[lqiu96]

Summary & Bottleneck

• Target files:
  • java-bigquery/google-cloud-bigquery/src/main/java/com/google/cloud/bigquery/ConnectionImpl.java (L547-L570, L601-L619, L1030)
  • java-bigquery/google-cloud-bigquery/src/main/java/com/google/cloud/bigquery/BigQueryResultImpl.java (L195-L225)
• Scaling cardinality: $O(\text{rows})$ node allocations and $O(\text{rows})$ lock/unlock cycles across producer and consumer threads.
• Bottleneck details:
  1. ConnectionImpl used LinkedBlockingDeque for bufferFvl, bufferRow, pageCache, and rpcResponseQueue. Every buffer.put(row) allocated a 24-byte LinkedBlockingDeque$Node object on the heap and acquired a single shared ReentrantLock contended by both the producer and consumer threads.
  2. BigQueryResultImpl.BigQueryResultSet.next() called buffer.take() one row at a time, acquiring and releasing the queue lock on every single row.

Changes

• Bounded ArrayBlockingQueue ring buffers in ConnectionImpl:
  • Replaced LinkedBlockingDeque with ArrayBlockingQueue for bufferFvl, bufferRow, pageCache, and rpcResponseQueue in tableDataList, processQueryResponseResults, and highThroughPutRead.
• Consumer-side batch draining in BigQueryResultSet.next():
  • Added a consumer-local ArrayDeque<T> localBatch (capacity 256) in BigQueryResultSet.
  • When localBatch is empty, next() takes the first available element via buffer.take() and immediately calls buffer.drainTo(localBatch, 256) under a single lock acquisition. Subsequent next() calls serve rows from localBatch without lock contention.

Expected Impact

• Eliminates $R$ LinkedBlockingDeque$Node allocations (24 MB per 1M rows) by pre-allocating a contiguous ring buffer array.
• Amortizes consumer ReentrantLock acquisitions in ResultSet.next() by up to 256x when buffered rows are available.

Verification

• Ran mvn test -pl java-bigquery/google-cloud-bigquery -Dtest=BigQueryResultImplTest,ConnectionImplTest (26 tests run, 0 failures).
• Added BigQueryResultImplTest.testResultSetBatchDrainingPreservesFifoOrderAndEof to verify FIFO ordering and EOF handling across 600 rows (spanning multiple 256-row drain batches).