bool", "versions": "1+", "default": "false", "ignorable": true,
"about": "Whether the replica should have existed on the broker or not." },
// -------- Properties added by this KIP ---------
{ "name": "IsUnclean", "type": "boolean", "versions": "6+", "default": "false", "ignorable": true,
"about": "Whether the leader was elected using the unclean leader election strategy." }
]}
]
}
Response Schema
No changes required to the response.
AlterIsr RPC
Request Schema
The schema will include a new property for the leader to indicate to the controller when it finished handling unclean leader election.
{
"apiKey": 56,
"type": "request",
"listeners": ["zkBroker", "controller"],
"name": "AlterIsrRequest",
"validVersions": "0-1",
"flexibleVersions": "0+",
"fields": [
{ "name": "BrokerId", "type": "int32", "versions": "0+", "entityType": "brokerId",
"about": "The ID of the requesting broker" },
{ "name": "BrokerEpoch", "type": "int64", "versions": "0+", "default": "-1",
"about": "The epoch of the requesting broker" },
{ "name": "Topics", "type": "[]TopicData", "versions": "0+", "fields": [
{ "name": "Name", "type": "string", "versions": "0+", "entityType": "topicName",
"about": "The name of the topic to alter ISRs for" },
{ "name": "Partitions", "type": "[]PartitionData", "versions": "0+", "fields": [
{ "name": "PartitionIndex", "type": "int32", "versions": "0+",
"about": "The partition index" },
{ "name": "LeaderEpoch", "type": "int32", "versions": "0+",
"about": "The leader epoch of this partition" },
{ "name": "NewIsr", "type": "[]int32", "versions": "0+", "entityType": "brokerId",
"about": "The ISR for this partition"},
// ----- Start of properties added by this KIP -----
{ "name": "NewIsUnclean", "type": "boolean", "versions": "1+", "default": "false", "ignoreable": true,
"about": "Whether the leader is recovering from unclean leader election." },
// ----- End of properties added by this KIP -----
{ "name": "CurrentIsrVersion", "type": "int32", "versions": "0+",
"about": "The expected version of ISR which is being updated"}
]}
]}
]
}
Response Schema
Add a property to indicate to the leader of the topic partition that it was elected using unclean leader election.
{
"apiKey": 56,
"type": "response",
"name": "AlterIsrResponse",
"validVersions": "0-1",
"flexibleVersions": "0+",
"fields": [
{ "name": "ThrottleTimeMs", "type": "int32", "versions": "0+",
"about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
{ "name": "ErrorCode", "type": "int16", "versions": "0+",
"about": "The top level response error code" },
{ "name": "Topics", "type": "[]TopicData", "versions": "0+", "fields": [
{ "name": "Name", "type": "string", "versions": "0+", "entityType": "topicName",
"about": "The name of the topic" },
{ "name": "Partitions", "type": "[]PartitionData", "versions": "0+", "fields": [
{ "name": "PartitionIndex", "type": "int32", "versions": "0+",
"about": "The partition index" },
{ "name": "ErrorCode", "type": "int16", "versions": "0+",
"about": "The partition level error code" },
{ "name": "LeaderId", "type": "int32", "versions": "0+", "entityType": "brokerId",
"about": "The broker ID of the leader." },
{ "name": "LeaderEpoch", "type": "int32", "versions": "0+",
"about": "The leader epoch." },
{ "name": "Isr", "type": "[]int32", "versions": "0+", "entityType": "brokerId",
"about": "The in-sync replica IDs." },
// ----- Start of properties added by this KIP -----
{ "name": "IsUnclean", "type": "boolean", "versions": "1+", "default": "false", "ignoreable": true,
"about": "Whether the assigned leader that was elected using the unclean leader election strategy." },
// ----- End of properties added by this KIP -----
{ "name": "CurrentIsrVersion", "type": "int32", "versions": "0+",
"about": "The current ISR version." }
]}
]}
]
}
Cluster Metadata Records
PartitionRecord
Add a property to indicate to the leader of the topic partition that it was elected using unclean leader election.
{
"apiKey": 3,
"type": "metadata",
"name": "PartitionRecord",
"validVersions": "0",
"flexibleVersions": "0+",
"fields": [
{ "name": "PartitionId", "type": "int32", "versions": "0+", "default": "-1",
"about": "The partition id." },
{ "name": "TopicId", "type": "uuid", "versions": "0+",
"about": "The unique ID of this topic." },
{ "name": "Replicas", "type": "[]int32", "versions": "0+", "entityType": "brokerId",
"about": "The replicas of this partition, sorted by preferred order." },
{ "name": "Isr", "type": "[]int32", "versions": "0+",
"about": "The in-sync replicas of this partition" },
{ "name": "RemovingReplicas", "type": "[]int32", "versions": "0+", "entityType": "brokerId",
"about": "The replicas that we are in the process of removing." },
{ "name": "AddingReplicas", "type": "[]int32", "versions": "0+", "entityType": "brokerId",
"about": "The replicas that we are in the process of adding." },
{ "name": "Leader", "type": "int32", "versions": "0+", "default": "-1", "entityType": "brokerId",
"about": "The lead replica, or -1 if there is no leader." },
// ----- Start of properties added by this KIP -----
{ "name": "IsUnclean", "type": "boolean", "default": "false", "versions": "0+", "taggedVersions": "0+", "tag": 0,
"about": "Whether the assigned leader was elected using the unclean leader election strategy." },
// ----- End of properties added by this KIP -----
{ "name": "LeaderEpoch", "type": "int32", "versions": "0+", "default": "-1",
"about": "The epoch of the partition leader." },
{ "name": "PartitionEpoch", "type": "int32", "versions": "0+", "default": "-1",
"about": "An epoch that gets incremented each time we change anything in the partition." }
]
}
PartitionChangeRecord
Add a property to indicate to the leader of the topic partition that it was elected using unclean leader election.
{
"apiKey": 5,
"type": "metadata",
"name": "PartitionChangeRecord",
"validVersions": "0",
"flexibleVersions": "0+",
"fields": [
{ "name": "PartitionId", "type": "int32", "versions": "0+", "default": "-1",
"about": "The partition id." },
{ "name": "TopicId", "type": "uuid", "versions": "0+",
"about": "The unique ID of this topic." },
{ "name": "Isr", "type": "[]int32", "default": "null", "entityType": "brokerId",
"versions": "0+", "nullableVersions": "0+", "taggedVersions": "0+", "tag": 0,
"about": "null if the ISR didn't change; the new in-sync replicas otherwise." },
{ "name": "Leader", "type": "int32", "default": "-2", "entityType": "brokerId",
"versions": "0+", "taggedVersions": "0+", "tag": 1,
"about": "-1 if there is now no leader; -2 if the leader didn't change; the new leader otherwise." },
{ "name": "Replicas", "type": "[]int32", "default": "null", "entityType": "brokerId",
"versions": "0+", "nullableVersions": "0+", "taggedVersions": "0+", "tag": 2,
"about": "null if the replicas didn't change; the new replicas otherwise." },
{ "name": "RemovingReplicas", "type": "[]int32", "default": "null", "entityType": "brokerId",
"versions": "0+", "nullableVersions": "0+", "taggedVersions": "0+", "tag": 3,
"about": "null if the removing replicas didn't change; the new removing replicas otherwise." },
{ "name": "AddingReplicas", "type": "[]int32", "default": "null", "entityType": "brokerId",
"versions": "0+", "nullableVersions": "0+", "taggedVersions": "0+", "tag": 4,
"about": "null if the adding replicas didn't change; the new adding replicas otherwise." },
// ----- Properties added by this KIP -----
{ "name": "IsUnclean", "type": "boolean", "default": "-1", "versions": "0+", "taggedVersions": "0+", "tag": 5,
"about": "-1 if it didn't change; 0 if the leader was elected from the ISR; 1 if the leader that was elected using unclean leader election." }
]
}
Compatibility, Deprecation, and Migration Plan
At a high-level this change is backwards compatible because the default value for all of the "is unclean" properties in the protocol is "false". When thinking about backward compatibility it is important to note that if the "is unclean" field is true then the ISR is guarantee to have a size of 1. The topic partition leader will not increase the ISR until it has recovered from the unclean leader election and has set the "is unclean" field to false.
Controller
If the ZooKeeper controller supports the feature then when performing an unclean leader election it will write "true" in ZK and it will set "false" in the LeaderAndIsr request. If the broker doesn't support this feature, the "is unclean" field in the LeaderAndIsr request will be ignored and behave as it currently does. When the broker sends the AlterIsr request to the controller, the controller will interpret the "is unclean" fields as "false" because the default value is false. A similar logic applies to the KRaft controller.
Rejected Alternatives
An alternative solution is to store the leader epoch when the unclean leader election was performed instead of storing a boolean in the "is unclean" field. The topic partition would perform unclean recovery when the unclean leader epoch is equal to the current leader epoch. One issue with this solution is that the controller changes the leader epoch when the leader goes offline. This means that the controller would have to also reset the unclean leader epoch when the leader goes offline.