perf(bigquery): [T1-12] replace Channels.newChannel bounce buffer with direct ByteBuffer channel on Arrow batches

[lqiu96]

Summary & Bottleneck

• Target files:
  • java-bigquery/google-cloud-bigquery/src/main/java/com/google/cloud/bigquery/ArrowDeserializer.java (L165-L169)
  • java-bigquery/google-cloud-bigquery/src/main/java/com/google/cloud/bigquery/ArrowQueryResultImpl.java (L528-L530)
  • java-bigquery/google-cloud-bigquery/src/main/java/com/google/cloud/bigquery/ConnectionImpl.java (L1115-L1118, L1139-L1142)
• Scaling cardinality: $O(\text{batches})$ allocations and double-copies across all Arrow IPC record batches.
• Bottleneck details:
  1. ArrowDeserializer.loadArrowRows and ArrowQueryResultImpl.VectorBatchIterator.loadBatch(ByteString) wrapped batch.getSerializedRecordBatch().newInput() in Channels.newChannel(...). In JDK's Channels.ReadableByteChannelImpl, read(ByteBuffer dst) allocates a byte[8192] transfer buffer, acquires synchronized (readLock) on every read, copies bytes from InputStream into byte[8192], and then copies from byte[8192] into dst.
  2. ConnectionImpl.ArrowRowReader called batch.getSerializedRecordBatch().toByteArray() and arrowSchema.getSerializedSchema().toByteArray(), allocating a full byte[] copy of every serialized Arrow schema and record batch payload before wrapping it in ByteArrayReadableSeekableByteChannel.

Changes

• Direct ByteBuffer channel over ByteString.asReadOnlyByteBuffer():
  • Added ArrowDeserializer.newByteStringChannel(ByteString) returning a lightweight, unsynchronized ByteBufferReadableByteChannel that copies slices directly from byteString.asReadOnlyByteBuffer() into the target ByteBuffer.
• Replaced Channels.newChannel and toByteArray() across Arrow readers:
  • Updated ArrowDeserializer.loadArrowRows, ArrowQueryResultImpl.VectorBatchIterator.loadBatch(ByteString), and ConnectionImpl.ArrowRowReader to use ArrowDeserializer.newByteStringChannel(...).

Expected Impact

• Eliminates the byte[8192] bounce buffer allocation, double memory copy, and synchronized (readLock) overhead per Arrow record batch in ArrowDeserializer and ArrowQueryResultImpl.
• Eliminates full-batch byte[] array copies (toByteArray()) per Arrow record batch in ConnectionImpl.ArrowRowReader.

Verification

• Ran mvn test -pl java-bigquery/google-cloud-bigquery -Dtest=ArrowDeserializerTest,ArrowQueryResultTest,ArrowQueryPageFetcherTest,ConnectionImplTest (57 tests run, 0 failures).
• Added ArrowDeserializerTest.testNewByteStringChannelReadsIncrementalSlicesAndCloses to verify incremental slice reads, EOF semantics, and channel closure behavior.