GLOBAL RIGHT/FULL JOIN samples wrong table after side swap, dropping rows

Why we believe this is a bug: buildJoinTreeQueryPlan (src/Planner/PlannerJoinTree.cpp:3436) calls tryRewriteGlobalRightJoinAsLeftJoin, which at :3421 swaps the join's table expressions, and at :3599 the new has_global_join_preserving_broadcast_rows branch wraps the leftmost table in a subquery. Either way the join is no longer executed per shard against a GLOBAL temporary table, so the non-Distributed side becomes an ordinary in-query MergeTree read -> MergeTreeDataSelectExecutor::getSampling -> the else branch at src/Storages/MergeTree/MergeTreeDataSelectExecutor.cpp:289 -> ASTSelectQuery::sampleSize (src/Parsers/ASTSelectQuery.cpp:414-421) returns the sample of the AST's FIRST table expression.

Affected locations:

• src/Planner/PlannerJoinTree.cpp:3421 — tryRewriteGlobalRightJoinAsLeftJoin swaps the join's table expressions; gate at :3380 is getShardCount() < 2
• src/Planner/PlannerJoinTree.cpp:3599 — has_global_join_preserving_broadcast_rows forces should_wrap_left_table for FULL / non-swappable RIGHT joins
• src/Storages/MergeTree/MergeTreeDataSelectExecutor.cpp:289 — getSampling falls back to the query AST's first table expression when the node carries no TableExpressionModifiers

Impact: Silent row loss on SAMPLE + GLOBAL RIGHT/FULL JOIN over a sharded Distributed table, or outright SAMPLING_NOT_SUPPORTED refusal of a query that works today. Reachable without the GLOBAL keyword whenever distributed_product_mode = 'global' or prefer_global_in_and_join = 1 is set, since isGlobalJoin (:3339) treats those settings as global.

-- SAMPLE must apply to the table it is written on, also under GLOBAL RIGHT / FULL JOIN.

DROP TABLE IF EXISTS t_sampled_05183;
DROP TABLE IF EXISTS t_dim_05183;
SET enable_analyzer = 1;
SET enable_parallel_replicas = 0;
SET max_parallel_replicas = 1;
SET prefer_localhost_replica = 1;

DROP TABLE IF EXISTS t_sampled_05183;
DROP TABLE IF EXISTS t_dim_05183;

CREATE TABLE t_sampled_05183 (k UInt32) ENGINE = MergeTree ORDER BY intHash32(k) SAMPLE BY intHash32(k);
CREATE TABLE t_dim_05183 (k UInt32) ENGINE = MergeTree ORDER BY intHash32(k) SAMPLE BY intHash32(k);

INSERT INTO t_sampled_05183 SELECT number FROM numbers(1000);
INSERT INTO t_dim_05183 SELECT number FROM numbers(1000);

SELECT count()
FROM remote('127.0.0.1', currentDatabase(), t_sampled_05183) AS l SAMPLE 0.5
GLOBAL RIGHT JOIN t_dim_05183 AS r ON l.k = r.k;

-- Subquery keeps the join on the initiator and defines the correct result for two shards.
SELECT count()
FROM (SELECT * FROM remote('127.0.0.1,127.0.0.1', currentDatabase(), t_sampled_05183) SAMPLE 0.5) AS l
RIGHT JOIN t_dim_05183 AS r ON l.k = r.k;

SELECT count()
FROM remote('127.0.0.1,127.0.0.1', currentDatabase(), t_sampled_05183) AS l SAMPLE 0.5
GLOBAL RIGHT JOIN t_dim_05183 AS r ON l.k = r.k;

SELECT count()
FROM remote('127.0.0.1,127.0.0.1', currentDatabase(), t_sampled_05183) AS l SAMPLE 0.5
GLOBAL FULL JOIN t_dim_05183 AS r ON l.k = r.k;

DROP TABLE t_sampled_05183;
DROP TABLE t_dim_05183;

Either skip the rewrite and the subquery wrap when the leftmost table expression carries TableExpressionModifiers with a sample ratio (trading the sampling regression back for the row-duplication this PR fixes), or fix the root cause in MergeTreeDataSelectExecutor::getSampling / ReadFromMergeTree::isQueryWithSampling so that under the analyzer the sample is taken only from the node's own TableExpressionModifiers and never from ASTSelectQuery::sampleSize.

Open risks:

• ReadFromMergeTree::isQueryWithSampling (src/Processors/QueryPlan/ReadFromMergeTree.cpp:4243) and StorageMerge (src/Storages/StorageMerge.cpp:1116) read the same AST fallback, so a Merge table on the non-sampled side is likely affected in the same way; not measured here.
• The rewrite also moves the join off the shards for RIGHT/FULL joins whenever distributed_product_mode = 'global' or prefer_global_in_and_join = 1 is set, including when the other side is a purely local table that buildQueryTreeForShard never globalizes. On a real cluster that changes which shard's local table is read; not reproducible on a single-server sandbox.

Found during automated review of PR #118611; whether that PR introduced it could not be established, so nobody is tagged. Severity P2 · Finding h_pr118611_201