perf(types): cache gaskv.Store wrappers in Context

[pdrobnjak]

Summary

• Cache gaskv.Store wrappers in a map[StoreKey]KVStore on Context, eliminating repeated heap allocations from ctx.KVStore(key) and ctx.TransientStore(key)
• Cache invalidated on WithMultiStore / WithGasMeter (the only methods that change gaskv inputs)

Problem

Every ctx.KVStore(key) and ctx.TransientStore(key) call allocates a new gaskv.Store on the heap (5-field struct, ~40 bytes). Each wraps the same underlying KVStore + same gas meter + same gas config — pure waste to reallocate.

Profiling after #2806 (30s, M4 Max, 1000 EVM transfers/block):

Metric	Value
gaskv.NewStore alloc_space	18,892 MB (#9 top allocator)
gaskv.NewStore alloc_objects	~143M objects
Top callers	AccountKeeper.GetAccount (104.5M), params.Subspace.kvStore (46.6M), bank.getAccountStore (20.7M)

Changes

sei-cosmos/types/context.go (single file):

1. New field: gaskvStores map[StoreKey]KVStore on Context struct
2. NewContext: Initialize cache with make(map[StoreKey]KVStore, 8)
3. WithMultiStore / WithGasMeter: Invalidate cache (new empty map)
4. KVStore / TransientStore: Check cache first; populate on miss

Benchmark Results (M4 Max, 1000 EVM transfers/block, 30s profile)

Metric	Before (after #2806)	After	Delta
TPS (steady-state range)	8,000–8,800	7,800–9,000	median ~8,400
gaskv.NewStore alloc_space	18,892 MB	0 MB	-18,892 MB (eliminated)
gaskv.NewStore alloc_objects	~143M	0	eliminated from top 200
memclrNoHeapPointers CPU	6.96s	2.11s	-4.85s
btree.NewFreeListG alloc	36,662 MB	20,943 MB	-15,719 MB
sync.Map nodes alloc	33,216 MB	19,103 MB	-14,113 MB
cachemulti.newStoreWithoutGiga alloc	39,081 MB	22,471 MB	-16,610 MB
Total alloc_space (30s)	811,276 MB	457,143 MB	-354,133 MB (-44%)

Note: TPS uplift is modest because freed CPU shifts to idle (usleep/kevent up) — workers complete faster but the bottleneck has shifted elsewhere (likely OCC scheduling / commit pipeline).

pprof -alloc_space -diff_base highlights

-18,298 MB  gaskv.NewStore               (direct savings from this PR)
-17,416 MB  cachemulti.newStoreWithoutGiga (cascading — fewer CMS recreations)
-16,483 MB  btree.NewFreeListG            (cascading — fewer cachekv stores)
-14,798 MB  sync.newIndirectNode          (cascading — fewer sync.Map allocs)
-11,050 MB  state.(*DBImpl).Snapshot      (cascading — less snapshot overhead)
-10,969 MB  sync.newEntryNode             (cascading — fewer sync.Map entries)
 -9,964 MB  vm.NewEVM                     (cascading — less GC pressure)
 -9,241 MB  multiversion.NewVersionIndexedStore

pprof -top -diff_base highlights (CPU)

 -4.85s  runtime.memclrNoHeapPointers (fewer allocs to zero)
 -2.37s  syscall.syscall              (less CGO overhead)
 -1.71s  syscall.syscall6
 -1.46s  runtime.heapBitsSmallForAddr (less heap bookkeeping)
 +3.19s  runtime.usleep               (workers idle more — finished faster)
 +1.61s  runtime.kevent               (same — less work to do)

Why safe

• Within a single Context, MultiStore(), GasMeter(), StoreTracer() are stable
• KVGasConfig() / TransientGasConfig() return constants
• WithMultiStore and WithGasMeter are the only methods that change gaskv inputs — both invalidate
• Map is reference type: copies share cache, but invalidation allocates a fresh map
• Thread safety: each OCC worker has its own Context

Test plan

• go test github.com/cosmos/cosmos-sdk/types -count=1 — pass
• go test ./giga/tests/... -count=1 — pass (all 14 giga tests)
• go test ./sei-cosmos/store/cachemulti/... -count=1 — pass
• gofmt -s -w clean
• Benchmark 2000+ blocks with heap profiling

🤖 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 11, 2026, 1:20 PM

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 56.67%. Comparing base (d52d4a1) to head (de7b59d).

Additional details and impacted files

@@                      Coverage Diff                      @@
##           perf/lazy-cachemultistore    #2808      +/-   ##
=============================================================
+ Coverage                      52.28%   56.67%   +4.39%     
=============================================================
  Files                           1030     2031    +1001     
  Lines                          85199   165964   +80765     
=============================================================
+ Hits                           44546    94065   +49519     
- Misses                         36523    63657   +27134     
- Partials                        4130     8242    +4112     

Flag	Coverage Δ
sei-chain	41.53% <100.00%> (?)
sei-cosmos	48.15% <100.00%> (+0.02%)	⬆️
sei-db	68.72% <ø> (ø)
sei-tendermint	58.08% <ø> (ø)

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

Files with missing lines	Coverage Δ
sei-cosmos/types/context.go	94.48% <100.00%> (+19.19%)	⬆️

... and 1142 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.

[pdrobnjak]

This PR was auto-merged by GitHub during a stack reorder (branch force-push made head appear merged into base). No code was actually merged to main. This PR has been superseded — see the restacked PR chain starting at #2849.