suvodeep-pyne commented on code in PR #19738:
URL: https://github.com/apache/pinot/pull/19738#discussion_r4298079056


##########
pinot-common/src/main/java/org/apache/pinot/common/utils/helix/IdealStateSingleCommit.java:
##########
@@ -151,7 +149,6 @@ private static class IdealStateWrapper {
   }
 
   public static IdealState cloneIdealState(IdealState idealState) {

Review Comment:
   Done in c22e31a770: removed `IdealStateSingleCommit.cloneIdealState` and 
call `HelixHelper.cloneIdealState` directly. It had no other callers.



##########
pinot-common/src/test/java/org/apache/pinot/common/utils/helix/HelixHelperTest.java:
##########
@@ -88,4 +106,105 @@ public void testRemoveDisabledPartitions() {
     assertTrue(HelixHelper.removeDisabledPartitions(instanceConfig));
     assertTrue(instanceConfig.getDisabledPartitionsMap().isEmpty());
   }
+
+  @Test
+  public void testCloneIdealState() {
+    IdealState idealState = new IdealState("myTable_REALTIME");
+    ZNRecord record = idealState.getRecord();
+    record.setSimpleField("REPLICAS", "2");
+    record.setBooleanField("enableCompression", true);
+    // Insertion order that differs from the sorted order, as in a 
deserialized ZNRecord
+    Map<String, Map<String, String>> mapFields = new LinkedHashMap<>();
+    Map<String, String> instanceStateMap = new LinkedHashMap<>();
+    instanceStateMap.put("Server_2", "CONSUMING");
+    instanceStateMap.put("Server_1", "CONSUMING");
+    mapFields.put("segment_2", instanceStateMap);
+    mapFields.put("segment_1", new LinkedHashMap<>(Map.of("Server_1", 
"ONLINE")));
+    record.setMapFields(mapFields);
+    record.setListField("segment_2", new ArrayList<>(List.of("Server_2", 
"Server_1")));
+    record.setRawPayload(new byte[]{1, 2, 3});
+    record.setVersion(5);
+
+    // Expected result of cloning through a serialization round trip
+    ZNRecordSerializer serializer = new ZNRecordSerializer();
+    ZNRecord expected =
+        new IdealState((ZNRecord) 
serializer.deserialize(serializer.serialize(record))).getRecord();
+    ZNRecord copy = HelixHelper.cloneIdealState(idealState).getRecord();
+
+    // Same content and iteration order
+    assertEquals(copy.getId(), expected.getId());
+    assertEquals(copy, expected);
+    assertEquals(new ArrayList<>(copy.getMapFields().keySet()), new 
ArrayList<>(expected.getMapFields().keySet()));
+    assertEquals(new ArrayList<>(copy.getMapField("segment_2").keySet()),
+        new ArrayList<>(expected.getMapField("segment_2").keySet()));
+    assertEquals(copy.getRawPayload(), expected.getRawPayload());
+    assertEquals(copy.getVersion(), expected.getVersion());
+
+    // Modifying the copy does not modify the original
+    assertNotSame(copy.getRawPayload(), record.getRawPayload());
+    copy.getSimpleFields().put("REPLICAS", "3");
+    copy.getMapField("segment_2").put("Server_2", "ONLINE");
+    copy.getMapFields().remove("segment_1");
+    copy.getListField("segment_2").add("Server_3");
+    assertEquals(record.getSimpleField("REPLICAS"), "2");
+    assertEquals(record.getMapField("segment_2").get("Server_2"), "CONSUMING");
+    assertTrue(record.getMapFields().containsKey("segment_1"));
+    assertEquals(record.getListField("segment_2"), List.of("Server_2", 
"Server_1"));
+  }
+
+  /// cloneIdealState copies a ZNRecord field by field, so it must know every 
field an IdealState stores in ZK. If a
+  /// Helix upgrade fails this test, make cloneIdealState copy the new state 
(or leave it out if it is not stored, like
+  /// the ZK stat fields), then update the expected values.
+  @Test
+  public void testCloneIdealStateKnowsAllStoredFields()
+      throws IOException {
+    // Instance fields, including inherited ones, sorted
+    assertEquals(getInstanceFieldNames(ZNRecord.class),
+        List.of("_creationTime", "_deltaList", "_ephemeralOwner", 
"_modifiedTime", "_serializer", "_version", "id",
+            "listFields", "mapFields", "rawPayload", "simpleFields"));
+    assertEquals(getInstanceFieldNames(IdealState.class), List.of("_record", 
"_stat"));
+
+    // Of these, only the following are stored in ZK: the properties the 
serializer's mapper can write, whatever their
+    // values, and the properties it writes for a record with all of them set
+    Set<String> storedFields = Set.of("id", "simpleFields", "mapFields", 
"listFields", "rawPayload");
+    ObjectMapper mapper = ZNRecordSerializerMapper.get();
+    
assertEquals(mapper.getSerializationConfig().introspect(mapper.constructType(ZNRecord.class)).findProperties()
+        .stream()
+        .filter(BeanPropertyDefinition::couldSerialize)
+        .map(BeanPropertyDefinition::getName)
+        .collect(Collectors.toSet()), storedFields);
+    ZNRecord record = new ZNRecord("myTable_REALTIME");
+    record.setSimpleField("REPLICAS", "2");
+    record.setMapField("segment_1", Map.of("Server_1", "ONLINE"));
+    record.setListField("segment_1", List.of("Server_1"));
+    record.setRawPayload(new byte[]{1});
+    byte[] bytes = new ZNRecordSerializer().serialize(record);
+    if (GZipCompressionUtil.isCompressed(bytes)) {
+      bytes = GZipCompressionUtil.uncompress(new ByteArrayInputStream(bytes));
+    }
+    Set<String> writtenFields = new HashSet<>();
+    
JsonUtils.bytesToJsonNode(bytes).fieldNames().forEachRemaining(writtenFields::add);
+    assertEquals(writtenFields, storedFields);
+  }
+
+  /// Keeps duplicates, so that a field hiding an inherited one is caught.
+  private static List<String> getInstanceFieldNames(Class<?> clazz) {
+    List<String> fieldNames = new ArrayList<>();
+    for (Class<?> c = clazz; c != Object.class; c = c.getSuperclass()) {
+      for (Field field : c.getDeclaredFields()) {
+        if (!Modifier.isStatic(field.getModifiers()) && !field.isSynthetic()) {
+          fieldNames.add(field.getName());
+        }
+      }
+    }
+    fieldNames.sort(null);
+    return fieldNames;
+  }
+
+  /// Exposes the ObjectMapper that ZNRecordSerializer writes ZNRecords with.
+  private static class ZNRecordSerializerMapper extends ZNRecordSerializer {

Review Comment:
   Yes, that's the intended tripwire. `cloneIdealState` copies the record field 
by field. If a Helix upgrade added stored state to `ZNRecord`, the copy would 
drop it, and the next IdealState write would remove it from ZK without any 
error.
   
   It should rarely trip. In Helix's history:
   - the only lasting change to the stored `ZNRecord` properties was adding 
`rawPayload` in 2013;
   - the instance fields last changed in 2018 (`_ephemeralOwner` on `ZNRecord`) 
and 2017 (`_stat` on `HelixProperty`);
   - `IdealState` has never had fields of its own;
   - the `protected static mapper` has been on `ZNRecordSerializer` since 2023.
   
   A lighter check, such as only the JSON keys written for a sample record, 
would miss a new property that is null in the sample, since the serializer 
omits nulls. That's why it also introspects with the serializer's own mapper. 
If it does trip on an upgrade, the fix is to update `cloneIdealState` and the 
expected lists.



##########
pinot-common/src/main/java/org/apache/pinot/common/utils/helix/HelixHelper.java:
##########
@@ -78,9 +78,31 @@ private HelixHelper() {
 
   public static final String BROKER_RESOURCE = 
CommonConstants.Helix.BROKER_RESOURCE_INSTANCE;
 
+  /// Returns a deep copy of the IdealState with the same content as cloning 
it through a serialization round trip of
+  /// its ZNRecord: id, simple fields, map fields, list fields and raw 
payload, with the same container types and
+  /// iteration order. Like the round trip, the ZK stat fields (e.g. version) 
are not copied. Unlike the round trip, the
+  /// serializer's write policies (list field bound, size limit) are not 
applied; they still apply when the IdealState
+  /// is written.
+  ///
+  /// Copying the fields directly avoids serializing (and compressing) the 
whole ZNRecord and parsing it back, which
+  /// takes seconds for an IdealState with hundreds of thousands of segments 
and widens the window for version
+  /// conflicts when updating it.
   public static IdealState cloneIdealState(IdealState idealState) {
-    return new IdealState(
-        (ZNRecord) 
ZN_RECORD_SERIALIZER.deserialize(ZN_RECORD_SERIALIZER.serialize(idealState.getRecord())));
+    ZNRecord record = idealState.getRecord();
+    IdealState copy = new IdealState(record.getId());
+    ZNRecord copyRecord = copy.getRecord();
+    copyRecord.setSimpleFields(new TreeMap<>(record.getSimpleFields()));
+    TreeMap<String, Map<String, String>> mapFields = new 
TreeMap<>(record.getMapFields());
+    mapFields.replaceAll((key, value) -> value != null ? new 
LinkedHashMap<>(value) : null);
+    copyRecord.setMapFields(mapFields);
+    TreeMap<String, List<String>> listFields = new 
TreeMap<>(record.getListFields());
+    listFields.replaceAll((key, value) -> value != null ? new 
ArrayList<>(value) : null);
+    copyRecord.setListFields(listFields);

Review Comment:
   Agreed. `cloneIdealState` only takes IdealStates, and Pinot never sets 
`ZNRecord.LIST_FIELD_BOUND` on them. The bound would still apply when a record 
is written, so even a reused clone stays bounded in ZK.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to