Commit Graph
2 Commits
Author SHA1 Message Date
684de2e691 perf(query): use UNION of index seeks for ListUsers login equality (#12701)
<!-- CURSOR_AGENT_PR_BODY_BEGIN -->
# Which Problems Are Solved

Login V2 `ListUsers` lookups of the form username OR email (and email OR
phone) hash-join the full `users14` / `users14_humans` tables, then
filter. That cost grows with instance size (~40ms at 10k users, ~650ms
at 200k).

Top-level login-name equals already used an indexed join, but the same
login-name filter inside an `OrQuery` fell back to the slow view-based
exists query.

# How the Problems Are Solved

**TL;DR:** equality lookups on username, email, phone, or login name
become `INNER JOIN (UNION of indexed ID seeks) AS matches`. Mixed /
contains / name searches keep the old WHERE shape.

### Reviewer map

| Read first | Why |
| --- | --- |
| `internal/query/user_login_equality.go` | Typed filters + extractor +
UNION join |
| `internal/query/user.go` (`prepareUsersQuery`, constructors) | Wiring;
username/email/phone equals return `lowerEqualsFilter` |
| `cmd/setup/76.go` + `cmd/setup/76/*.sql` | Concurrent indexes for
existing deployments; drop invalid leftovers |
| `internal/query/projection/user.go` | Same indexes in `Init()` for new
databases |
| `internal/query/user_login_name_exists_test.go` | SQL-shape coverage |
| `internal/api/grpc/user/v2/integration_test/query_test.go` | Nested
username AND/OR user results |

### Query rewrite

- Extract one equals leaf, a flat `OrQuery` of those leaves, or the same
nested under `AndQuery`. Remaining AND conjuncts stay as WHERE. Equality
nested under `OrQuery` / `NotQuery` is left as WHERE.
- Login-name equals is a UNION arm that reuses `login_names3_users` SQL,
so it works inside OR as well as at the top level.
- EQUALS seeks use `LOWER(col)` for the index, then recheck `col =
value` so case-sensitive searches do not match extra rows.
EQUALS_IGNORE_CASE is unchanged.
- Empty OR arms are not skipped (that would change OR meaning); mixed
empty-OR falls back to WHERE. Login UI already omits empty arms.

### Indexes

- Setup 76: `CREATE INDEX CONCURRENTLY IF NOT EXISTS` on
`users14_username_lower_idx` `(instance_id, LOWER(username))` and
`users14_humans_phone_lower_idx` `(instance_id, LOWER(phone))`. Not
wrapped in a transaction.
- Before create, drop any leftover `pg_index.indisvalid = false` indexes
of those names (failed/cancelled `CREATE INDEX CONCURRENTLY`). `IF NOT
EXISTS` would otherwise skip rebuild.
- `Init()` creates the same indexes without `CONCURRENTLY` (empty
tables). Existing `users14_username` and email lower indexes are
unchanged.

# Additional Changes

- Unit tests assert UNION SQL for username OR email, email OR phone,
login name OR email, username OR username, nested `AndQuery(loginName,
org)`, and nested `AndQuery(Or(...), org)`; that contains / display-name
/ mixed OR / nested `OrQuery(AndQuery(...), ...)` do not UNION; that
email EQUALS rechecks the raw column; and that empty-OR is not
extracted.
- Setup 76 sqlmock coverage: missing `users14` skips work; valid/absent
indexes only create; invalid leftover username index is dropped then
both creates run.
- `TestServer_ListUsers_UsernameAndOr` checks live results: org AND
(alice OR bob) returns both users, sibling username AND is empty, nested
`AndQuery(Or(...), org)` returns both.

# Additional Context

- Base branch: `main`
- Extends the login-name equals rewrite from #12460 to
username/email/phone OR trees used by Login V2 search.

<!-- CURSOR_AGENT_PR_BODY_END -->

<div><a
href="https://cursor.com/agents/bc-1f043b1c-6168-45ca-9e9e-c18451994207?cursor_ref=pr_footer&cursor_cta=open_in_web"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://cursor.com/assets/images/open-in-web-dark.png"><source
media="(prefers-color-scheme: light)"
srcset="https://cursor.com/assets/images/open-in-web-light.png"><img
alt="Open in Web" width="114" height="28"
src="https://cursor.com/assets/images/open-in-web-dark.png"></picture></a>&nbsp;<a
href="https://cursor.com/background-agent?bcId=bc-1f043b1c-6168-45ca-9e9e-c18451994207&cursor_ref=pr_footer&cursor_cta=open_in_cursor"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://cursor.com/assets/images/open-in-cursor-dark.png"><source
media="(prefers-color-scheme: light)"
srcset="https://cursor.com/assets/images/open-in-cursor-light.png"><img
alt="Open in Cursor" width="131" height="28"
src="https://cursor.com/assets/images/open-in-cursor-dark.png"></picture></a>&nbsp;</div>

---------

Co-authored-by: Cursor Agent <[email protected]>
Co-authored-by: Silvan <[email protected]>
2026-09-08 13:11:16 +00:00
318acd4bc3 perf(query): speed up ListUsers login name equality filters (#12460)
<!-- CURSOR_AGENT_PR_BODY_BEGIN -->
## Summary

`ListUsers` with `LoginNameQuery` + equals/equals-ignore-case was very
slow on large orgs (e.g. ~330k users) because the generated SQL filtered
the `projections.login_names3` **view** on the computed
`login_name_lower` expression and correlated that subquery on
`users14.id`. Postgres nested-looped every user and never used
`login_names3_users_search (instance_id, user_name_lower)`.

This change rewrites the **query planner** for that hot path: when an
equals/equals-ignore-case login-name filter is present (and not under
`OR`/`NOT`), the user list query **INNER JOINs** an indexed matches
subquery instead of filtering via the view expression. The matches SQL
mirrors `user_by_login_name.sql` (`user_name_lower` / domain paths +
`preferred` / `is_primary`).

Non-equals methods and OR combinations keep the previous view-based
filter so semantics stay unchanged.

Also adds a k6 use case that mirrors login v2 discovery
(`loginNameQuery` EQUALS_IGNORE_CASE + `organizationIdQuery`, `limit:
2`):

```bash
cd benchmark
make users_by_login_name USER_AMOUNT=100000 VUS=10 DURATION=60s
```

## Approach

1. `NewLoginNameSearchQuery` for equals / equals-ignore-case returns a
marker `loginNameEqualsFilter` (other methods unchanged).
2. `prepareUsersQuery` extracts that marker when safe, then:
- builds the usual `sq.SelectBuilder` **without** the login-name view
predicate
- adds `JoinClause` to `user_login_name_matches(.sql)` /
`_case_sensitive.sql` as `login_name_matches`
- keeps metadata JOIN/`DISTINCT` only when metadata filters are present
(same as before)
3. Embedded SQL files under `internal/query/` for the matches subquery.

Local smoke against ~330k synthetic users: baseline ~1392ms → rewritten
path ~0.4ms for a single equals-ignore-case lookup.

## Test plan

- [x] `go test ./internal/query/ -run
'TestLoginName|TestUsers|TestUserByLoginName'` (after generate-stubs)
- [x] Existing `user_test` expected SQL updated (no always-on metadata
join; login-name equals uses JOIN)
- [ ] Run k6 before/after on a large `USER_AMOUNT` (e.g. 50k–100k+) and
compare `list_users_duration` p50/p95/p99
- [x] Manual login v2 username discovery against a large org
- [x] Confirm OR / NOT / CONTAINS login-name queries still return
expected results

<!-- CURSOR_AGENT_PR_BODY_END -->

<div><a
href="https://cursor.com/agents/bc-252883bd-48d9-492e-b619-5ccfa93cf9c3"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://cursor.com/assets/images/open-in-web-dark.png"><source
media="(prefers-color-scheme: light)"
srcset="https://cursor.com/assets/images/open-in-web-light.png"><img
alt="Open in Web" width="114" height="28"
src="https://cursor.com/assets/images/open-in-web-dark.png"></picture></a>&nbsp;<a
href="https://cursor.com/background-agent?bcId=bc-252883bd-48d9-492e-b619-5ccfa93cf9c3"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://cursor.com/assets/images/open-in-cursor-dark.png"><source
media="(prefers-color-scheme: light)"
srcset="https://cursor.com/assets/images/open-in-cursor-light.png"><img
alt="Open in Cursor" width="131" height="28"
src="https://cursor.com/assets/images/open-in-cursor-dark.png"></picture></a>&nbsp;</div>

---------

Co-authored-by: Cursor Agent <[email protected]>
Co-authored-by: Silvan <[email protected]>
2026-07-28 11:56:07 +00:00