[improve][broker] Add reset cursor latency metric

[Technoboy-]

Motivation

Follow-up for investigating slow subscription cursor resets. When reset cursor is slow or fails, brokers currently lack a direct latency metric for the operation, and the binary protocol seek error path can dereference a missing cause while building the error response.

Modifications

• Add pulsar_subscription_reset_cursor_latency_ms summary with topic, subscription, and result labels.
• Record reset cursor latency for both message-id and timestamp reset paths, including early fenced failures.
• Make broker seek error responses null-safe by unwrapping completion exceptions and falling back when the message is missing.
• Recommend enabling managedLedgerReadEntryTimeoutSeconds by setting the broker/standalone/template defaults to 120 seconds and documenting 0 as the disable value.

Verifying this change

• ./gradlew :pulsar-broker-common:compileJava :pulsar-broker:compileJava
• ./gradlew :pulsar-broker:test --tests org.apache.pulsar.broker.service.SubscriptionSeekTest.testSeek --tests org.apache.pulsar.broker.service.SubscriptionSeekTest.testSeekByTimestamp --tests org.apache.pulsar.broker.service.SubscriptionSeekTest.testConcurrentResetCursor

[Copilot]

Pull request overview

This PR improves broker observability and robustness around subscription cursor resets by adding a Prometheus latency metric for reset-cursor operations, hardening the binary protocol seek error response path against null causes, and enabling a managed-ledger read-entry timeout by default to reduce the risk of stuck reads blocking cursor operations.

Changes:

• Added pulsar_subscription_reset_cursor_latency_ms Summary metric (labels: topic, subscription, result) and recorded latency for both timestamp- and position-based reset cursor paths, including early fenced failures.
• Made seek/reset-cursor error responses safer by unwrapping completion exceptions and using a null-safe message fallback.
• Changed managedLedgerReadEntryTimeoutSeconds default to 120 seconds and updated template configs/comments to recommend keeping it enabled (with 0 as the disable value).

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
pulsar-broker/src/main/java/org/apache/pulsar/broker/service/ServerCnx.java	Unwraps completion exceptions in seek error handling and adds null-safe message formatting.
pulsar-broker/src/main/java/org/apache/pulsar/broker/service/persistent/PersistentSubscription.java	Introduces and records a Prometheus Summary metric for reset cursor latency (success/failure).
pulsar-broker-common/src/main/java/org/apache/pulsar/broker/ServiceConfiguration.java	Updates the default managed-ledger read-entry timeout and expands its documentation.
deployment/terraform-ansible/templates/broker.conf	Updates the default managed-ledger read-entry timeout and documents 0 as disable.
conf/standalone.conf	Updates the default managed-ledger read-entry timeout and documents 0 as disable.
conf/broker.conf	Updates the default managed-ledger read-entry timeout and documents 0 as disable.

[void-ptr974]

This metric may leak high-cardinality label children.

RESET_CURSOR_LATENCY.labels(topicName, subName, result) creates a Summary child per topic/subscription/result, and each Summary.Child also registers a DataSketchesSummaryLogger in the static rotation list. I don't see any cleanup when the subscription/topic is closed or deleted, so subscriptions that ever perform reset cursor can remain in the collector and rotation list indefinitely. On brokers with many short-lived topics/subscriptions, this can grow memory usage, periodic rotation cost, and Prometheus series count.

Could we either avoid topic/subscription labels for this static summary, or record this through subscription-scoped stats with lifecycle cleanup? If keeping these labels is required, we should add cleanup on subscription/topic deletion that removes both the collector child and its associated summary logger.