Skip to content

Graphsight - #7

Merged
Kcodess2807 merged 2 commits into
mainfrom
graphsight
Jul 28, 2026
Merged

Graphsight#7
Kcodess2807 merged 2 commits into
mainfrom
graphsight

Conversation

@Kcodess2807

Copy link
Copy Markdown
Owner

No description provided.

A live ingest of pallets/click (15 PRs, 212 typed edges) showed --files had
made things worse: giving PRs their real TOUCHES edges pushed 7 of 15 past
MAX_DEGREE, so the hub filter skipped them entirely and traversal collapsed to
one hop. Wide diffs are exactly the PRs worth traversing.

Hubness is now judged per relation. A relation fanning out past MAX_DEGREE from
one node carries no signal and is dropped; the node's other edges survive. A
node whose every relation fans out that wide is still skipped, so repo:, which
TOUCHES all 45 neighbours, stays excluded.

On that live graph, PR #3695 (degree 18, previously invisible) now expands to
AUTHORED -> Rowlando13, RESOLVES -> #3099, REVIEWED -> davidism, and the same
query goes from 1 hop to 15.
…s, serial fetch

100-PR live ingests of fastapi/fastapi and pallets/click found four problems.

REVIEWED was reading `requested_reviewers`, which is a *pending* review request
that empties out once someone actually reviews. So it both missed real reviewers
and credited people who never opened the code — fastapi produced 0 REVIEWED
edges across 100 PRs. Now read from /pulls/N/reviews (--reviews), deduped, with
self-approval excluded: fastapi 0 -> 10, click 10 -> 13.

The per-PR file list was fetched as a single page, silently dropping everything
past the first 100 files on a wide PR. Now paged.

parse_ts read a tz-less timestamp as local time, shifting ages by the machine's
UTC offset. Now assumed UTC.

Per-PR detail fetches were serial and dominated wall time (110s of 143s). Now
concurrent, and paging stops on a short page instead of paying for an empty
round trip: 100 PRs went from 142.8s to 98.4s while fetching twice as much.

Also measured and unchanged: p50 18ms / p95 21ms query latency on a 1274-node
graph, 115 q/s across 32 threads with zero errors, and no crash on any hostile
payload (null user, 2000 files, 50k-char query, emoji, null byte). A Cypher
injection attempt left all 1274 nodes intact — the parameterized queries hold.

51 engine tests.
@vercel

vercel Bot commented Jul 28, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
trace-rag Ready Ready Preview, Comment Jul 28, 2026 1:49pm

@Kcodess2807
Kcodess2807 merged commit 90e1959 into main Jul 28, 2026
10 of 14 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant