[BUG] BWC Rolling upgrade tests fails for SparseEncoder Processor during Batch Ingestion

[vibrantvarun]

What is the bug?

BatchIngestionIT.testBatchIngestion_SparseEncodingProcessor_E2EFlow test is failing with the following error.

> Task :qa:rolling-upgrade:testAgainstOneThirdUpgradedCluster
REPRODUCE WITH: ./gradlew ':qa:rolling-upgrade:testAgainstOneThirdUpgradedCluster' --tests "org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT.testBatchIngestion_SparseEncodingProcessor_E2EFlow" -Dtests.seed=801B4B74838557A -Dtests.security.manager=false -Dtests.bwc.version=2.19.0-SNAPSHOT -Dtests.locale=sw-TZ -Dtests.timezone=America/Bahia_Banderas -Druntime.java=21
Suite: Test class org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT
  2> Jan 24, 2025 6:12:44 PM org.apache.lucene.internal.vectorization.VectorizationProvider lookup
  2> WARNING: Java vector incubator module is not readable. For optimal vector performance, pass '--add-modules jdk.incubator.vector' to enable Vector API.
  2> SLF4J: Failed to load class "org.slf4j.impl.StaticLoggerBinder".
  2> SLF4J: Defaulting to no-operation (NOP) logger implementation
  2> SLF4J: See http://www.slf4j.org/codes.html#StaticLoggerBinder for further details.
  2> REPRODUCE WITH: ./gradlew ':qa:rolling-upgrade:testAgainstOneThirdUpgradedCluster' --tests "org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT.testBatchIngestion_SparseEncodingProcessor_E2EFlow" -Dtests.seed=801B4B74838557A -Dtests.security.manager=false -Dtests.bwc.version=2.19.0-SNAPSHOT -Dtests.locale=sw-TZ -Dtests.timezone=America/Bahia_Banderas -Druntime.java=21
  2> java.lang.AssertionError: expected:<10> but was:<8>
        at __randomizedtesting.SeedInfo.seed([801B4B74838557A:8487AE9ACE1A1513]:0)
        at org.junit.Assert.fail(Assert.java:89)
        at org.junit.Assert.failNotEquals(Assert.java:835)
        at org.junit.Assert.assertEquals(Assert.java:647)
        at org.junit.Assert.assertEquals(Assert.java:633)
        at org.opensearch.neuralsearch.BaseNeuralSearchIT.validateDocCountAndInfo(BaseNeuralSearchIT.java:1535)
        at org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT.testBatchIngestion_SparseEncodingProcessor_E2EFlow(BatchIngestionIT.java:46)
  2> NOTE: leaving temporary files on disk at: /home/runner/work/neural-search/neural-search/qa/rolling-upgrade/build/testrun/testAgainstOneThirdUpgradedCluster/temp/org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT_801B4B74838557A-001
  2> NOTE: test params are: codec=Asserting(Lucene912): {}, docValues:{}, maxPointsInLeafNode=22, maxMBSortInHeap=5.7175746038185356, sim=Asserting(RandomSimilarity(queryNorm=true): {}), locale=sw-TZ, timezone=America/Bahia_Banderas
  2> NOTE: Linux 6.8.0-1020-azure amd64/Azul Systems, Inc. 21.0.6 (64-bit)/cpus=4,threads=3,free=453753024,total=536870912
  2> NOTE: All tests run in this JVM: [BatchIngestionIT]
BatchIngestionIT > testBatchIngestion_SparseEncodingProcessor_E2EFlow FAILED
    java.lang.AssertionError: expected:<10> but was:<8>
        at __randomizedtesting.SeedInfo.seed([801B4B74838557A:8487AE9ACE1A1513]:0)
        at org.junit.Assert.fail(Assert.java:89)
        at org.junit.Assert.failNotEquals(Assert.java:835)
        at org.junit.Assert.assertEquals(Assert.java:647)
        at org.junit.Assert.assertEquals(Assert.java:633)
        at org.opensearch.neuralsearch.BaseNeuralSearchIT.validateDocCountAndInfo(BaseNeuralSearchIT.java:1535)
        at org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT.testBatchIngestion_SparseEncodingProcessor_E2EFlow(BatchIngestionIT.java:46)
  1> [2025-01-24T12:12:44,811][INFO ][o.o.n.b.r.BatchIngestionIT] [testBatchIngestion_SparseEncodingProcessor_E2EFlow] before test
  1> [2025-01-24T12:12:45,122][INFO ][o.o.n.b.r.BatchIngestionIT] [testBatchIngestion_SparseEncodingProcessor_E2EFlow] initializing REST clients against [http://[::1]:[410](https://github.com/opensearch-project/neural-search/actions/runs/12942873461/job/36138496383?pr=1140#step:4:411)85, http://127.0.0.1:40513, http://[::1]:35229, http://127.0.0.1:41063, http://[::1]:41591, http://127.0.0.1:34731]
  1> [2025-01-24T12:12:49,676][INFO ][o.o.n.b.r.BatchIngestionIT] [testBatchIngestion_SparseEncodingProcessor_E2EFlow] There are still tasks running after this test that might break subsequent tests [cluster:admin/opensearch/ml/undeploy_model, cluster:admin/opensearch/mlinternal/syncup, indices:admin/seq_no/global_checkpoint_sync, indices:admin/seq_no/global_checkpoint_sync[p], indices:data/write/bulk, indices:data/write/bulk[s]].
  1> [2025-01-24T12:12:49,722][INFO ][o.o.n.b.r.BatchIngestionIT] [testBatchIngestion_SparseEncodingProcessor_E2EFlow] after test

How can one reproduce the bug?

Run the test locally or raise a PR on neural to see it in github CI check.

What is the expected behavior?

Test should pass successfully.

Do you have any additional context?

https://github.com/opensearch-project/neural-search/actions/runs/12942873461/job/36138496383?pr=1140

[heemin32]

Is it flaky test?

[zhichao-aws]

It's a serialization/deserialization issue from OpenSearch core, all query builders fail to serialization/deserialization between nodes.

To reproduce:

1. set up a two node cluster with 1 3.0.0 node + 1 2.19 snapshot node.
2. create index

PUT test
{
    "settings": {
      "number_of_shards": 2,
      "number_of_replicas": 0
    },
    "mappings": {
      "properties": {
        "passage_embedding": {
          "type": "rank_features"
        },
        "passage_text": {
          "type": "text"
        }
      }
    }
}

3. ingest doc

POST _bulk
{"index":{"_index":"test"}}
{"passage_embedding":{"hello":1.1,"world":1.2}, "passage_text": "hello world"}
{"index":{"_index":"test"}}
{"passage_embedding":{"hello":1.1,"world":1.2}, "passage_text": "hello world"}
{"index":{"_index":"test"}}
{"passage_embedding":{"hello":1.1,"world":1.2}, "passage_text": "hello world"}
{"index":{"_index":"test"}}
{"passage_embedding":{"hello":1.1,"world":1.2}, "passage_text": "hello world"}
{"index":{"_index":"test"}}
{"passage_embedding":{"hello":1.1,"world":1.2}, "passage_text": "hello world"}

4. do search

GET test/_search

here we can use any query(if leave it empty, it's match_all), and send request to any nodes. we get shard failure of serialization/deserialization error in response:

{'took': 63,
 'timed_out': False,
 '_shards': {'total': 2,
  'successful': 1,
  'skipped': 0,
  'failed': 1,
  'failures': [{'shard': 1,
    'index': 'test2',
    'node': 'AEmZn6uASnGjAU0SCF-EXg',
    'reason': {'type': 'illegal_state_exception',
     'reason': 'unexpected byte [0x3f]'}}]},
 'hits': {'total': {'value': 3, 'relation': 'eq'},
  'max_score': 1.0,
  'hits': [ ... ]}}

[vibrantvarun]

Hey @zhichao-aws It can be due the recent breaking changes in 3.0. Can you try the same with 2.18, 2.19?

[vibrantvarun]

@martin-gaievski You did some deep-dive on the breaking changes coming from 3.0, can you share some insights here as we do not see this issue in bwc tests running on 2.x.

[martin-gaievski]

I checked the lucene 10 changes, drafted PR to address them #1141.

If that would be related to lucene 10 then code wouldn't even compile, there are some incompatible API changes. At the first glance that is related to the internal logic of the SparseEncoder, or maybe some resource that we setup in this particular test

[zhichao-aws]

Hey @zhichao-aws It can be due the recent breaking changes in 3.0. Can you try the same with 2.18, 2.19?

Do you mean the bwc test between 2.18, 2.19 or bwc test between main and 2.18?

[zhichao-aws]

I checked the lucene 10 changes, drafted PR to address them #1141.

If that would be related to lucene 10 then code wouldn't even compile, there are some incompatible API changes. At the first glance that is related to the internal logic of the SparseEncoder, or maybe some resource that we setup in this particular test

It's not due to the lucene 10 changes. It's caused by serialization/deserialization error from core opensearch-project/OpenSearch#17125

[zhichao-aws]

Based on the comment, the issue should be fixed after we bump to 3.0.0-alpha1

[martin-gaievski]

I've drafted PR for 3.0-alpha #1141, we can monitor progress over there, I see BWC are failing for now os it doesn't look like simple adoption of latest core takes care. maybe some additional steps are needed on neural search side

[zhichao-aws]

I've drafted PR for 3.0-alpha #1141, we can monitor progress over there, I see BWC are failing for now os it doesn't look like simple adoption of latest core takes care. maybe some additional steps are needed on neural search side

Based on the error log, it seems to fail due to another reason (ml model). While the previous reason is related to query serialization/deserialization between nodes.

REPRODUCE WITH: ./gradlew ':qa:rolling-upgrade:testAgainstOldCluster' --tests "org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT.testBatchIngestion_SparseEncodingProcessor_E2EFlow" -Dtests.seed=34D952AC3611A812 -Dtests.security.manager=false -Dtests.bwc.version=2.20.0-SNAPSHOT -Dtests.locale=bg-Cyrl-BG -Dtests.timezone=Pacific/Gambier -Druntime.java=23

BatchIngestionIT > testBatchIngestion_SparseEncodingProcessor_E2EFlow FAILED
    java.lang.RuntimeException: Model t0NmApUBfHInJ7oBTwbr failed to load after 30 attempts
        at __randomizedtesting.SeedInfo.seed([34D952AC3611A812:B85F4881B033E87B]:0)
        at org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT.waitForModelToLoad(BatchIngestionIT.java:108)
        at org.opensearch.neuralsearch.bwc.rolling.BatchIngestionIT.testBatchIngestion_SparseEncodingProcessor_E2EFlow(BatchIngestionIT.java:36)

[zhichao-aws]

The bug is fixed after switch to 3.0.0-alpha1.