Share a lazy source index across cache hits

This commit is contained in:
2026-11-22 09:42:06 +00:00
parent 2429f0ae6a
commit 80148ef7e4
4 changed files with 65 additions and 4 deletions

View File

@@ -10,6 +10,7 @@ import (
"path"
"path/filepath"
"strings"
"sync"
"time"
"gitea.maximumdirect.net/eric/notarius/internal/core/source"
@@ -70,6 +71,8 @@ type RunInput struct {
stepID string
references map[referenceTargetKey]contracts.ReferenceSet
referenceReuseEligibility map[referenceTargetKey]bool
// Shared by operation decoders, but built only when a cache hit needs it.
cacheSourceIndex func() source.DocumentIndex
}
type RunOutput struct {
@@ -182,6 +185,9 @@ func (r *Runner) Run(ctx context.Context, input RunInput) (output RunOutput, err
if err := source.ValidateDocument(doc); err != nil {
return failOutput(output), fmt.Errorf("validate source document: %w", err)
}
input.cacheSourceIndex = sync.OnceValue(func() source.DocumentIndex {
return source.NewDocumentIndex(doc)
})
if err := writeDebugTimed(debugRecorder, "source/output.json", debugTimedEnvelope{
Stage: "source",

View File

@@ -3,6 +3,7 @@ package pipeline
import (
"context"
"fmt"
"sync"
"gitea.maximumdirect.net/eric/notarius/internal/core/source"
"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
@@ -12,7 +13,11 @@ func artifactProducerCache(input RunInput, typed *preparedTypedLane, class contr
if input.Cache == nil || class != contracts.ExecutionClassLLMBacked {
return nil
}
index := source.NewDocumentIndex(doc)
sourceIndex := input.cacheSourceIndex
if sourceIndex == nil {
// Standalone operation construction can also defer indexing until decode.
sourceIndex = sync.OnceValue(func() source.DocumentIndex { return source.NewDocumentIndex(doc) })
}
return &contracts.ProducerCache[any]{
Store: input.Cache,
Encode: typed.codec.encode,
@@ -26,9 +31,12 @@ func artifactProducerCache(input RunInput, typed *preparedTypedLane, class contr
if err != nil {
return nil, err
}
for _, ref := range refs {
if err := index.ValidateRef(ref); err != nil {
return nil, fmt.Errorf("cached artifact source reference: %w", err)
if len(refs) > 0 {
index := sourceIndex()
for _, ref := range refs {
if err := index.ValidateRef(ref); err != nil {
return nil, fmt.Errorf("cached artifact source reference: %w", err)
}
}
}
}

View File

@@ -6,6 +6,7 @@ import (
"errors"
"reflect"
"strings"
"sync/atomic"
"testing"
"gitea.maximumdirect.net/eric/notarius/internal/core/source"
@@ -13,6 +14,47 @@ import (
"go.maximumdirect.net/promptkit"
)
func TestRunnerChecksCurrentSourceOnConcurrentCacheHits(t *testing.T) {
_, cache := filesystemProducerCache(t)
var calls atomic.Int32
engine := nativeProducerEngine(t, producerClientFunc(func(context.Context, promptkit.GenerateRequest) (*promptkit.GenerateResponse, error) {
calls.Add(1)
return &promptkit.GenerateResponse{Content: `{"value":1}`}, nil
}))
const chunkCount = 4
prepared := preparedConcurrentPipeline(t, chunkCount)
for i := range prepared.Steps[0].lanes {
lane := &prepared.Steps[0].lanes[i]
if lane.resolved.ArtifactKind != "test/notes" {
continue
}
lane.resolved.ExtractExecutionClass = contracts.ExecutionClassLLMBacked
lane.typed.cacheEvidence = func(any) ([]source.SourceRef, error) {
return []source.SourceRef{{SourceID: "source", StartUnitID: 1, EndUnitID: 1}}, nil
}
lane.typed.extract = func(ctx context.Context, _ any, request contracts.TypedExtractionRequest[any]) (erasedTypedResult, error) {
initial := nativeProducerInitial()
initial.Inputs["source"] = promptkit.Inline(string(request.Chunk.Content))
value, outcome, err := RunProducer(ctx, engine, request.Operation, initial,
func(context.Context, *promptkit.RunResult) (*any, []string, error) {
var value any = codecNotes{Items: []string{"accepted"}}
return &value, nil, nil
}, nil, nil)
return erasedTypedResult{Value: value, Outcome: outcome}, err
}
}
for run := 0; run < 2; run++ {
// The filesystem store is concurrency-safe; its counting test wrapper is not.
output, err := New().Run(context.Background(), RunInput{Prepared: prepared, Cache: cache.inner, RawInput: []byte("input"), Workers: chunkCount})
if err != nil || len(output.normalizeOutputs) != 2 {
t.Fatalf("run %d: outputs=%d error=%v", run, len(output.normalizeOutputs), err)
}
}
if calls.Load() != chunkCount {
t.Fatalf("cache hits executed model work: calls=%d", calls.Load())
}
}
func TestRunnerReusesAcceptedPredecessorAfterDownstreamFailure(t *testing.T) {
_, store := filesystemProducerCache(t)
probe := newRunnerProbe()