[Ronitsabhaya75]

Problem

getRangeStream was registering the entire requested key range [min(b, begin), max(e, end)] as a read conflict range upfront before reading any data. This resulted in false conflict aborts if the client only read part of the stream and committed, if the range was empty/inverted (b >= e), or if an error occurred before reaching later shards.

Solution

• Track streaming conflict ranges in TransactionState and extend the range incrementally in getRangeStreamImpl as each chunk is yielded to the caller.
• Append accumulated stream conflict ranges into tr.transaction.read_conflict_ranges when preparing commitMutations().
• Return immediately on empty ranges (b >= e) without adding a conflict range.

Testing

• Validated lifecycle against NativeAPI conflict range handling and TransactionState reset/commit paths.

[saintstack]

LG. Local LLM found an issue:

An empty resolved range (b >= e) now records no read conflict range. The old code recorded [min(b, begin.getKey()), max(e, end.getKey())) before this check. Resolving the selectors already read keys between begin.getKey() and b and between e and end.getKey(). A concurrent insert there could make the range non-empty, and the transaction would no longer conflict. Non-stream getRange still records a range for an empty result (getRangeFinished). The early return in Transaction::getRangeStream, before any resolution, is fine because nothing has been read yet. Fix: when !snapshot, add [rangeBegin, rangeEnd) before the end_of_stream return in ::getRangeStream.

[saintstack]

Here are another few you might to look at:

• Reverse branches: the reverse logic at NativeAPI.cpp:3430 and :3448 is wrong, but unreachable. Transaction::getRangeStream asserts !reverse.
• readConflictRanges() lifetime: it uses a shallow push_back of arena-backed ranges (NativeAPI.h:493-500). It could dangle after reset, but nothing in the RYW layer calls getRangeStream.
• Unbounded range growth: one range is added per chunk, never merged, and added after the transaction_too_large size check. This is a scalability regression rather than a correctness bug.
• Abandoned streams: an abandoned stream with an end selector of offset <= 0 skips the end-selector region.
• Tests and style: the PR adds no tests, and the getRangeStreamImpl call at line 3528 is over 120 columns.

[Ronitsabhaya75]

@saintstack thank you for your review I've update pr according to your review