Optimize ImmutableOpenMap.Builder

[joegallo]

Related to #77466 (also kindof related to #82708, in that that's what got me thinking about this). Also shoutout to @jakelandis and also @original-brownbear for daydreaming about this with me earlier today.

Consider Metadata.Builder, it builds four ImmutableOpenMap.Builders from four ImmutableOpenMaps:

elasticsearch/server/src/main/java/org/elasticsearch/cluster/metadata/Metadata.java

Lines 1186 to 1189 in 21a5a55

	this.indices = ImmutableOpenMap.builder(metadata.indices);
	this.aliasedIndices = ImmutableOpenMap.builder(metadata.aliasedIndices);
	this.templates = ImmutableOpenMap.builder(metadata.templates);
	this.customs = ImmutableOpenMap.builder(metadata.customs);

In the current code (before this branch), each of those builder(someMap) invocations results in an ObjectObjectHashMap#clone() behind the scenes. Then, later, those ObjectObjectHashMap clones are used to construct new ImmutableOpenMaps in Metadata.Builder#build(). None of that is predicated on actually touching those maps, so Metadata.builder(metadata).version(2).build() will clone and copy four maps.

This PR changes that in two ways. First, it makes a no-op trip of some map through an ImmutableOpenMap.Builder return the exact same reference, so myMap == ImmutableOpenMap.builder(myMap).build(). Second, it defers cloneing of a passed-in map until the point where a mutable map reference is needed for either a read or a write, so the runtime cost in memory and cpu of a no-op trip through a builder is essentially zero.

Given the way that we frequently use ImmutableOpenMap.Builders in other Builders to refer to maps that often do not change (see grep output that follows), this should make many cluster state updates quite a bit cheaper.

joegallo@galactic:~/Code/elastic/elasticsearch/server/src/main/java/org/elasticsearch/cluster $ git grep 'ImmutableOpenMap.builder(' | grep -v 'ImmutableOpenMap.builder()'
ClusterState.java:            this.customs = ImmutableOpenMap.builder(state.customs());
DiffableUtils.java:            ImmutableOpenMap.Builder<K, T> builder = ImmutableOpenMap.builder(map);
RestoreInProgress.java:        final ImmutableOpenMap.Builder<String, Entry> entriesBuilder = ImmutableOpenMap.builder(count);
SnapshotsInProgress.java:                final ImmutableOpenMap.Builder<RepositoryShardId, ShardSnapshotStatus> byRepoShardIdBuilder = ImmutableOpenMap.builder(
block/ClusterBlocks.java:            ImmutableOpenMap.Builder<String, Set<ClusterBlock>> indicesBuilder = ImmutableOpenMap.builder(indices.size());
metadata/IndexMetadata.java:            this.aliases = ImmutableOpenMap.builder(indexMetadata.aliases);
metadata/IndexMetadata.java:            this.customMetadata = ImmutableOpenMap.builder(indexMetadata.customData);
metadata/IndexMetadata.java:            this.rolloverInfos = ImmutableOpenMap.builder(indexMetadata.rolloverInfos);
metadata/IndexTemplateMetadata.java:            mappings = ImmutableOpenMap.builder(indexTemplateMetadata.mappings);
metadata/IndexTemplateMetadata.java:            aliases = ImmutableOpenMap.builder(indexTemplateMetadata.aliases());
metadata/Metadata.java:            this.indices = ImmutableOpenMap.builder(metadata.indices);
metadata/Metadata.java:            this.aliasedIndices = ImmutableOpenMap.builder(metadata.aliasedIndices);
metadata/Metadata.java:            this.templates = ImmutableOpenMap.builder(metadata.templates);
metadata/Metadata.java:            this.customs = ImmutableOpenMap.builder(metadata.customs);
metadata/MetadataDeleteIndexService.java:            ImmutableOpenMap.Builder<String, ClusterState.Custom> builder = ImmutableOpenMap.builder(customs);
node/DiscoveryNodes.java:        ImmutableOpenMap.Builder<String, DiscoveryNode> nodes = ImmutableOpenMap.builder(dataNodes);
node/DiscoveryNodes.java:        ImmutableOpenMap.Builder<String, DiscoveryNode> nodes = ImmutableOpenMap.builder(this.nodes);
routing/allocation/AllocationService.java:                ImmutableOpenMap.Builder<String, ClusterState.Custom> customsBuilder = ImmutableOpenMap.builder(allocation.getCustoms());

For example, in a very ordinary ILM cluster state update, we change the customMetadata on an IndexMetadata but not the aliases or the rolloverInfos, and then we put that into a new Metadata so we'd be changing the indices but not the templates or customs. Currently, cluster state updates that touch that kind of object tree are paying a price for these things whether they change anything in them or not. With this PR, all those things we don't touch now cost nothing.

[elasticmachine]

Pinging @elastic/es-core-infra (Team:Core/Infra)

[elasticsearchmachine]

Hi @joegallo, I've created a changelog YAML for you.

[original-brownbear]

LGTM, benchmarked this and it offers some unexpected speedups indeed by avoiding copying the large indices map in the metadata builder whenever we're updating a separate part of the metadata (I think that's where most of the speedups are from). Thanks Joe!