Honor Amazon embedding retry options
Passed SeaTunnel's common model invocation options into the Amazon embedding client and model so configured attempts and backoff apply consistently with other providers.
apache/seatunnel · #12193
AI pipeline reliability fix
SeaTunnel's Amazon embedding transform now honors configured retry and backoff settings for transient model failures.
Problem
The Amazon branch discarded `model_retry_max_attempts` and related invocation settings when constructing `BedrockModel`. Even when three attempts were configured, a retryable model response failed the transform after the first SeaTunnel attempt.
Approach
Routes existing invocation options through the client factory and model constructor while preserving defaults, credentials, endpoints, batching, schemas, ownership, and AWS SDK retry behavior. Documentation distinguishes SeaTunnel attempts, possible model charges, and the unchanged timeout limitation.
Impact and scope
- Allows embedding pipelines to recover from bounded transient model failures as their configuration specifies.
- Aligns Amazon embedding behavior with SeaTunnel's common provider-neutral retry contract.
- Preserves interruption and nonretryable error behavior without introducing hidden retries or API changes.
Validation
- Added eight end-to-end transform regressions for recovery, exhaustion, defaults, authentication, malformed responses, missing vectors, interruption, and pre-client validation.
- The complete transform suite passed 1,117 tests on both Java 8 and Java 11 using the real AWS SDK against local synthetic HTTP responses.
- A SeaTunnel maintainer approved the fix; final checks passed and the GitHub-verified merge commit credits Goutam Adwant and closed GH-12190.