Blake Eggleston created CASSANDRA-21100:
-------------------------------------------
Summary: CEP-45: Handle rf changes / new DC, etc
Key: CASSANDRA-21100
URL: https://issues.apache.org/jira/browse/CASSANDRA-21100
Project: Apache Cassandra
Issue Type: Improvement
Components: Consistency/Streaming
Reporter: Blake Eggleston
When replicas are added via increasing the RF or adding DCs, operators need to
run repair to populate the new replicas with data.
This will cause problems when dealing with keypaces using mutation tracking.
With the current behavior, new replicas will immediately start participating in
reads. This will cause correctness issues in cases where logs have been
truncated since the new replica will not be aware that it's missing data. If
it's selected as a data nore for a read, it will return incomplete data and not
be aware of it. Even if the entire write set is not truncated, it could cause a
significant amount of data to be synced as part of read reconciliation.
Replicas using mutation tracking that gain ranges as part of rf changes or new
dcs should start in something resembling a pending node during a token move,
where they accept writes, don't participate in reads (actively refusing or
forwarding requests), and use some mechanism (optionally operator initiated) to
stream missing data.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
