Repository navigation
perf: establish the simpler service source baseline - #2
Merged
Merged
Conversation
load_include serves readers outside the parse, such as join comparing a target the parse never read (docs/20 § 2). It composes through _compose_include, which stored the decoded text in raml.source_texts whenever retain_text was on, contrary to its documented contract that such reads enter none of the parse's caches. A post-parse write could grow the read set of an already built snapshot. Add a store_text switch to _compose_include; load_include passes False.
RAML 1.0 deprecates the schemas: table and the schema: facet. The decoders that accept those spellings now record each accepted use in Raml.syntax_aliases, keyed by the declaring entity's ID (docs/04 § 5, docs/05 § 3). The record holds the authored URI, the key's position and the spelling itself; it keeps no YAML node (docs/13 § 3). deprecated-schemas reads those records instead of walking every declaration's retained tree, so it is no longer a source-sensitive rule (docs/18 § 6). No rule in the default set now requires retained trees; the language service's snapshots therefore parse with text alone unless an opt-in style rule is enabled (docs/21 § 2). The retired docs/15 paragraph proposed exactly this recording. The three opt-in style rules that do read trees are named in docs/18. Bench (ab 2b6502e, 3 rounds): hover time +27.1% (noise 3.3%), peak +16.6%, kept +4.0%; inlays time +21.9% (noise 4.8%), peak +12.9%, kept +3.3%. The snapshots no longer retain parse trees, so the first hover or inlay per snapshot composes the queried file on demand, about a third of a parse of it; repeated queries are cached. Memory: large/service peak -20.5%, kept -24.4%; effective-types kept -27.5%. Accepted with a revisit note in docs/15 § 2.
…lection Folding and selection composed the buffer's whole document on every request (docs/21 § 4). The workspace now composes a file's current text once and holds the tree, invalidated with the buffer on change, close or a disk change; Workspace.source answers both request kinds. The text-based folding_ranges and selection_ranges keep their one-shot, per-call composition. The source-structure reach test also counts compose calls: each measured pass composes the text once, shared by its fifteen requests. Bench (ab 2b6502e, 3 rounds): no existing workload moves beyond its noise; the hover and inlays first-use deltas belong to the preceding commit and are documented in docs/15 § 2. New source-structure workload, linearity over 3 repeats: time 1.004, peak 0.998, retained 1.000. See docs/reports/2026-10-10/service-source-simple-rework.md.
The parse-side compatibility recording (docs/18 § 6) only runs on `schemas:` and `schema:`, which no bench corpus used, so the parse workloads never measured it. Every other `large` library now declares its whole table under the `schemas:` alias (a fragment takes one table or the other), every other inherited type uses the `schema:` facet, the `large` root carries a `schemas:` table of its own, and `endpoints` declares its payload type in a `schemas:` table with every fourth body typed by `schema:`. A reach test pins that both corpora record both spellings. Bench (ab 2b6502e, 3 rounds, on the spelling-carrying corpora): time within noise on large (all eight configurations), endpoints and small (parse, unwrap, unwrap+lint). Allocation peak and kept move +0.4 to +0.8 percent on the configurations that hold the model — the syntax_aliases records themselves, about 1.3 thousand of them, roughly 120 KB on large. large/service: time noise, peak -20.1%, kept -24.0%.
Merge the portable benchmark and profiling plan without changing the simpler production implementation. Record the master-controlled mixed-request result: service-session time -49.8% beyond 4.4% noise, peak +2.1%, kept +26.8%; source-first time -52.7% beyond 18.9% noise, peak -10.9%, kept +26.9%. Duplicate source ownership blocks acceptance and is the next isolated fix. The Windows gate passes: 5824 passed, 68 skipped, 1 xfailed. Measurement-only PR 1 passed all CI checks and merged as d2b8a02.
Use one workspace-owned plain tree for compatible hover/inlay and folding/selection inputs. Cache failures and include depth/backend policy in compatibility. Snapshots borrow the owner weakly; stale input misses cannot replace current entries, while unrelated edits and unchanged dependencies can reuse source. Preserve full-source originals and separate JSON normalization. Against 94ebc60, service-session/unwrap time -12.5% beyond 2.8% noise, peak -26.5%, kept -16.9%; source-first time within 22.0% noise, peak -22.7%, kept -16.9%. All three relevant linearity checks pass. The Windows gate passes: 5838 passed, 68 skipped, 1 xfailed. The user accepts the remaining cold-query and post-query allocation trade-off for merge-then-recover.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Acceptance
The remaining trade-offs are accepted for merge-then-recover. Against measurement-only master d2b8a02 (rewritten as e81d361):
Verification
Full Windows gate: 5838 passed, 68 skipped, 1 xfailed; strict mypy, Ruff and formatting pass. Source-structure and both mixed linearity checks pass. Fourteen focused source-ownership cases cover request order, errors, policy, old snapshots and release. Linux, pure-Python YAML, TCK and full CI must pass before merge.
Plan and evidence
See docs/research/2026-10-10/service-recovery-plan.md and docs/reports/2026-10-10/service-baseline-integration.md. The untracked handoff is excluded.