CASSANDRA-21605: Improve algorithmic complexity for finding intersecting SSTables in LCS

[cheeeee]

Jira Issue

• CASSANDRA-21605

Motivation

In LeveledCompactionStrategy, SSTables in levels L1 through L8 are stored in a TreeSet and are strictly disjoint and sorted by first key. However, LeveledManifest.getCandidatesFor(level) previously performed an $O(N)$ linear scan over all SSTables in the next level, constructing an intermediate HashMap of bounds for the entire level (genBounds) on every candidate iteration. On large tables with thousands of SSTables per level, this resulted in $O(N \times M)$ comparisons and heavy garbage generation.

Changes

1. $O(\log N + K)$ Binary Search (LeveledGenerations): Added getOverlapping(level, sstable). For levels $\ge 1$, queries the level's TreeSet using sortedLevel.floor(sstable) to locate the sole predecessor that could potentially overlap, then iterates forward via tailSet, breaking immediately once candidate.first.token > sstable.last.token.
2. Zero-Garbage Selection (LeveledManifest): Eliminated the intermediate genBounds HashMap creation and invoked generations.getOverlapping(level + 1, sstable).
3. Complexity Reduction: Candidate intersection checks drop from $O(N \times M)$ to $O(N \log M + N \times K)$, where $K \ll M$ is the number of overlapping tables.

Testing

• LeveledGenerationsTest: 5/5 tests pass (added testGetOverlapping covering exact boundaries, multi-table spans, gaps between tables, floor-only, successor-only, and empty levels).
• LeveledCompactionStrategyTest: full test suite passes cleanly (18/18 tests).