[
https://issues.apache.org/jira/browse/CASSANDRA-762?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jonathan Ellis updated CASSANDRA-762:
-------------------------------------
Component/s: Core
Fix Version/s: 0.6
As I said in IRC, a loadbalancing node doesn't compute its destination token
until after all its data has been transferred off. (We *should* add a sleep
until load is rebroadcast, but that will only affect which node gets picked to
pull half the range from.) So if there is a bug here it is in the code, not
the algorithm.
Can you reproduce this?
> Load balancing does not account for the load of the moving node
> ---------------------------------------------------------------
>
> Key: CASSANDRA-762
> URL: https://issues.apache.org/jira/browse/CASSANDRA-762
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Affects Versions: 0.5
> Reporter: Stu Hood
> Priority: Minor
> Fix For: 0.6
>
>
> Given a node A (with load 10 gb) and a node B (with load 20 gb), running the
> loadbalance command against node A will:
> 1. Remove node A from the ring
> * Recalculates pending ranges so that node B is responsible for the entire
> ring
> 2. Pick the most loaded node
> * node B is still reporting 20 gb load, because that is all it has locally
> 3. Choose a token that divides the range of the most loaded node in half
> Since the token calculation doesn't take into account the load that node B is
> 'inheriting' from node A, the token will divide node B's load in half and
> swap the loads. Instead, the token calculation needs to pretend that B has
> already inherited the 10 gb from node A, for a total of 30 gb. The token that
> should be chosen falls at 15 gb of the total load, or 5 gb into node B's load.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
