diff --git a/ui/src/plugins/dev.perfetto.GpuByProcess/index.ts b/ui/src/plugins/dev.perfetto.GpuByProcess/index.ts
index fe7d4d64a82..baa9ef57ba4 100644
--- a/ui/src/plugins/dev.perfetto.GpuByProcess/index.ts
+++ b/ui/src/plugins/dev.perfetto.GpuByProcess/index.ts
@@ -418,6 +418,10 @@ export default class implements PerfettoPlugin {
           trace: ctx,
           uri,
           dataset: t.dataset,
+          // Root table for SQL-event resolution: without it these tracks
+          // are invisible to selectSqlEvent(), so ts/dur deep links into
+          // GPU streams zoom but never select the slice.
+          rootTableName: 'slice',
           detailsPanel: () => new ThreadSliceDetailsPanel(ctx),
         }),
       });
diff --git a/ui/src/trace_processor/dataset.ts b/ui/src/trace_processor/dataset.ts
index 9cc44715e01..8984838615c 100644
--- a/ui/src/trace_processor/dataset.ts
+++ b/ui/src/trace_processor/dataset.ts
@@ -491,7 +491,14 @@ function getJoinSignature(dataset: SourceDataset): string {
   if (!dataset.joins) {
     return '';
   }
-  return Object.keys(dataset.joins).sort().join(',');
+  // The signature must identify the join definition, not just its alias:
+  // two datasets whose 'depth' joins point at different tables must not be
+  // merged, or the merged query joins every row through one table and
+  // silently drops the rows only present in the others.
+  return Object.entries(dataset.joins)
+    .map(([id, join]) => `${id}:${join.from}:${join.unique ?? false}`)
+    .sort()
+    .join(',');
 }
 
 function mergeFilters(filters: InFilter[]): InFilter | undefined {
diff --git a/ui/src/trace_processor/dataset_unittest.ts b/ui/src/trace_processor/dataset_unittest.ts
index 9d034278a49..9293f3cd172 100644
--- a/ui/src/trace_processor/dataset_unittest.ts
+++ b/ui/src/trace_processor/dataset_unittest.ts
@@ -396,6 +396,39 @@ FROM (slice)
 WHERE id IN (123, 456)`);
 });
 
+test('union dataset does not merge datasets whose joins differ', () => {
+  // Mirrors the multi-track slice tracks created for grouped/split tracks:
+  // each track's dataset joins its own layout-depth table under the same
+  // 'depth' alias. Merging them would join every row through one table and
+  // silently drop the rows only present in the others.
+  const makeDepthDataset = (table: string, trackIds: number[]) =>
+    new SourceDataset({
+      src: 'slice',
+      schema: {id: NUM, depth: NUM},
+      select: {
+        id: 'id',
+        depth: {join: 'depth', expr: 'depth.depth'},
+      },
+      joins: {
+        depth: {from: `${table} USING (id)`, unique: true},
+      },
+      filter: {
+        col: 'track_id',
+        in: trackIds,
+      },
+    });
+
+  const dataset = UnionDataset.create([
+    makeDepthDataset('__depth_1', [1, 2]),
+    makeDepthDataset('__depth_2', [3, 4]),
+  ]);
+
+  const query = dataset.query();
+  expect(query).toContain('__depth_1');
+  expect(query).toContain('__depth_2');
+  expect(query).toContain('UNION ALL');
+});
+
 test('union dataset batches large numbers of unions', () => {
   const datasets = [];
   for (let i = 0; i < 800; i++) {