-
[OPIK-5664] [SDK] feat: global client management API with context-wise propagation and batching warning (#6123)
frostbyte_neo 发布于
2026-04-08 13:08:55 +00:00 - [OPIK-5664] [SDK] feat: add global client management API with context-wise propagation
Introduce get_global_client(), set_global_client(), reset_global_client(),
and get_current_client_raw() to replace the lru_cache-based get_client_cached()
singleton. The new API supports context-local client overrides via contextvars,
enabling EvaluationSuite.run() and the evaluation engine to propagate an
explicit Opik client to @track decorators and integrations in worker threads.Key changes:
- EvaluationSuite stores and propagates the parent Opik client
- StreamingExecutor sets the client in each worker thread via set_global_client(context_wise=True)
- All integrations resolve the client lazily at runtime via @property
- ADK patchers no longer require an opik_client argument
- get_client_cached() kept as a thin delegate for backwards compatibility
Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
- test: add client propagation tests for evaluate_suite
- Explicit client is used for experiment creation (not global)
- EvaluationSuite.run() forwards its stored client
- Worker threads see the same client via get_global_client()
Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
- fix: skip .end() in test teardown to avoid network calls
The autouse fixture was calling reset_global_client() which ends the
singleton client. Tests that accidentally created a real Opik client
would block on network flush during teardown. Using end_client=False
just clears the slot without network I/O.Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
- feat: add public batching param to Opik and warn on update with batching
- Add public
batching: bool = Trueparameter to Opik.init - Deprecate internal
_use_batchingparam - Warn when Span.end()/update(), Trace.end()/update(),
Opik.update_span(), Opik.update_trace() are called with batching
enabled — updates can be silently dropped causing data loss - Add Streamer.use_batching property to check batching state
- Replace all _use_batching=True callsites with batching=True
Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
- chore: lint fixes and version bump to 1.11.0
Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
- fix: keep opik_client param on OpikTracer for backwards compatibility
The parameter is accepted but ignored — client resolves lazily via
the _opik_client property.Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
- fix: address PR review comments
- Fix batching=False bug: _use_batching default back to False so
Opik(batching=False) actually disables batching - Switch batching warnings from LOGGER.warning to warnings.warn so
they deduplicate automatically (once per callsite per session)
Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
-
Fix lint errors
-
fix: preserve opik_client arg order in OpikTracer for backwards compatibility
Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
- fix: set context client in single-worker evaluation path
The workers==1 branch runs tasks inline without set_global_client,
so @track decorators inside the task would fall back to the global
singleton instead of using the explicit client.Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com
Co-authored-by: Claude Opus 4.6 (1M context) noreply@anthropic.com
下载附件