mpt: make trie update stats monotonic with a snapshot API

[maxkozlovsky]

TrieUpdateCollectedStats was reset at the start of every upsert, so a sampler outside the upsert path saw counters that vanished at the next upsert's start. Keep the totals for the lifetime of the UpdateAux instead, expose them through stats_snapshot() / last_upsert_stats(), and recover per-upsert figures by subtracting two snapshots via delta_since().

This allows exporting db stats through prometheus or similar metrics.

Also drop MONAD_MPT_COLLECT_STATS. It was hard-wired to 1 with no way to set it to 0, so its #else branches were never compiled and had rotted -- flipping it failed to build. Collection is now unconditional, which is what every build has always done.

[Copilot]

Pull request overview

This PR makes trie update statistics monotonic across the lifetime of UpdateAux, enabling external samplers/metrics exporters (e.g., Prometheus) to observe counters without them being reset at the start of each upsert. It also removes the unused MONAD_MPT_COLLECT_STATS compile-time switch and makes stats collection unconditional.

Changes:

• Replace per-upsert resettable stats with lifetime totals (stats_) plus snapshot/delta tracking (prev_upsert_stats_, last_upsert_stats_) and expose them via stats_snapshot() / last_upsert_stats().
• Compute per-upsert figures as delta_since() between consecutive snapshots and plumb those deltas into logging and compaction speed-control logic.
• Add a unit test asserting that last_upsert_stats() matches snapshot deltas across consecutive upserts (including when compaction=false).

Verdict: CORRECT

Generated with Claude Code

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
category/mpt/update_aux.cpp	Computes/stores per-upsert deltas from monotonic lifetime totals; updates logging and compaction logic to consume the delta.
category/mpt/trie.hpp	Adds snapshot APIs and moves stats members to private lifetime-tracking fields.
category/mpt/trie.cpp	Removes the per-upsert reset_stats() call.
category/mpt/test/compaction_test.cpp	Adds coverage ensuring last_upsert_stats() isolates the latest upsert and matches snapshot differences.
category/mpt/detail/collected_stats.hpp	Removes MONAD_MPT_COLLECT_STATS, makes the struct always full-sized, and adds delta_since() API.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Reviewed the change. The refactor keeps TrieUpdateCollectedStats as a lifetime accumulator inside UpdateAux, computes each upsert's contribution as a delta between snapshots at the end of do_update, and threads the delta explicitly into print_update_stats. That preserves the semantics that mattered:

• advance_compact_offsets needs "bytes compacted to slow during the previous upsert" to set gc_efficiency. Under the old code, it read stats.compacted_bytes_in_slow before upsert()'s reset_stats() fired, so it saw the previous upsert's total. last_upsert_stats_.compacted_bytes_in_slow is now written unconditionally at end of do_update and is the exact same quantity, since every collect_* call happens inside upsert() and cannot fire between do_update invocations.
• The unconditional assignment before the if (compaction) block correctly ensures a compaction=false upsert's counters can't leak into the next compacted upsert's delta — and the new test pins that behavior.
• delta_since() covers all 16 named fields (including both [2] arrays via the loop), and the sizeof == 160 and sizeof(UpdateAux) == 120 + 3 * sizeof(TrieUpdateCollectedStats) static_asserts will fire if a field is added, forcing the author to look at delta_since() too.
• Grep confirms no external consumers of the removed aux.stats, reset_stats, or MONAD_MPT_COLLECT_STATS; the #include <new> drop matches the placement-new removal.
• The added test uses the existing FillDBWithChunksGTest fixture and validates the monotonicity + delta-isolation invariants that a reintroduced reset would break.

Style, naming, and PR hygiene (single commit, #pragma once, trailing-underscore private members, east-const on parameter definitions) all check out.

Verdict: CORRECT

🤖 Generated with Claude Code