issue #547: wire jvector readRangeAsync to RemoteFileServiceClient for parallel FusedPQ IO

[eolivelli]

Fixes #547.

Changes

• .github/workflows/ci.yml + kubernetes-tests.yml: pin jvector checkout to parallel-fusedpq-io branch while eolivelli/jvector#8 is unmerged. Both workflows had ref: main; must track the PR branch to pick up the new readRangeAsync API. TODO: revert to ref: main once the PR merges.

• SegmentBlockCache: add AsyncBlockLoader functional interface + getBlockAsync(path, offset, length, AsyncBlockLoader) method. Uses a ConcurrentHashMap<BlockKey, CompletableFuture<ByteBuf>> inFlightAsync for single-flight deduplication: concurrent misses for the same block share one loader call, each getting an independent retained slice. Hit path is an atomic computeIfPresent retain under the Caffeine map lock — the same pattern as getBlock. Cache insertion uses compute() to handle the rare race with a concurrent sync getBlock call. RefCnt ownership rules (refCnt=2 on insert: cache ref + caller ref) are identical to the sync path.

• RemoteRandomAccessReader: override readRangeAsync(long, int) from jvector's RandomAccessReader. Does NOT touch position, blockBuffer, or bufferedBlockIndex (async reads bypass the sliding-window cursor). Single-block fast path calls one getBlockAsync; multi-block path fires all covering blocks in parallel with allOf, assembling a heap ByteBuffer in the completion callback. All ByteBuf slices are released in a finally block (both success and failure paths) so no off-heap memory escapes. fetchBlockFromRemoteAsync helper mirrors fetchBlockFromRemote — clamps length to avoid reading past EOF, updates rfs_client_read_* counters and VectorSearchRequestContext accounting.

• PersistentVectorStore: add static volatile boolean searchAsyncPipelineEnabled with setSearchAsyncPipelineEnabled() / isSearchAsyncPipelineEnabled(). Mirrors the setStreamingCompactionEnabled pattern. Default false until a production profile confirms a net win.

• VectorSegment.search(): after getting or creating the ThreadLocal-cached GraphSearcher, call searcher.setAsyncPipelineEnabled(PersistentVectorStore.searchAsyncPipelineEnabled) on every search invocation. Cheap field write; enables runtime toggle without discarding cached searcher instances.

• IndexingServerConfiguration: add PROPERTY_VECTOR_SEARCH_ASYNC_PIPELINE_ENABLED / SYSPROP_VECTOR_SEARCH_ASYNC_PIPELINE_ENABLED constants (default false).

• IndexingServiceEngine: read the new property at IS startup and call PersistentVectorStore.setSearchAsyncPipelineEnabled(); config key takes precedence over system property. Logs the resolved value.

Tests

• RemoteRandomAccessReaderAsyncTest (11 cases, Netty PARANOID leak detector): readRangeAsync returns identical bytes to seek+readFully for single-block, cross-block, three-block, and end-of-file ranges; position is invariant after async reads; 16 concurrent async reads alongside a synchronous readFully loop on the same reader instance produce correct bytes with no refcount exceptions; cache warm-up via sync path is visible to subsequent async reads; disabled-cache pass-through works.

• SegmentBlockCacheAsyncTest (7 cases, Netty PARANOID leak detector): pass-through on disabled cache; miss populates then next call is a hit; sync miss populates then async call is a hit; 16 concurrent async misses share exactly one loader call; each concurrent caller receives an independent ByteBuf slice; loader failure propagates to all waiting callers; multi-threaded async loads are byte-correct.

🤖 Implemented by the pr-worker agent.