Code search's next page can repeat the last one's file

What happened

searching_code in crates/knotless-serve/tests/web.rs fails now and then, on the spaces net. The second page of needle_&limit=1 repeats the first page's repository:

thread 'searching_code' panicked at crates/knotless-serve/tests/web.rs:3527:5:
assertion `left != right` failed

That line is assert_ne!(next["files"][0]["repo"], first["files"][0]["repo"]).

It failed once and then passed at one head of knotless-serve answers repositories' sites on their own domain, and failed twice in a row at its next head, 8c5a70cce2b3. Neither head touches code search.

The cause is that the cursor is an offset into an ordering that is not stable:

Code::search (crates/knotless-serve/src/code.rs) asks Zoekt for MaxDocDisplayCount = skip + limit files, and web/search.rs skips skip of them.
Zoekt sorts files by score alone. Files with equal scores, as the test's two repositories' needle_ files are, come back in whatever order its shards finish.
So the first page (1 file asked for) and the second (2 asked for, the first skipped) can rank the same two files differently.

How to see it

Run searching_code a few times on the spaces net, or read the test log of that pull at 8c5a70cce2b3. It's quarantined meanwhile in harness/quarantine.json.

What you expected

Each page continues the last, with no file shown twice. That takes a stable order: score, then repository and path. A tie at a page's edge is settled only by seeing every file that ties, so a fix asks past upto (to DEEPEST, or until the score drops), then sorts and slices. Measure what that costs first.

Where

knotless-serve's dev.knotless.searchCode, as main has it at 54981405.

chadtmiller.com/knotless Switch repository Loading…