Record algorithmic scaling opportunities

This commit is contained in:
2026-11-22 09:28:59 +00:00
parent fdaf39162b
commit e823baf3af

View File

@@ -10,14 +10,18 @@ Integration follow-up baseline: `fd134e85e54c6e0f74aab5c77df5bada7c1b1e6a`.
The follow-up baseline adds only the initial review findings; application code
is unchanged between these revisions.
Algorithmic follow-up baseline: `fdaf39162b46b528fddcd8a4cad18a09e1daf661`.
Only this findings document changed after the application baseline above.
This review followed the 46 stages in [the implementation plan](implementation.md)
against the [native migration target](promptkit.md), current owning code,
focused tests, maintained examples, and canonical documentation. The migration
is substantially complete. Findings 1–2 come from the initial review. Findings
3–5 come from a second, stage-ordered pass through the interaction of the native
migration with existing domain, validation, and diagnostic behavior. None calls
for another compatibility layer or application execution abstraction. This
document records findings, not implemented fixes.
3–5 come from an integration pass. Findings 6–9 and the complexity note on
Finding 1 come from a stage-ordered algorithmic pass. The latter are refactoring
opportunities for consideration, not measured production bottlenecks or new
performance guarantees. None calls for another compatibility layer or application
execution abstraction. This document records findings, not implemented fixes.
## Finding 1: Retain Genuine Native Timeouts During Concurrent Error Selection
@@ -69,6 +73,18 @@ application error taxonomy is needed.
the existing native timeout fixture where useful. Run the pipeline tests and
their concurrency/race checks.
### Algorithmic Follow-up: Select A Minimum Instead Of Sorting
With `E` collected errors, the current selector comparison-sorts all eligible
records by `(stage, lane, chunk)` solely to return the first. It performs
`O(E log E)` comparisons; stable sorting can also incur additional superlinear
data movement. After eligibility classification, a minimum scan takes `O(E)`
comparisons and constant auxiliary space; error-chain classification costs remain
separate. Keep the first occurrence on equal keys, matching stable-sort behavior.
This can accompany the correctness
fix without changing error precedence or cancellation semantics. Its practical
impact is smaller than the document/cache indexing findings below.
## Finding 2: Narrow The Documented Empty-Validator-Chain Guarantee
Priority: P2. Relevant stages: 11, 38, 39, and 42.
@@ -281,6 +297,160 @@ assessment/capture error types. Do not bridge native errors into these retired
sentinels or add aliases for compatibility. Verify no current callers remain and
run contracts, pipeline, CLI, and module integration tests after removal.
## Finding 6: Use Indexed Source Validation In The Chunk-Plan Hot Path
Priority: P2. Relevant stages: 16, 24, 28–29, and 37.
### Evidence And Complexity
The supporting [source validation code](../../internal/core/source/validation.go)
predates this migration, but the new chunk-plan assessment/cache path repeatedly
uses it. `ValidateDocument`, lines 95–101, validates each unit's self-reference
with `ValidateRef`. That helper resolves both endpoints through the linear
`UnitIndex` scan. For a valid document of `N` units, the second validation pass
alone makes `2 * (1 + ... + N) = N(N+1)` unit-ID comparisons: `Theta(N^2)` work.
[`ValidateChunkPlan` and `MaterializeChunkPlan`](../../internal/core/source/chunk_plan.go)
also repeatedly scan the source for each of `C` ranges. A materialization thus
performs `O(N^2 + CN)` source-position work in addition to field checking,
annotation processing, metadata ownership, and payload encoding.
The pipeline's
[`validateAndMaterializeChunkPlan`](../../internal/framework/pipeline/chunk_validation.go)
validates once explicitly and then calls a materializer that validates again.
The [new plan path](../../internal/framework/pipeline/runner_chunk_plan.go)
invokes this during assessment, cache encode/decode, and final delivery. Cached
plans therefore retain this deterministic cost even without inference.
### Potential Refactoring
Use the existing immutable `source.DocumentIndex` for bulk endpoint lookups.
Build/reuse a unit-position map while validating the document, then check its
self-references in linear expected time. Share resolved positions through a
small private validation/materialization helper so materialization does not
rescan endpoints or repeat the same validation unnecessarily. Keep standalone
public validation/materialization entry points correctly validating their inputs.
The attainable position-lookup cost is `O(N + C)` expected work with `O(N)`
index space. Field/payload inspection, annotation canonicalization, metadata
ownership, and JSON encoding remain separate costs. This does not promise linear
total output size: permitted overlapping generic ranges can themselves require
superlinear copied output. Repeated materialization of accepted plans can be
reduced where already-owned checked chunks are available, without retaining a
whole-run ledger or removing necessary current-source checks.
### Preserve And Verify
Preserve document-order semantics for nonconsecutive/nonmonotonic unit IDs,
duplicate/missing/wrong-source/reversed-reference rejection, annotation handling,
deep ownership, and deterministic chunk IDs. Generic plans still permit their
existing gaps/overlaps; D&D coverage rules remain with scene interpretation and
configured validators. No stricter plan format or new size limit is needed.
Use the existing source validation/materialization, chunk-plan operation/cache,
and chunk-map tests. Add only missing behavioral regressions; source inspection
or optional synthetic operation counts can verify the scaling change without
wall-clock thresholds or production data.
## Finding 7: Share A Lazy Source Index Across Artifact Cache Decoders
Priority: P2. Relevant stages: 11, 21–22, and 27–28.
### Evidence And Complexity
[`artifactProducerCache`](../../internal/framework/pipeline/runner_cache.go),
line 17, eagerly creates a full `DocumentIndex` for each model-backed operation
whenever a store collaborator exists. It does so before the producer helper
checks bypass/refresh mode, incoming reuse eligibility, or whether a module takes
a legitimate no-model branch. Cold misses also never use the resulting index.
For `J` such operations against `N` source units, this adds `Theta(JN)` index
construction work. With a fixed number of lanes and chunks proportional to
transcript length, that cache-specific term becomes quadratic in `N`.
Concurrent operations can retain `O(min(J,W)N)` cache index entries while waiting
for model work, where `W` is the active worker bound. This is separate from
codec decoding and domain/validator input ownership costs.
### Potential Refactoring
Supply the cache decoders with one run-scoped, immutable source-position index.
Acquire it lazily only when a decoded hit actually needs evidence checks, using
ordinary concurrency-safe initialization. The cache-specific work becomes
`O(N + R)` for all hits, where `R` is the total number of checked cached source
references, with `O(N)` shared index space. Runs with no such checks need not
construct the cache index at all.
This is a small in-memory collaborator, not another persistent cache or source
identity framework. Keep the native message-content key, modes, hit validation,
fallback eligibility, and no-semantic-reassessment policy. Finding 3's scene
identity adaptation is a separate usability concern; shared indexing must not
replace it. Other intentional domain/validator snapshots need not be weakened.
### Preserve And Verify
Retain wrong-source/missing/reversed citation rejection on hits, normal misses,
refresh/bypass/ineligible/no-model behavior, and canceled-hit delivery. Exercise
concurrent hits with the existing cache/runner tests and race checks. Do not add
a public index-build-count contract or a timing-based release gate.
## Finding 8: Index Debug Sensitivity Paths Once
Priority: P3. Relevant stages: 30–31 and 33.
### Evidence And Complexity
[`newRunEvidenceRecorder`](../../internal/cli/run_capture.go), lines 64–91,
collects all `H` historical configuration paths, then scans that full list for
each of `P` profiles/backends across root and collection resources. Prefix
matching therefore performs `O(PH)` work, plus path-string processing. With a
roughly fixed number of fields per resource, `H` grows with `P`, yielding a
quadratic resource-count term.
The caller correctly constructs this recorder only when capture is enabled.
This finding concerns debug startup for large native configurations, not an
unconditional cost on ordinary runs.
### Potential Refactoring
Walk metadata paths once and group relevant sensitive suffixes by their exact
resource prefix. Then associate those groups with root/qualified resource IDs in
one resource traversal. This takes `O(H + P)` path/resource visits, plus processing
the path characters and retained masks, rather than rescanning every unrelated
configuration path for every resource.
Retain native sensitivity classification, correct JSON-pointer segment unescaping,
root/collection distinctions, nested maps/arrays, and nonmutation of native
snapshots. Do not replace exact metadata with credential-like string searches or
broaden omission policy. Use the existing configuration/native capture tests;
add a mixed-resource case only if needed to verify equivalent mask routing.
## Finding 9: Traverse Dense Chunk Results Without Sorting Their Keys
Priority: P3. Relevant stages: 11, 25, and 37.
### Evidence And Complexity
[`finalizeLaneExtract`](../../internal/framework/pipeline/runner_concurrent.go),
lines 333–349, copies result-map keys to a new slice and sorts them before
assembling the lane. For `C` chunks this costs `O(C log C)` comparisons and
`O(C)` temporary key storage per lane.
The runner's chunks are already assigned contiguous zero-based indexes by
`MaterializeChunkPlan`; successful finalization occurs only after all of that
lane's jobs return. The required order is therefore known, not arbitrary.
### Potential Refactoring And Verification
Traverse the known ordinal range directly for `O(C)` expected map-lookup work
and no temporary key slice, or use an indexed result slice if it simplifies
storage. Keep all artifact/diagnostic ordering, fallback counts, eligibility,
failure handling, cancellation draining, and ready-job scheduling unchanged.
No new chunker restriction is required: dense ordinals are already produced by
the existing materializer, regardless of transcript unit-ID spacing.
The current runner concurrency and ready-job tests cover out-of-order completion
and public result order. Reuse those checks; do not couple tests to a particular
map/slice representation or add another scheduler.
## Review Coverage And Verification
Owning boundaries were inspected in implementation order:
@@ -295,7 +465,7 @@ Owning boundaries were inspected in implementation order:
| 35–37 | Removal inventory, maintained examples, assembled recovery and subprocess coverage. |
| 38–46 | Successor ADRs and historical-body preservation, canonical references, operations/consumer/migration/internal/integration guides, orientation and final acceptance evidence. |
Checks completed during review:
Checks completed during the initial review:
- `GOWORK=off GOPROXY=off go test ./...` passed.
- `GOWORK=off GOPROXY=off go vet ./...` passed.
@@ -308,7 +478,7 @@ Checks completed during review:
- The migration diff contains no changes to model-facing JSON schemas or D&D
intrinsic codec schema files. Historical ADR decision bodies are retained;
supersession edits are confined to status descriptions.
- Temporary Go-overlay probes reproduced both findings without modifying
- Temporary Go-overlay probes reproduced Findings 1–2 without modifying
repository code or tests. Their deliberately failing assertions are review
evidence, separate from the passing default suite.
@@ -331,6 +501,40 @@ application code and dependencies were unchanged. Additional checks were:
These probes live only in a temporary review directory. No production code,
test, configuration, schema, or existing canonical documentation was changed.
The algorithmic pass followed the same stage order, inspecting actual loop/call
costs rather than treating nested-loop graph metadata as proof of a bottleneck:
- Stages 1–7: native loading/delegation, bounded invocation-selection reuse,
resource/input construction, and absence of a second application load.
- Stages 8–11: candidate serialization/isolation, complete-chain traversal,
native helper control flow, and runner/handoff state.
- Stages 12–24: fixed extraction mappings, bounded reconciliation preparation,
interval coalescing, indexed duplicate/group application, direct validators,
deterministic paths, and the source/materialization costs in Finding 6.
- Stages 25–29: constant-work ready-job dispatch, result/error ordering,
filesystem byte traversal, source-index lifetime, and reuse/auxiliary assembly.
- Stages 30–34: incremental capture, sensitivity-path routing, preflight,
command composition, diagnostics, and publication.
- Stages 35–46: retained/removal boundaries, example/assembled coverage, and
documentation consistency with the algorithms and selected policies above.
Explicitly selected behavior was not reopened: complete validation on every
candidate, fresh normalization application to fixed originals, native growing
history, content-only cache identity, no semantic reassessment on hits, the ready-
job scheduling policy, and self-contained debug records even when content repeats.
The earlier expressly recommended per-handoff output index was also left alone.
Sorting genuinely unordered citations/intervals and collision-safe indexed domain
grouping are not findings merely because they contain loops or comparisons.
Complexity claims in Findings 6–9 follow from the current source and call sites;
they are not throughput estimates. Position-map costs use ordinary expected
hash-map complexity; retained field/payload processing is accounted for separately.
The existing
offline suites for `internal/core/source`, `internal/framework/pipeline`,
`internal/framework/artifactcache`, `internal/framework/semanticreconcile`, and
`internal/cli` passed again. No production workload or performance data was needed,
and no new benchmark, timing assertion, or permanent gate was added.
No live provider inference was performed. This review does not establish
production model quality or performance. Intentional changes such as the hard
configuration/CLI cutover, native retry ownership, content-only cache identity,