Cap external scale IT at 100k and fix external-cluster test flakiness

.github/workflows/java-playwright-external.yml

@@ -396,8 +396,10 @@ jobs:
 
   scale-it-external:
     # Runs after search-it (shared instance, singleton reindex app). max-parallel: 1 so the matrix
-    # entries (e.g. 100k then 500k) seed + reindex one at a time — two concurrent reindexes on one
-    # cluster collide and the db/es counts would mix cohorts.
+    # entries seed + reindex one at a time — two concurrent reindexes on one cluster collide and the
+    # db/es counts would mix cohorts. Default cohort is 100k only; a 500k run took >5h and was
+    # cancelled every night without ever completing. Larger sizes can still be requested on demand
+    # via the scaleSeeds dispatch input (e.g. [100000, 500000]).
     needs: [search-it-external]
     if: ${{ always() && github.event.inputs.runScaleIt != 'false' }}
     runs-on: ubuntu-latest
@@ -406,7 +408,7 @@ jobs:
       fail-fast: false
       max-parallel: 1
       matrix:
-        seed: ${{ fromJSON(github.event.inputs.scaleSeeds || '[100000, 500000]') }}
+        seed: ${{ fromJSON(github.event.inputs.scaleSeeds || '[100000]') }}
     steps:
       - name: Free Disk Space (Ubuntu)
         uses: jlumbroso/free-disk-space@main

openmetadata-integration-tests/src/test/java/org/openmetadata/it/tests/search/SearchRelevancyPreviewIT.java

@@ -121,20 +121,27 @@ void searchFieldsSelectWhichFieldMatches(final TestNamespace ns) {
 
     final SearchSettings base = currentSettings();
 
+    // Field selection is proven by which seeded table is IN vs OUT: name-only must match the
+    // name-carrier and exclude the description-carrier (and vice versa). Use
+    // contains/doesNotContain
+    // rather than containsExactly — on the shared cluster a foreign ngram-colliding table can add a
+    // hit without changing which of OUR two tables the searched field selects.
     SearchSettings nameOnly = SearchSettingsTestHelper.copyOf(base);
     SearchSettingsTestHelper.setOnlySearchField(nameOnly, TABLE_INDEX, NAME_FIELD, 5.0);
     nameOnly = SearchSettingsTestHelper.withRankingDisabled(nameOnly, TABLE_INDEX);
     assertThat(SearchSettingsTestHelper.previewIds(server, query, TABLE_INDEX, nameOnly, 10))
-        .as("searching only 'name' must return only the table whose name carries the token")
-        .containsExactly(tokenInName.getId().toString());
+        .as("searching only 'name' must match the name-carrier, not the description-carrier")
+        .contains(tokenInName.getId().toString())
+        .doesNotContain(tokenInDescription.getId().toString());
 
     SearchSettings descriptionOnly = SearchSettingsTestHelper.copyOf(base);
     SearchSettingsTestHelper.setOnlySearchField(
         descriptionOnly, TABLE_INDEX, DESCRIPTION_FIELD, 5.0);
     descriptionOnly = SearchSettingsTestHelper.withRankingDisabled(descriptionOnly, TABLE_INDEX);
     assertThat(SearchSettingsTestHelper.previewIds(server, query, TABLE_INDEX, descriptionOnly, 10))
-        .as("searching only 'description' must return only the table whose description carries it")
-        .containsExactly(tokenInDescription.getId().toString());
+        .as("searching only 'description' must match the description-carrier, not the name-carrier")
+        .contains(tokenInDescription.getId().toString())
+        .doesNotContain(tokenInName.getId().toString());
   }
 
   @Test
@@ -193,7 +200,8 @@ void matchTypeExactRequiresTheFullKeyword(final TestNamespace ns) {
     final DatabaseSchema schema = DatabaseSchemaTestFactory.createSimple(ns);
     final Table exactTable =
         RelevancyFixtures.createTable(schema, exactName, "plaindescription", null);
-    RelevancyFixtures.createTable(schema, exactName + "extra", "plaindescription", null);
+    final Table prefixedTable =
+        RelevancyFixtures.createTable(schema, exactName + "extra", "plaindescription", null);
     awaitIndexed(exactName, 2);
 
     final SearchSettings base = currentSettings();
@@ -210,9 +218,13 @@ void matchTypeExactRequiresTheFullKeyword(final TestNamespace ns) {
     SearchSettingsTestHelper.setOnlySearchField(
         standard, TABLE_INDEX, NAME_FIELD, 5.0, FieldBoost.MatchType.STANDARD);
     standard = SearchSettingsTestHelper.withRankingDisabled(standard, TABLE_INDEX);
+    // STANDARD is a fuzzy name.ngram match, so on the shared cluster a foreign ngram-colliding
+    // table can add hits beyond the two we seeded. Assert both seeded ids are present rather than
+    // an exact size (mirrors maxResultHitsClampsTheReturnedHits), which such a foreign hit
+    // inflates.
     assertThat(SearchSettingsTestHelper.previewIds(server, exactName, TABLE_INDEX, standard, 10))
         .as("matchType=standard must match both the exact and the prefixed name")
-        .hasSize(2);
+        .contains(exactTable.getId().toString(), prefixedTable.getId().toString());
   }
 
   @Test

openmetadata-integration-tests/src/test/java/org/openmetadata/it/util/ExternalTokenRefresher.java

@@ -29,9 +29,6 @@
 public final class ExternalTokenRefresher implements AutoCloseable {
 
   private static final Logger LOG = LoggerFactory.getLogger(ExternalTokenRefresher.class);
-  // Renew this far ahead of expiry so a flaky login endpoint has room for several 60s retries
-  // before the token actually dies (login TTL is ~1h, so 10m early is safe).
-  private static final Duration EXPIRY_BUFFER = Duration.ofMinutes(10);
   // Bounds BOTH the connect and the read phase of each login. Without a read timeout a slow or
   // half-open gateway would block http.send() forever on the single-thread scheduler, and the
   // token would never renew for the rest of the run.
@@ -125,12 +122,18 @@ private String login() {
     }
   }
 
+  // Renew at half the token's remaining lifetime (~30m on the external cluster's ~1h token) rather
+  // than a fixed buffer before expiry. That leaves ~half the TTL as retry runway if the login
+  // endpoint is unreachable at renewal time: a ~10m login-endpoint outage that lands on the renewal
+  // window (observed 2026-07-17) burned through a fixed 10m buffer and let the token expire
+  // mid-seed; half-life gives the 60s retries far more room to catch the endpoint before the token
+  // actually dies.
   private long refreshDelaySeconds(final String token) {
     long delaySeconds = FALLBACK_REFRESH_SECONDS;
     final Long exp = expiryEpochSeconds(token);
     if (exp != null) {
       final long now = System.currentTimeMillis() / 1000;
-      delaySeconds = Math.max(60, exp - now - EXPIRY_BUFFER.toSeconds());
+      delaySeconds = Math.max(60, (exp - now) / 2);
     }
     return delaySeconds;
   }

openmetadata-integration-tests/src/test/java/org/openmetadata/playwright/ui/pages/DataQualityListPage.java

@@ -187,8 +187,17 @@ public DataQualityListPage clickPrevious() {
 
   /** Open the page-size dropdown and assert the standard 3 options render. */
   public DataQualityListPage assertPageSizeOptionsCount(final int expected) {
-    openMenu(byTestId("page-size-selection-dropdown"), page.locator(".ant-dropdown-menu"));
-    PlaywrightAssertions.assertThat(page.locator(".ant-dropdown-menu-item")).hasCount(expected);
+    // Scope to the OPEN overlay: every Ant <Dropdown> on the page (nav bar, help, etc.) portals a
+    // .ant-dropdown-menu to <body> and keeps it in the DOM while hidden, so a bare
+    // .ant-dropdown-menu first-matches a foreign hidden menu and the visibility wait times out.
+    // :not(.ant-dropdown-hidden) pins it to the just-opened page-size menu (mirrors the
+    // .ant-select-dropdown idiom above).
+    final Locator pageSizeMenu =
+        page.locator(".ant-dropdown:not(.ant-dropdown-hidden) .ant-dropdown-menu");
+    openMenu(byTestId("page-size-selection-dropdown"), pageSizeMenu);
+    PlaywrightAssertions.assertThat(
+            page.locator(".ant-dropdown:not(.ant-dropdown-hidden) .ant-dropdown-menu-item"))
+        .hasCount(expected);
     return this;
   }