ContextQMD
Back to Pulsar

PIP-422: Support global topic-level policy: replicated clusters and new API to delete topic-level policies

pip/pip-422.md

4.2.15.5 KB
Original Source

PIP-422: Support global topic-level policy: replicated clusters and new API to delete topic-level policies

Background knowledge

Users have two choices when configuring the metadata store when using the feature Geo-Replication: shared or not. All clusters share data that are stored in the global config metadata store if users choose to use a shared metadata store. By the way, the share config metadata store is often built across multiple zones/regions to ensure fault tolerance.

Since the data that is stored in the shared metadata store contains the following pulsar resources: tenants, namespaces, partitioned topics, the Admin API can not be used to delete these pulsar resources for a single cluster. Pulsar has a feature that is used to delete namespace-level resources for a specific cluster when using a shared metadata store, it works as follows:

  • The original namespace-level policy that named replicated clusters contains two clusters c1 and c2.
  • The topics under the cluster c1 will be deleted automatically when users remove c1 from the namespace-level policy.

The feature above is very useful in the following scenarios:

  • Transfer a namespace from an old cluster to a new cluster: at last, the namespace should be deleted from the old cluster.
  • Delete topics: since Pulsar does not allow deleting topics when enabled Geo-Replication, users should disable replication first.

Motivation

When using a shared metadata store and enabling Geo-Replication, there is no mechanism for the following scenarios:

  • Transfer a partitioned topic from an old cluster to a new cluster.
  • Delete a partitioned topic from a specified cluster.

Goals & Designs

Add two new APIs

  • pulsar-admin topicPolicies --set-replication-cluster -c <clusters> --g <topic>
    • Pulsar only supports setting a local topic-level replicated clusters, and does not support global policies so far.
  • pulsar-admin topicPolicies delete <topic>
    • To delete topic level policies, including global and local policies, it can be called even if the topic has been deleted.

Old Apis

  • Prohibits users from using local policies to delete the local cluster.

The new APIs will be used for the following scenarios:

A. Delete a partitioned topic under all clusters, if enabled Geo-Replication with a shared config metadata store.

  • Disable global topic level replicated clusters, in other words, set replicated clusters with a single list.
    • The clusters that are not contained in the replicated clusters will delete the sub-topics automatically, such as the namespace-level implementation.
    • The schemas and local topic policies of the topic will be deleted after the latest sub-topic is deleted.
  • Delete the partitioned topic since we have disabled Geo-Replication.
  • Manually delete global topic-level policies by the new API pulsar-admin topicPolicies delete <topic> for other clusters that are not contained in replicated clusters.
  • Note: If there is any active producers or consumers connected to the cluster which been removed from the replicated cluster list, they will be rejected from reconnecting because the connected cluster is not listed in replicated clusters.

B. Delete a partitioned topic under a specified cluster.

  • Support setting up a global topic-level replicated clusters policy, which does not contain the cluster that you want to remove.
    • Broker will delete the subtopics automatically.
    • The schemas and local topic policies of the topic will be deleted after the latest sub-topic is deleted.
  • Note:
    • The global topic-level policy will not be removed, to avoid the namespace-level replicated cluster from taking effect.
    • Once you want to delete the topic under all clusters, please use Solution A to remove the global topic-level policy left.

Question: Why not add a mechanism that brokers clear global topic-level policies so that the related topic is deleted regularly, which avoids calling pulsar-admin topicPolicies delete <topic>?

Answer: When users quickly delete and create topics, it will result in the previous policies not being deleted by the scheduled cleaning mechanism, thereby causing confusion.

Public API

new Pulsar Admin API

  • pulsar-admin topicPolicies --set-replication-cluster <clusters> --global <topic>
  • pulsar-admin topicPolicies delete <topic>
  • The API pulsar-admin topics --set-replication-cluster <clusters> <topic> never supports to delete the local cluster anymore.

Binary protocol

Nothing.

Configuration

Nothing.

CLI

Nothing.

Metrics

Nothing.

Monitoring

Nothing.

Security Considerations

Nothing.

Backward & Forward Compatibility

Nothing.

Upgrade

Nothing.

Downgrade / Rollback

Nothing.

Pulsar Geo-Replication Upgrade & Downgrade/Rollback Considerations

Nothing.

Alternatives

Passed solutions:

  1. Disable namespace-level Geo-Replication then delete the partitioned topic: this will trigger a deletion for other topics, which is not expected.
  2. Disable topic-level Geo-Replication under each cluster, and delete the partitioned topic by Admin API: it will delete the shared partitioned topic metadata, which was also deleted from another cluster.
  3. Disable topic-level Geo-Replication under each cluster, and delete the sub-topics manually by Admin API: this is the best solution without the current PIP, the defect is that the schema data and topic-level policies are still there.

General Notes

Links