- 10月 12, 2022
-
-
由 Sushant Mahajan 创作于
* TRUST -> POOL * POLICY -> FILTER * SUBJECT_CLAIM -> IDENTITY_CLAIM
-
由 Manikumar Reddy 创作于
Currently ConfluentAuditLogProvider instance is not getting started in Kraft controller. This is due to error in authorizer startup. We are not configuring httpServerBinder, due to this we are getting NPE during authorizer startup. Due to this we are not emitting the audit events to audit event pipeline.
-
由 Sushant Mahajan 创作于
-
由 Niket 创作于
KAFKA-14275; KRaft Controllers should crash after failing to apply any metadata record (#12709) (#7711) Make all faults in metadata processing on standby controllers be fatal. This is the same behavior-wise as the active controller. This prevents a standby controller from eventually becoming active with incomplete state. Reviewers: Colin Patrick McCabe <cmccabe@apache.org>, Jason Gustafson <jason@confluent.io>
-
由 Bob Barrett 创作于
* Only prepend LKC if hostname starts with '-' * Add tests
-
由 Aadithya Chandra 创作于
MINOR: Flaky test fix. Wait for broker to be kicked out of the metadata cache after killing it. (#7684)
- 10月 11, 2022
-
-
由 Sanjana Kaundinya 创作于
-
由 Jason Gustafson 创作于
Fixes a lingering issue from the incompatible changes to confluent.metadata.version in the cherry pick of KAFKA-14177. Controllers which have been updated will use a finalized confluent.metadata.version which matches the new range (101+). Brokers which are on the old version will register with support for the range 1-7. Currently this fails on the controller since 101+ is outside the supported range. The patch fixes the issue by verifying the range against the apache feature level of the current metadata version when the old confluent range is advertised by a broker. Reviewers: Colin P. McCabe <cmccabe@apache.org>, José Armando García Sancio <jsancio@jsancio@gmail.com>
-
- 10月 10, 2022
-
-
由 Colin Patrick McCabe 创作于
Add builders for LocalLogManagerTestEnv and QuorumControllerTestEnv, since the constructor overloads were starting to get unwieldy. Make indentation more consistent in QuorumControllerTest. Take advantage of the fact that you can initialize multiple resources in a Java try-with-resources block to avoid excessive indentation in a few cases. Reviewers: José Armando García Sancio <jsancio@apache.org> Conflicts: Fix some conflicts in QuorumControllerTest related to ce-only features.
-
The error message reported when advertised.listeners is used in controller only is confusing. When the KRaft server is configured to controller only the following must be true: 1. `advertised.listeners` is not set 2. `listeners` contains a listener for every name in `controller.listener.names` 3. `controller.listener.names` contains a name for every listener in `listeners` Reviewers: Jason Gustafson <jason@confluent.io>, Igor Soarez <i@soarez.me> Conflicts: In KafkaConfigTest, ce-kafka was using Properties#put, whereas AK was using Properties#setProperty. This PR resolves that divergence by simply using setProperty everywhere. The use of Properties#put is "strongly discouraged" according to the Java documentation because it can compromise type safety. Also in KafkaConfigTest, ce-kafka was explicitly disabling cluster linking since that used to be incompatible with KRaft. It isn't any more, so this PR drops that divergence as well.
- 10月 09, 2022
-
-
由 Aadithya Chandra 创作于
A replica is eligible to be in the ISR, if all of the following are true: 1. The replica is active (i.e. it is not fenced, not in controlled shutdown nor offline). 2. If the replica is an observer, one of the following must be true: a. The leader is an observer. b. The observer replica is eligible for promotion. 3. If the replica is degraded, one of the following must be true: a. The replica is an observer. b. ExcludeDegradedBrokers is false. c. The leader is a degraded broker.
- 10月 08, 2022
-
-
由 David Mao 创作于
Reset the state of mocked class variables
-
由 Jeff Kim 创作于
-
由 David Mao 创作于
-
由 Aadithya Chandra 创作于
-
由 andymg3 创作于
KGLOBAL-2033: Consolidate on one reason code instead of having both a reason code and an exception in LinkStateInfo (#7672)
-
由 Jeff Kim 创作于
-
由 David Mao 创作于
Significant conflicts: Reverted KAFKA-14198 follow-up commit e43e59fc RetryWithToleranceOperatorTest, RetryWithToleranceOperator - reconciled with upstream changes connect_distributed_test.py - Reconciled with downstream * confluentinc-kafka/master: (119 commits) KAFKA-14209 : Rewrite self joins to use single state store 2/3 (#12644) MINOR: Add initial property tests for StandardAuthorizer (#12703) MINOR; Use 3.3.1 release for system test (#12714) MINOR: Fix typo in selector documentation (#12710) KAFKA-14270: Fix generated Kafka Streams version file name (#12700) KAFKA-14247; Define event handler interface and events (#12663) MINOR: Fix delegation token system test (#12693) Fix ByteBufferSerializer#serialize(String, ByteBuffer) not roundtrip input with ByteBufferDeserializer#deserialize(String, byte[]) (#12704) KAFKA-14133: Replace EasyMock with Mockito in streams tests (#12527) KAFKA-14248; Fix flaky test PlaintextAdminIntegrationTest.testCreateTopicsReturnsConfigs (#12669) ...
-
由 Aniket Shrimal 创作于
-
由 Jason Gustafson 创作于
Auto balancing of leaders by the controller should not generate change records unless there is a leader change. This does not work for linked topics because `PartitionChangeBuilder.build()` will always set the linked leader epoch and the link state fields even when there is no change. When there are many linked partitions (more than 1000) in a cluster, then this bug can cause a livelock where the controller continuously writes partition change records for linked partitions without any actual changes. Every iteration of the logic will reach the batch limit of 1000 and trigger re-execution. But since no progress is made, the controller gets stuck. This patch fixes the logic in `build()` so that it properly checks for changes to the linked state fields. One note: the link state field requires special treatment since there is no sentinel indicating no change. We always have to s et the field in the change record and we always have to compare against the previous value to decide if there is a change. Finally, this patch increases the retry logic when leader balancing cannot be completed in the current round from 10ms to 100ms. This just loosens the loop in case we encounter another condition like this in the future. Reviewers: Akhilesh C <akhileshchg@users.noreply.github.com>, José Armando García Sancio <jsancio@users.noreply.github.com>
- 10月 07, 2022