Paulo Motta created CASSANDRA-16138:
---------------------------------------
Summary: Refactor Local Ring Management
Key: CASSANDRA-16138
URL: https://issues.apache.org/jira/browse/CASSANDRA-16138
Project: Cassandra
Issue Type: Improvement
Components: Cluster/Membership, Feature/Virtual Nodes,
Legacy/Distributed Metadata
Reporter: Paulo Motta
Assignee: Paulo Motta
Token ring management is one of the most critical parts of Cassandra, yet one
of the most overlooked. Some of the problems include but are not limited to:
* Complexity (ie. [pending range
calculation|https://github.com/apache/cassandra/blob/8ba163f25a56cb507e621b89b6928c2aef0ecc57/src/java/org/apache/cassandra/locator/TokenMetadata.java#L878])
* Inefficiency (ie. [pending range
calculation|https://github.com/apache/cassandra/blob/8ba163f25a56cb507e621b89b6928c2aef0ecc57/src/java/org/apache/cassandra/locator/TokenMetadata.java#L878],
[AbstractReplicationStrategy.getAddressReplicas|https://github.com/apache/cassandra/blob/33eada06a6dd3529da644377dba180795f522176/src/java/org/apache/cassandra/locator/AbstractReplicationStrategy.java#L233])
* Prone to race conditions (ie.
[here|https://github.com/apache/cassandra/blob/8ba163f25a56cb507e621b89b6928c2aef0ecc57/src/java/org/apache/cassandra/locator/ReplicaLayout.java#L198])
* Poor modularity and consistency (ie. natural replicas computed from
[NetworkTopologyStrategy|https://github.com/apache/cassandra/blob/8ba163f25a56cb507e621b89b6928c2aef0ecc57/src/java/org/apache/cassandra/locator/NetworkTopologyStrategy.java]
and pending replicas computed from
[TokenMetadata|https://github.com/apache/cassandra/blob/8ba163f25a56cb507e621b89b6928c2aef0ecc57/src/java/org/apache/cassandra/locator/TokenMetadata.java#L1271])
* Insufficient testing (due to complexity and poor modularity)
These limitations make it difficult to reliably fix bugs like properly
supporting node replacement with the same IP address (CASSANDRA-12344), add
improvements such as safe ring membership changes, support for networking via
identity instead of IP (CASSANDRA-15823) or add new features such as dynamic
virtual nodes.
This ticket aims at refactoring the ring management sub-module (namely
TokenMetadata and related classes) to address most of its current limitations
in order to support further improvements and new features.
Some of the requirements of the proposed refactoring are:
# Make node-local ring representation fully immutable and snapshottable.
# Add content-based versioning to uniquely identify a ring snapshot throughout
the cluster.
# Make token ring management vnode-centric to support membership operations on
individual tokens and simplify token assignment calculations.
# Primarily identify ring endpoints by node ID to decouple a node’s identity
from its IP address.
# Add a local publish/subscribe mechanism for ring change notifications, so
other modules can subscribe to it and receive the newest snapshot of the ring
after membership changes.
# Add testing framework to verify correctness of ring membership operations.
# Ensure the refactored sub-module does not change current behavior via
comprehensive testing.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
