Documentation
¶
Overview ¶
Package realcorpus is the pinned real-repo manifest for the `publish` benchmark mode: the freshly-built Go `codegraph` binary, run over real, publicly available source trees rather than the synthetic PERF-02/INDX-06 corpus (tools/bench/gencorpus), to produce absolute throughput / query-latency / peak-RSS numbers for publication.
Every entry MUST carry a commit SHA — never just a branch or tag name — so a published measurement stays reproducible run-to-run and machine-to-machine (CONTEXT.md D-04's reproducibility discipline, extended from the synthetic corpus to real repos). This reuses the exact provenance shape established by tools/spike/testdata/ATTRIBUTION.md (recoverable via `git show e5da8e7:tools/spike/testdata/ATTRIBUTION.md`) and by this project's own testdata/golden/README.md "Corpus (D-06a)" table, which already pins one of these two entries (weft-go) for the Phase-3 golden test suite.
This package performs no network I/O: Resolve only reports a local checkout path when one already exists (via an env var override or the conventional sibling-checkout directory) or returns ErrNeedsClone otherwise. Any actual shallow clone is the caller's responsibility (tools/bench/runner, or the bench.yml CI job), and MUST be pinned to exactly the entry's CommitSHA — never at HEAD (V5: no attacker-controlled or unpinned ref is ever resolved into a subprocess argument).
Index ¶
Constants ¶
This section is empty.
Variables ¶
var ErrNeedsClone = errors.New("realcorpus: no local checkout found; shallow-clone SourceURL at CommitSHA")
ErrNeedsClone is returned by Resolve when no local checkout of an Entry is present via either its EnvVar override or its conventional sibling-checkout directory. The caller (tools/bench/runner or a CI job) is responsible for shallow-cloning Entry.SourceURL and checking out exactly Entry.CommitSHA — never HEAD — into a scratch directory when this is returned. realcorpus itself never touches the network.
Functions ¶
This section is empty.
Types ¶
type Entry ¶
type Entry struct {
// Name is a short, filesystem-safe identifier used for cache
// directory names and reported bench.Metrics.Repo values.
Name string
// SourceURL is the canonical repo URL (https, no ".git" suffix)
// a caller should shallow-clone if no local checkout is found.
SourceURL string
// Tag is the human-readable release ref the pin corresponds to
// (e.g. "v2.1.6"). May be empty when the pin was taken at a
// repo's default-branch HEAD at capture time rather than a
// tagged release — CommitSHA is authoritative either way.
Tag string
// CommitSHA is the exact pinned commit every `publish`-mode run
// MUST resolve to. This is the one field that makes a published
// measurement reproducible (see package doc).
CommitSHA string
// License is the repo's SPDX-style license identifier at
// CommitSHA.
License string
// SelectionRule documents exactly what subset of the repo's tree
// is indexed, mirroring tools/spike/testdata/ATTRIBUTION.md's
// "Selection" field discipline.
SelectionRule string
// QueryTerms is the fixed query set (D-05) run against this
// entry's index for the query-latency metric: real symbol names
// confirmed present in the source tree at CommitSHA, so every
// measured query is a genuine (non-empty) lookup rather than an
// arbitrary string that may or may not match.
QueryTerms []string
// EnvVar, if set, is checked first by Resolve: an operator can
// point it at an existing local checkout to skip cloning
// entirely.
EnvVar string
// SiblingDir, if set, is a conventional local checkout path
// (relative to this repo's parent directory) Resolve checks
// second, before reporting ErrNeedsClone.
SiblingDir string
}
Entry is one pinned real-repo benchmark input. Every field below (aside from Tag, which may legitimately be empty for a pin taken at a non-tagged commit) is required so a published measurement is always traceable back to an exact source, license, and file set — the same discipline tools/spike/testdata/ATTRIBUTION.md established.
func Corpora ¶
func Corpora() []Entry
Corpora returns the fixed, pinned real-repo manifest for the `publish` benchmark mode. Order is deliberate: compact-and-already- vendored-adjacent first, larger-scale second (per CONTEXT.md D-04's "plus a few larger real repos" + this plan's "at least one larger real repo pinned by commit SHA to exercise scale" acceptance criterion).
func (Entry) Resolve ¶
Resolve returns a local filesystem path containing e's source tree, checked in this order:
- e.EnvVar, if set and pointing at an existing directory (an operator override).
- e.SiblingDir, if set, resolved relative to this repo's parent directory (the "../weft next to this repo" convention).
Resolve does not verify the checkout is actually pinned at e.CommitSHA — callers that need that guarantee (as tools/bench/runner.resolveOrClone's pinnedAt check does for `publish`-mode measurement, WR-02 Phase 8 re-review) should check it themselves; this function's job is only path discovery, kept deliberately simple so it has no network or git-plumbing surface of its own beyond a single local `git rev-parse --show-toplevel` to locate the sibling directory.
Source Files