Performance: successful commit path re-deserializes the just-inserted BSON document

[AGiorgetti]

Summary

After a successful insert, Commit() returns commitDoc.ToCommit(_serializer). This means the write path serializes the commit to BSON for insert, then immediately deserializes the same BSON back to ICommit to return it — pure redundant work.

Why this matters

• Post-insert CPU and allocation overhead on every successful commit.
• The cost is independent of MongoDB round-trip latency.
• Likely noise relative to total write-path time, but worth quantifying before dismissing.

Proposed investigation

• Measure the post-insert deserialization cost as a fraction of total write-path time using MemoryDiagnoser.
• Evaluate building the return ICommit from the original in-memory object instead of round-tripping through BSON.
• Only act if the cost is measurable (> ~1% of total write time).

Reference

See docs/Performance-Investigation.md → Finding #5 for full context.

[AGiorgetti]

Write-side follow-up for #77:

I inspected the successful commit return path in both sync and async MongoPersistenceEngine.Commit implementations. Both insert the BSON commit document, then return commitDoc.ToCommit(_serializer).

Important detail:

• The obvious shortcut would be to construct Commit directly from CommitAttempt plus the generated checkpoint.
• That is not clearly behavior-preserving because the current returned ICommit is materialized from the serialized BSON document. A direct CommitAttempt-based return would hand back the caller's original EventMessage instances instead of serializer round-tripped event messages.
• Commit hooks receive this returned ICommit, so the difference is observable outside the persistence engine.

Benchmark note:

• I attempted to stabilize and run the broad WriteToStreamBenchmarks slice for this question, but the 10,000-commit case exceeded the command timeout and produced no usable benchmark archive.
• I reverted that benchmark job change rather than leaving a slower/unusable benchmark configuration behind.

Decision: put #77 in standby. Do not change the runtime commit return path without a dedicated microbenchmark that isolates commitDoc.ToCommit(_serializer) and tests that define acceptable return materialization semantics.