perf(store): lazy-init sortedCache in cachekv.Store

[pdrobnjak]

Summary

• Defer MemDB (btree + FreeList) allocation in cachekv.Store until the first iterator is actually requested
• During EVM transaction execution, sortedCache is never accessed — only Get/Set/Delete are called, which use store.cache (sync.Map), not sortedCache
• This eliminates ~42,000 wasted MemDB allocations per block (1000 txs × ~2 snapshots × 21 module stores) for the EVM transfer workload

Problem

Profiling the Giga OCC executor on origin/main showed:

Bottleneck	CPU %	Root Cause
Heap allocator lock contention (runtime.lock2)	23.0%	24 OCC workers all allocating MemDB/btree simultaneously
GC pressure (runtime.gcDrain)	21.6%	Scanning 9.5 GB/s of allocations
Actual EVM execution (evmone)	11.5%	The useful work

Top allocators over 30s: btree.NewFreeListG (57 GB) + tm-db.NewMemDB (70 GB) = 127 GB of the 286 GB total — all from cachekv.NewStore() eagerly creating a sortedCache that is never used.

sortedCache is only read in the iterator() → dirtyItems() → clearUnsortedCacheSubset() path. For EVM transfers, iteration never happens — the stores only do point reads/writes. Iteration only occurs during EndBlock (receipt flush, ante surplus, pruning) and Cosmos native transactions (gov, staking, bank queries), which account for a tiny fraction of stores created.

Changes

4 lines changed in sei-cosmos/store/cachekv/store.go:

1. NewStore: sortedCache: nil instead of dbm.NewMemDB()
2. Write(): store.sortedCache = nil instead of dbm.NewMemDB()
3. New ensureSortedCache(): allocates MemDB on first use
4. iterator(): calls ensureSortedCache() before dirtyItems()

Benchmark Results (M4 Max, EVMTransfer 1000 txs/batch)

Metric	Before	After	Delta
TPS (median steady-state)	~7,800	~8,200	+5%
runtime.lock2 (heap contention)	23.0% CPU	15.4% CPU	-13.45s
runtime.gcDrain	21.6% CPU	19.0% CPU	-6.95s
runtime.(*mheap).allocSpan	18.4% CPU	11.4% CPU	-12.02s
btree.New / NewMemDB / FreeList	~11s combined	~0s	eliminated
Total CPU samples (30s)	142.2s	124.9s	-12%

Test plan

• cd giga/tests && go test ./... -count=1 — all 14 tests pass
• gofmt -s -l sei-cosmos/store/cachekv/store.go — clean
• Benchmark: TPS improved ~5%, no regressions
• CPU profile diff confirms allocation reduction
• CI passes

🤖 Generated with Claude Code

[github-actions]

The latest Buf updates on your PR. Results from workflow Buf / buf (pull_request).

Build	Format	Lint	Breaking	Updated (UTC)
✅ passed	✅ passed	✅ passed	✅ passed	Feb 6, 2026, 3:29 PM

[codecov]

Codecov Report

❌ Patch coverage is 77.77778% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 56.57%. Comparing base (5c26262) to head (4be37df).

Files with missing lines	Patch %	Lines
sei-cosmos/store/cachekv/store.go	77.77%	0 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2804      +/-   ##
==========================================
- Coverage   56.58%   56.57%   -0.01%     
==========================================
  Files        2033     2033              
  Lines      166263   166267       +4     
==========================================
- Hits        94074    94065       -9     
- Misses      63926    63946      +20     
+ Partials     8263     8256       -7     

Flag	Coverage Δ
sei-chain	41.49% <77.77%> (-0.02%)	⬇️
sei-cosmos	48.13% <77.77%> (+<0.01%)	⬆️
sei-db	68.72% <ø> (ø)
sei-tendermint	57.91% <ø> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
sei-cosmos/store/cachekv/store.go	88.82% <77.77%> (+0.24%)	⬆️

... and 21 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.