fix(query): count grouped, distinct and aggregate reads by their rows

[abnegate]

Problem

count() handed the adapter only the filters and joins. For a groupBy, distinct or aggregate query, a list total counted the matching documents instead of the rows aggregate() or find() returns. Measured on PostgreSQL, count vs real rows:

Shape	count()	Rows
groupBy	9	4
distinct	9	4
plain aggregate	9	1
join + groupBy	8	3
distinct over a join	7	5

This held with and without document-level permissions. sum() dropped the same queries, so a sum over distinct values added up every document. Consumers page aggregate() and distinct find() results with these totals, so the pager was wrong for every such list.

Fix

• Database. For a query with aggregates, groupBy, having or distinct, count() and sum() build the adapter queries and collection with the same helpers fetch() uses (readQueries(), readCollection(), assertReadable()). They authorise joins the same way and skip permissions under the same rule (main collection granted, no joins). Plain, filtered and join-only counts take the existing path unchanged.
• SQL adapters. find()'s query split moves into a shared plan(). A row count builds the statement find() builds through the same builder setup, including the emulated full outer join paths on MariaDB, MySQL and SQLite. It drops order, cursor and page, bounds the statement by LIMIT max, and counts it as SELECT COUNT(1) AS sum FROM (...) table_count. A grouped or distinct sum() adds up a group or selected column of those rows.
• MongoDB, Memory and Redis. They now refuse these counts with the QueryException that find() and aggregate() already throw, rather than answer the document count.
• Overflow (follows fix(aggregates): refuse a floating aggregate that overflows a double #1018). A grouped or distinct sum() adds up a column of the derived table, which can be a dotted alias such as r.score. The dialect builder never compiles that SUM(), so it would miss fix(aggregates): refuse a floating aggregate that overflows a double #1018's + 0 range check. The adapter wraps it through rangeChecked(): + 0 on MariaDB and MySQL, unchanged on PostgreSQL and SQLite, which raise or return INF for Database::sum() to refuse.

Plain counts keep their fast path

I compiled the statements on main and on this branch for MySQL and PostgreSQL. Plain, filtered and join counts, with and without max, are byte-identical. For example, MySQL plain:

SELECT COUNT(1) AS `sum` FROM `database`.`namespace_products` AS `table_main` WHERE `_uid` IN (SELECT _document FROM `database`.`namespace_products_perms` WHERE _permission IN (?) AND _type = ?)

A grouped count now reads:

SELECT COUNT(1) AS `sum` FROM (SELECT COUNT(*) AS `n`, `category` FROM ... WHERE `_uid` IN (...perms...) GROUP BY `category`) AS `table_count`

Tests

• tests/unit/RowCountTest.php runs on SQLite, as admin (collection grant) and as viewer (document permissions). It covers groupBy, groupBy with having, distinct, plain aggregate, left join + groupBy, join + groupBy, distinct over a join, a distinct joined attribute, a distinct join without a select, and the emulated full outer join with groupBy and with distinct. Each count equals the rows the read returns, ignores the page, and is capped by max. It also checks that document counts (plain, filtered, join) are unchanged, that grouped and distinct sums are right, that summing an attribute the groups do not hold is refused, and that Memory refuses these counts.
• tests/unit/Adapter/RowSumStatementTest.php checks the SQL of a grouped and a distinct sum(): SUM(...) + 0 on MariaDB and MySQL, plain on PostgreSQL.
• tests/unit/SumOverflowTest.php checks that a grouped or distinct sum past the range of a double is refused on SQLite in both fetch modes, and that the largest finite one is still answered.
• tests/e2e/Adapter/Scopes/AggregationTests.php:
  • testCountMatchesTheRowsOfGroupedDistinctAndAggregateReads and testSumAddsUpTheRowsOfGroupedAndDistinctReads run on every adapter, as both roles. Adapters without aggregates must refuse with QueryException.
  • testSumOfGroupedOrDistinctRowsThatOverflowsIsRefused covers overflow above and below the range.

Against main's source, 13 of the 15 RowCountTest cases fail, the 4 MariaDB/MySQL RowSumStatementTest cases fail, and both count/sum E2Es fail on SQLite and PostgreSQL ("10 is identical to 4", "20 matches expected 15"). They all pass here. The two RowCountTest cases that pass on main are the unchanged document counts and the distinct join without a select, which is there for MySQL's duplicate-column rule.

Verified locally: Pint and both PHPStan configurations are clean. The new E2Es pass on SQLite and PostgreSQL 18. The unit suite passes apart from the Swoole coroutine pool tests, which segfault on this machine's PHP build on main too. The MySQL, MariaDB and MongoDB E2E matrix runs in CI.

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Next included review available in 11 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

• Configuration used: Repository: utopia-php/database/.coderabbit.yaml
• Review profile: CHILL
• Plan: Advanced
• Run ID: 4311c712-af91-487f-90e1-103bc1216da8

📥 Commits

Reviewing files that changed from the base of the PR and between 107dbf5 and 81d2fc0.

📒 Files selected for processing: 8

• src/Database/Adapter/MariaDB.php
• src/Database/Adapter/SQL.php
• src/Database/Adapter/SQL/Plan.php
• src/Database/Trait/Documents.php
• tests/e2e/Adapter/Scopes/AggregationTests.php
• tests/unit/Adapter/RowSumStatementTest.php
• tests/unit/RowCountTest.php
• tests/unit/SumOverflowTest.php

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}