[CASSANDRA-4650] RangeStreamer should be smarter when picking endpoints for streaming in case of N >=3 in each DC.

• Type: Improvement
• Status: Resolved
• Priority: Low
• Resolution: Fixed
• Fix Version/s: 4.0-alpha1, 4.0
• Labels: streaming
• Assignee: Sankalp Kohli
• Reporter: Sankalp Kohli

Authors

Sankalp Kohli

Reviewers

Marcus Eriksson

• Created: 12/Sep/12 00:01
• Updated: 15/May/20 07:59
• Resolved: 12/May/17 07:17

Description

getRangeFetchMap method in RangeStreamer should pick unique nodes to stream data from when number of replicas in each DC is three or more.
When N>=3 in a DC, there are two options for streaming a range. Consider an example of 4 nodes in one datacenter and replication factor of 3.
If a node goes down, it needs to recover 3 ranges of data. With current code, two nodes could get selected as it orders the node by proximity.
We ideally will want to select 3 nodes for streaming the data. We can do this by selecting unique nodes for each range.

Advantages:
This will increase the performance of bootstrapping a node and will also put less pressure on nodes serving the data.

Note: This does not affect if N < 3 in each DC as then it streams data from only 2 nodes.

Issue Links

• is related to: CASSANDRA-4525 https://issues.apache.org/jira/browse/CASSANDRA-4516
• relates to: CASSANDRA-13576 test failure in bootstrap_test.TestBootstrap.consistent_range_movement_false_with_rf1_should_succeed_test