[Metadata Immutability] Change different indices lookup objects from array type to lists

[sandeshkr419]

Is your feature request related to a problem? Please describe.
Metadata object as part of ClusterState is supposed to be immutable by design. However, certain objects within Metadata do not reflect this design.

The following arrays should be replaced with lists and should implement Collections.unmodifiableList() or any other immutable implementations. Assigning them as arrays leave them open for mutability.
https://github.com/opensearch-project/OpenSearch/blob/main/server/src/main/java/org/opensearch/cluster/metadata/Metadata.java#L268

 String[] allIndices,
 String[] visibleIndices,
 String[] allOpenIndices,
 String[] visibleOpenIndices,
 String[] allClosedIndices,
 String[] visibleClosedIndices,

Also, the usage of these arrays is prominently in IndexNameExpressionResolver where before these arrays are utilized, they are converted to lists before they are consumed.
It is thus unreasonable to keep these objects as arrays as opposed to lists.

Describe the solution you'd like
Store the above objects as lists, and implement them as Collections.unmodifiableList() in the Metadata constructor.

Describe alternatives you've considered
Briefly discussed this change here, but this change was out of scope of #7853

Additional context
this change will require changes in how the metadata or its diff is parsed as XContent as well.

[andrross]

What do you think about this comment:

// TODO: I think we can remove these arrays. it isn't worth the effort, for operations on all indices.
// When doing an operation across all indices, most of the time is spent on actually going to all shards and
// do the required operations, the bottleneck isn't resolving expressions into concrete indices.

While I completely agree with you that immutable lists are better than mutable arrays, my first thought when digging into this code was to remove the arrays all together, then I found the above comment. If I understand correctly, then the existing getters would change to compute the results on demand. For example, getConcreteVisibleIndices() would change to:

/**
 * Returns all the concrete indices that are not hidden.
 */
public List<String> getConcreteVisibleIndices() {
    return indices.entrySet().stream()
        .filter(e -> IndexMetadata.INDEX_HIDDEN_SETTING.get(e.getValue().getSettings()) == false)
        .map(Map.Entry::getKey)
        .collect(Collectors.toUnmodifiableList());
}

(This also changes the return type from String[] to List, which isn't strictly necessary since mutability is no longer a concern given that a new array would be computed each invocation. However, if this is always being used as a list then it should probably just return a list.)

Pre-computing the filtered results strike me as an unnecessary optimization and I would suggest doing the simpler thing unless we know for sure there are performance benefits to pre-computing.

[sandeshkr419]

@andrross Thanks for taking a look. I had visited the comment you mentioned in my previous PR #7853 which touched the same code base.

I don't support the idea of computing different indices lookups (arrays in concern) on demand is that these different different arrays are used for expression resolving which is invoked in each almost all search and index data modification related calls. Calculating these on demand will surely increase the latency of all search and indexing related operations, which we should avoid.
Computing these arrays with each cluster state update does not even takes time of the order of miliseconds. Plus trading off some memory/time in ClusterState update activity seems a better choice than to put it on indexing/search tasks where latency is required at minimal.

Also, I would like to point out that except the getConcreteAllIndices(), all the other arrays (5 of them) are only invoked at this resolveEmptyOrTrivialWildcard() in the entire code base, where they are used as lists directly.

[andrross]

Thanks @sandeshkr419! It makes total sense if these data structures are used in hot path indexing and search code then precomputing can be a good tradeoff. However, these TODO comments should be resolved or removed to avoid future confusion. The TODO comment referenced in #7853 actually does make a lot of sense (if I understand it correctly) as it is suggests to move the pre-computed structures into IndexNameExpressionResolver. That would likely make the code and intent more clear as the structures would more clearly be in hot path code (index name resolution). I haven't looked into what it would take to do that refactoring. Either way, I suggest addressing or removing the TODOs wherever possible.

[sandeshkr419]

@andrross Yes, we should remove these comments altogether. We did not removed it previously just to be sure that we have not misunderstood the correct context of these comments. Plus, there is also discussion which seems to be initiated in future by @shwetathareja on remote cluster state as in #7002 (comment)

For this issue, I would like to propose a minor enhancement only in terms of Metadata immutability.

[andrross]

@sandeshkr419 I'm all for incremental improvements. However, I do think you need to do something with the comment that says "TODO: I think we can remove these arrays..." as a part of this change since you're actually working with the arrays in question here.

[sandeshkr419]

@andrross Seems reasonable to remove the comment since we have valid reasons (potential performance regressions) of not implementing this.

@shwetathareja What are your thoughts on this (removing the TODOs - since we have reason of not implementing them).

[shwetathareja]

Like the comment said, these arrays are used during index resolution. We can benchmark to see how many cpu cycles are spent to generate these arrays during ingestion and search traffic. Personally I think even if there is small performance gain, we should keep them. @sandeshkr419 @andrross