perf(graph): build prepared relationship indexes with FxHash maps by jonathanong · Pull Request #833 · jonathanong/no-mistakes

jonathanong · 2026-09-05T06:39:51Z

Summary

Build PreparedRelationshipIndex with FxHashMap/FxHashSet (local interned keys, same policy as fx.rs) and replace per-name BTreeSet inserts with Vec + one sort/dedup. Public display order and alias grouping are unchanged.

Local Criterion (source of truth)

Machine: macOS aarch64, rustc 1.96.0, --sample-size 20, NO_MISTAKES_BENCH_SHARD=query. Same machine before/after. CodSpeed is not used as a gate.

bench	size	before (mean)	after (mean)	change
`index_construction`	4096	3.73 ms	3.41 ms	−22% n.s.
`index_construction`	16384	20.19 ms	10.78 ms	−45%
`scoped_projection`	4096	4.31 ms	3.85 ms	−8%
`scoped_projection`	16384	30.01 ms	23.21 ms	−30%
`unscoped_projection`	4096	1.14 ms	1.24 ms	+10% n.s.
`unscoped_projection`	16384	5.29 ms	4.24 ms	−19%

Shepherd Journal

CodSpeed extract regression reported different runtime environments; local Criterion showed index_construction[16384] −45%. codecov/project/rust: no coverage on head while Rust tests inProgress.
First-look CodSpeed comment already minimized; different runtime environments; local Criterion index_construction −45%.
Codex quota comments IC_kwDOSaSRXM8AAAABStKfzQ/IC_kwDOSaSRXM8AAAABStKhDA and Sourcery budget review PRR_kwDOSaSRXM8AAAABMTHn8g are bot-limit notices, not code findings; minimize only.
Accepted Codex P2 PRRT_kwDOSaSRXM6fhSsD: shrink public_names after the unique-node fill so raw-edge-count reservation is not retained. Rejected P2 PRRT_kwDOSaSRXM6fhXEp: bucket sort/dedup already unique the stored alias vectors; per-bucket sets would hash Node on every push on the construction path this PR is speeding up, and the duplicate clones are a temporary spike. Codex quota comments IC_kwDOSaSRXM8AAAABStOd_g/IC_kwDOSaSRXM8AAAABStQZvA are bot-limit notices; minimize only.
Accepted Codex P2 PRRT_kwDOSaSRXM6fhZii: restore SipHash HashSet in project_first_seen because projected outputs are public report edges from repo source, not interned analysis keys (fx.rs).
Codex quota comment IC_kwDOSaSRXM8AAAABStTOVQ is a bot-limit notice; minimize only.
Rejected Codex P2 PRRT_kwDOSaSRXM6fhdHG: grouped_nodes/nodes_by_name keys are public display strings of interned graph nodes, the same local-path universe fx.rs already hashes with FxHash (PathSet). That is analyzed-repo self-DoS, not network input. SipHash here would undo the measured index-construction win. codecov/project/rust "No coverage information found on head" while native/Rust jobs are still inProgress is upload lag.

Use FxHash maps/sets and a single sort/dedup per public-name bucket so index construction does not pay SipHash or BTreeSet insert on local keys.

Reserve by raw edge count for the first pass, then drop spare HashMap buckets so duplicate enqueue sites do not keep construction capacity in the retained PreparedRelationshipIndex.

project_first_seen dedupes public report edges whose paths and job names come from repository source. FxHash is reserved for interned analysis keys; restoring HashSet avoids HashDoS on crafted names.