perf: remove pending cache

[carneiro-cw]

Currently a single slot read in a transaction results in at least 3 cache reads, and at most 4 cache reads, 1 temp read, 1 perm read and 2 cache writes. These changes make it so at least it results in 1 temp-read, and at most 1 temp, 1 cache and 1 perm read and 0 cache writes. (with temp + latest cache covering 100% of the reads the pending cache potentially covered)

The temp now acts as the "pending cache" effectively, and the latest cache is updated with slots that were touched alongside slots that were modified when we commit the block. The latest cache is also still updated by eth_call.

Local benchmarking showed a 7% increase in throughput after these changes.

We're likely to see some more gains by optimizing the temp storage's State merge operation and reads.

PR Type

Enhancement, Tests

---

Description

• Rename evm_input/result to input/output fields

• Introduce Full stage for richer Changes API

• Remove pending cache and simplify storage flows

• Update miner, importer, and stratus_storage accordingly

---

File Walkthrough

	Relevant files
Enhancement	19 files mod.rs Rename transaction execution fields in EVM executor            +5/-5      transaction_execution.rs Parameterize `Changes` with `Full` stage                                  +8/-20    util.rs Use `input`/`output` in default_trace tracer                          +4/-4      mod.rs Switch to `output` field in metrics and revert logging      +3/-2      changes.rs Add `Full` stage and extend `Changes` methods                        +127/-74 transaction_execution.rs Rename `evm_input`/`result` to `input`/`output`                    +15/-15  fake_leader.rs Remove pending cache, simplify fake leader import                +5/-8      miner.rs Update block mining API with `Full` changes                            +14/-23  cache.rs Remove old caches and fix insert_if_missing                            +15/-56  rocks_permanent.rs Use `Changes` in genesis and block saves              +3/-2      rocks_state.rs Update batch execution with `Changes`                    +12/-10  stratus_storage.rs Drop pending cache, adjust resolve and cache logic              +42/-90  call.rs Apply `output` changes in in-memory call storage                  +6/-9      transaction.rs Switch to `input`/`output` in in-memory temp storage          +20/-14  block.rs Rename transaction timestamp and input usage                          +3/-3      transaction_mined.rs Map `input`/`output` in rocksdb conversion                              +4/-4      transaction_stage.rs Use `output` in transaction stage to_result                            +2/-2      metrics_definitions.rs Remove unused metric labels in finish_pending_block            +1/-1      events.rs Use `input`/`output` fields in event parsing                          +12/-12
Additional files	5 files mod.rs +1/-0      mod.rs +5/-4      transaction_mined.rs +17/-17  resolve_pending.rs +16/-24  mod.rs +11/-10

---

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 5 🔵🔵🔵🔵🔵

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Error Swallowing The new read_temp method returns Option<Self> and drops any StorageError from temporary storage reads, conflating errors with a missing value. This hides underlying storage failures and can lead to silent data inconsistencies. Errors should be propagated or at least logged with a reason field. fn read_temp(s: &StratusStorage, key: Self::Key, kind: ExecutionKind) -> Option<Self>; /// Reads from permanent storage at the resolved mined point. fn read_perm(s: &StratusStorage, key: Self::Key, point: MinedPointInTime<'_>) -> Result<Self, StorageError>; /// Caches the value as a latest (mined tip) entry, if not already cached. fn cache_latest_if_missing(s: &StratusStorage, key: Self::Key, value: Self); } impl EntityRead for Account { type Key = Address; fn read_temp(s: &StratusStorage, address: Address, kind: ExecutionKind) -> Option<Self> { tracing::debug!(storage = %label::TEMP, %address, "reading account"); timed(|| s.temp.read_account(address, kind)).with(|m| { if m.result.is_some() { metrics::inc_storage_read_account(m.elapsed, label::TEMP, PointInTime::Pending, true); } }) } Missing Error Logging In finish_pending_block, errors from self.temp.finish_pending_block() are no longer logged or surfaced. Without logging the reason, failures in finishing the pending block will go unnoticed. Add a structured error log (e.g., tracing::error!(reason = ?e)) inside the timed callback. pub fn finish_pending_block(&self) -> (PendingBlock, Changes<Full>) { #[cfg(feature = "tracing")] let _span = tracing::info_span!("storage::finish_pending_block", block_number = tracing::field::Empty).entered(); tracing::debug!(storage = %label::TEMP, "finishing pending block"); let result = timed(|| self.temp.finish_pending_block()).with(|m| { metrics::inc_storage_finish_pending_block(m.elapsed); }); Unstructured Tracing The tracing::warn! call in save_execution uses a formatted string "Failed transaction contains {} slot change(s)" with a positional placeholder. For structured tracing, emit slot_changes = total_slot_changes as a field rather than using {} formatting. if !tx.output.result.is_success() { let total_slot_changes: usize = changes.slots.len(); if total_slot_changes > 0 { tracing::warn!(?tx, "Failed transaction contains {} slot change(s)", total_slot_changes); }

[github-actions]

PR Code Suggestions ✨

Explore these optional code suggestions:

Category	Suggestion	Impact
Possible issue	Correctly extract slot values The code attempts to call a non-existent .value() on CompleteValue. Extract the inner SlotValue via .clone().take_value() before inserting. src/eth/storage/cache.rs [75-77] for ((address, index), value) in changes.slots.iter() { - slot_cache.insert((*address, *index), *value.value()); + let slot_val = value.clone().take_value(); + slot_cache.insert((*address, *index), slot_val); } Suggestion importance[1-10]: 7 __ Why: The code calls a non-existent .value() on CompleteValue; using .clone().take_value() properly retrieves the inner SlotValue for insertion and fixes the compilation and runtime behavior.	Medium

[stratus-benchmark]

Benchmark:
Run ID: bench-28d4a294

Git Info:

• Branch: remove_pending_cache
• Commit: 26a6932
• PR: perf: remove pending cache #2630

Leader Stats:
RPS Stats: Max: 8152.00, Min: 1480.00, Avg: 3037.73, StdDev: 351.83
TPS Stats: Max: 3557.00, Min: 213.00, Avg: 2992.01, StdDev: 300.85

Follower Stats:
Imported Blocks/s: Max: 8.00, Min: 3.00, Avg: 5.85, StdDev: 1.26
Imported Transactions/s: Max: 24309.00, Min: 8883.00, Avg: 17491.73, StdDev: 3826.05

Plots:

• View Leader Plot
• View Follower Plot

[stratus-benchmark]

Benchmark:
Run ID: bench-e064e3b5

Git Info:

• Branch: remove_pending_cache
• Commit: 8001df9
• PR: perf: remove pending cache #2630

Leader Stats:
RPS Stats: Max: 6747.00, Min: 1614.00, Avg: 2726.78, StdDev: 288.19
TPS Stats: Max: 3037.00, Min: 752.00, Avg: 2685.50, StdDev: 235.35

Follower Stats:
Imported Blocks/s: Max: 8.00, Min: 3.00, Avg: 6.47, StdDev: 1.41
Imported Transactions/s: Max: 22340.00, Min: 8215.00, Avg: 17370.06, StdDev: 3734.74

Plots:

• View Leader Plot
• View Follower Plot

[stratus-benchmark]

Benchmark:
Run ID: bench-9c4318e3

Git Info:

• Branch: remove_pending_cache
• Commit: 8001df9
• PR: perf: remove pending cache #2630

Leader Stats:
RPS Stats: Max: 10162.00, Min: 1495.00, Avg: 2875.98, StdDev: 456.51
TPS Stats: Max: 3211.00, Min: 195.00, Avg: 2821.09, StdDev: 275.64

Follower Stats:
Imported Blocks/s: Max: 9.00, Min: 2.00, Avg: 5.85, StdDev: 1.55
Imported Transactions/s: Max: 25916.00, Min: 5593.00, Avg: 16492.52, StdDev: 4581.75

Plots:

• View Leader Plot
• View Follower Plot