[
https://issues.apache.org/jira/browse/SOLR-16806?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17732597#comment-17732597
]
Houston Putman commented on SOLR-16806:
---------------------------------------
Including from the PR, since it is very important information:
{quote}This PR now shifts the existing PlacementPlugins to be subclasses of
OrderedNodePlacementPlugin, which gives default implementations of
{{computePlacements()}} and {{{}computeBalancing(){}}}. These subclasses merely
need to implement a {{getWeightedNodes()}} method, which returns a mapping from
{{Node}} to {{{}WeightedNode{}}}. Each plugin makes its own implementation of
{{{}WeightedNode{}}}, because the "weighting" is determined by what that plugin
wants to prioritize when making selections.
The {{WeightedNode}} abstract class has become a bit bloated, but basically it
allows a Node to keep track of a state, and make an overall "weight" for that
node with the given state. (It also has a method to project relevant weight
with a given replica added, that's what used for {{{}computePlacements(){}}}).
I'm sure I need to do a lot more explanation, but in general, the given plugins
should work almost exactly as they do today, with the added benefit of having
{{computeBalancing()}} come for "free". (though I'm sure the implementation of
that shared method can be improved with time)
This OrderedNodePlacementPlugin is just a class that the existing
PlacementPlugins extend, so it is back-compat with custom plugins. We should
mark the class experimental for 9.x, so that we can make improvements to
OrderedNodePlacementPlugin without having to worry about back-compat guarantees
while we are improving these APIs. (e.g., the WeightedNode class may need
additional methods if we want to make improvements to the sorting).
{quote}
> Add a BalanceReplicas API
> -------------------------
>
> Key: SOLR-16806
> URL: https://issues.apache.org/jira/browse/SOLR-16806
> Project: Solr
> Issue Type: Sub-task
> Reporter: Houston Putman
> Priority: Major
> Time Spent: 6h 40m
> Remaining Estimate: 0h
>
> Currently there is no way to move replicas to a new, empty, Solr node,
> without manually choosing the replicas yourself, and moving/recreating the
> replicas. The logic we would need for this is to choose the best replicas to
> move to an empty Solr node, and stop when that Solr node reaches an
> "equilibrium" with the rest of the cluster.
> This logic is very close to merely balancing the existing replicas across a
> given set of nodes. So instead of creating a similar command that solves a
> subset of use-cases, it is probably worth it to take a bit more effort and
> create the general balance-replicas API to solve all balancing use-cases.
> The API would be quite simple. It would be something like {*}PUT
> "/api/cluster/balanceReplicas"{*}, and it would take in a list of nodes:
> {"nodes": [ "node1", "node2" ]}
> If no nodes were provided in the request body, then Solr would default to
> balancing across all data nodes.
> I don't see a need to write a V1 API for this.
> This requires adding method signatures to the PlacementPlugin, but we can
> make them default to a NO-OP, preserving compatibility for existing
> custom-plugins.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
