perf(spanner): optimize chunk processing and row decoding in PartialResultStream

[olavloite]

Eliminate intermediate array allocations and array mutations during query stream processing in PartialResultStream:

• Process chunk values in-place using startIndex and endIndex cursors in _addChunk, eliminating the intermediate values array allocation and O(N) shift/unshift array operations during chunked value merging.
• Pre-allocate a single row buffer (_rowValues) per result set and track cell progress with an index cursor (_valueIndex), eliminating per-row array allocations in _addValue.
• Preserve the row write cursor (_valueIndexForResume) alongside pending chunked values on resume token checkpoints, ensuring retries across incomplete rows correctly resume at the exact column position.
• Guard this.emit('paused') to fire at most once per chunk upon entering the backpressured state.

[gemini-code-assist]

Code Review

This pull request refactors PartialResultStream to optimize row buffering and chunk processing by replacing dynamic array pushes with a pre-allocated array and a write cursor. It also structures chunk processing into five distinct stages and adds comprehensive unit tests covering backpressure, multi-chunk streaming, and buffer reuse. The review feedback highlights potential runtime errors and out-of-bounds access if an incoming chunk is empty during pending value merging or chunked value handling, and suggests simplifying a redundant ternary check when initializing the row values array.

[olavloite]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors PartialResultStream to use a pre-allocated row buffer (_rowValues) and a cursor index (_valueIndex) instead of a dynamic array, improving performance and memory usage during chunked streaming. It also updates chunk processing, adds checkpoint tracking for stream retries, and introduces new tests for backpressure and multi-chunk streaming. However, a critical bug was identified where _rowValues and _valueIndex are overwritten when a stream resumes and receives metadata, discarding the restored state and leading to data corruption. Removing the state reset from the metadata block is recommended to preserve the resume checkpoint.

[olavloite]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors PartialResultStream to use a pre-allocated row buffer (_rowValues) and index tracking (_valueIndex) instead of dynamically pushing to an array, improving memory efficiency and streaming robustness. It also adds comprehensive unit tests covering backpressure, multi-chunk streaming, and stream resumption. The review feedback recommends clearing stale elements in _rowValues from the reset _valueIndex to the end of the array during _resetPendingValues() to prevent potential memory retention of large decoded objects.

[olavloite]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors PartialResultStream to optimize row assembly and stream chunk processing. It replaces the dynamic _values array with a pre-allocated _rowValues array and a _valueIndex cursor, improving memory usage and avoiding cross-row contamination. Additionally, it enhances resume token checkpointing by preserving the exact column position (_valueIndexForResume) and pending chunked values during retries. Comprehensive unit tests have been added to verify backpressure handling, multi-chunk streaming, row buffer reuse, and state restoration upon stream resumption. No review comments were provided, so there is no feedback to address.