goutamadwant opened a new issue, #12190:
URL: https://github.com/apache/seatunnel/issues/12190
### Search before asking
- [x] Searched existing issues and PRs and found no issue or active PR for
this specific retry-forwarding defect. This is a focused follow-up to merged
#10863.
### What happened
The `AMAZON` path in `EmbeddingTransform.open()` does not pass the
configured invocation options into `BedrockModel`. The model therefore uses the
default single SeaTunnel attempt even when `model_retry_max_attempts` requests
additional attempts.
This matters for embedding pipelines using the common retry settings to
recover from transient model failures. Other providers already receive the
invocation options; the Amazon construction path drops them.
### SeaTunnel Version
Reproduced against `dev` at `8bea8c681cacccbd99fa642ca0853718980b9234`.
Released-version coverage has not been established.
### SeaTunnel Config
The reproduction constructs the actual transform with these options, a
single STRING `text` field, and a local HTTP endpoint allocated by the test:
```hocon
model_provider = AMAZON
model = "amazon.titan-embed-text-v2:0"
aws_region = "us-east-1"
api_key = "test-access-key"
secret_key = "test-secret-key"
dimension = 2
vectorization_fields { vector = text }
model_retry_max_attempts = 3
model_retry_backoff_ms = 0
model_retry_max_backoff_ms = 0
model_request_timeout_ms = 5000
```
These credentials and the two-dimensional vector are synthetic fixture
values, not a configuration to send to AWS. The test sets `api_path` to its
loopback server; no AWS account or paid model service is needed.
### Reproduction
1. Construct `EmbeddingTransform` with the options above and map one row
containing `test document`.
2. The local server queues two HTTP 424 responses with `x-amzn-errortype:
ModelErrorException` and a temporary-error message.
3. It then queues HTTP 200 with `{"embedding":[0.25,0.5]}`.
4. On the original production code, mapping fails after the first HTTP
request instead of reaching the successful third response. This was reproduced
on both Java 8 and Java 11.
HTTP 424 in this fixture is retryable by SeaTunnel's adapter, but is not
retried by the SDK's default HTTP retry policy. That separates SeaTunnel
attempts from SDK-internal retries.
### Running Command
The compare contains
`seatunnel-transforms-v2/src/test/java/org/apache/seatunnel/transform/embedding/AmazonEmbeddingInvocationTest.java`.
After the usual repository dependency setup:
```shell
./mvnw -B -ntp -pl seatunnel-transforms-v2 \
-Dskip.spotless=true \
'-Dtest=AmazonEmbeddingInvocationTest#shouldApplyConfiguredRetriesThroughTransform'
test
```
For a red/green comparison in an isolated baseline checkout, apply only that
regression test file, leaving production sources unchanged, and run the same
test. Then run it with the compare's production change.
### Error Exception
The baseline failure includes:
```text
java.lang.RuntimeException: Failed to data vectorization
Caused by: ModelInvocationException: Model invocation failed:
provider=BEDROCK, errorType=TEMPORARY_REMOTE_ERROR,
retryable=true, httpStatus=424
```
### Expected behavior and proposed fix
Pass the invocation options through the existing client factory and existing
model constructor. Read model ID, dimension, and batch-size configuration
before allocating the client.
Before: three configured attempts still fail on the first retryable response.
After: the same fixture makes three requests and returns the expected
vector. Exhaustion is bounded by the configured attempt count. Omitting retry
settings still uses one SeaTunnel attempt.
Existing public constructors, dependencies, credentials, endpoints,
batching, schema, client ownership, and SDK HTTP retry policy remain unchanged.
Explicitly configured additional attempts can incur additional model charges;
EN/ZH documentation and upgrade notes explain this behavior correction.
Timeout enforcement is outside this fix. `model_request_timeout_ms` remains
unapplied to Bedrock SDK calls. This is not a timeout or prompt-cancellation
fix.
### Zeta or Flink or Spark Version
Engine-independent transform/SDK reproduction. No full distributed engine
deployment was used.
### Java or Scala Version
Reproduced and validated on actual Java 8 (Corretto 1.8.0_504) and Java 11
(Temurin 11.0.32.1), not Java 21 substitutes.
### Validation
- Complete transform suite: 1,117 tests on each Java version, zero failures,
errors, or skips.
- Eight new tests cover recovery, exhaustion, the unchanged default,
nonretryable authentication/malformed/missing-vector errors, external-interrupt
preservation without another HTTP request, and config validation before client
construction.
- Tests use the actual transform and AWS SDK against local HTTP fixtures,
not just private-method mocks.
- Repository Spotless and full-repository `./mvnw -q -DskipTests verify`
passed. The latter ran on Java 11 and skips tests; it is separate from the Java
8/11 transform suites.
### Implementation
[Reviewable compare
branch](https://github.com/apache/seatunnel/compare/dev...goutamadwant:fix/amazon-embedding-retries).
A PR will be linked after review.
### Are you willing to submit PR?
- [x] Yes, I am willing to submit a PR.
### Code of Conduct
- [x] I agree to follow this project's [Code of
Conduct](https://www.apache.org/foundation/policies/conduct).
--
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]
