fix: Change attributes datasets and experiment on client to lazyloading and mark datasets as deprecated

[tassadarius]

• I have added tests that cover my changes.
• If adding a new instrumentation or changing an existing one, I've added screenshots from some observability platform showing the change.
• PR name follows conventional commits format: feat(instrumentation): ... or fix(instrumentation): ....
• (If applicable) I have updated the documentation accordingly.

Changes

• Import and load attributes datasets and experimental only once when actually requested
• Mark attribute datasets on Client as deprecated

Reasoning

Calling Traceloop.init() allocates ~35 MB of memory when pandas is
installed, either by the attributes datasets or experiment. Both of these attributes seem
not relevant for production

People interested in the Datasets class should IMO just create an
instance on their own.

Since it's already released I propose a deprecation process.

In general I think it is a bad idea to expose and initialize all these internal attributes directly on the public main class.

Here you can see the memory usages done with memray of this simple execution:

TRACELOOP_API_KEY=asdf uv  run --active memray run -c "from traceloop.sdk import Traceloop;Traceloop.init(app_name='joke_generation_service')

Code	Mem-usage	Link to report
Current main	70.1 MB	memray-flamegraph-before-patch.html
Patch	37.6 MB	memray-flamegraph-with-patch.html

---

Important

Optimize Client memory usage by lazy-loading datasets and experiment attributes, marking datasets as deprecated.

• Behavior:
  • datasets and experiment attributes in Client are now lazy-loaded, reducing initial memory usage.
  • datasets attribute is marked as deprecated with a warning, suggesting users create their own instances.
• Tests:
  • Added test_client_lazy_loads_datasets() and test_client_lazy_loads_experiment() in test_client.py to verify lazy loading.
  • Added test_datasets_deprecation_warnings() to check for deprecation warnings.
• Misc:
  • Removed direct imports of Datasets and Experiment in client.py and moved them inside the properties.

^{This description was created by for 2f53649. You can customize this summary. It will automatically update as commits are pushed.}

Summary by CodeRabbit

• Deprecations

  • The datasets property on the client is deprecated, including its setter and deleter; it will be removed in a future release.

• Changes

  • The experiment property now initializes on first access instead of during client creation.

• Tests

  • Added tests verifying lazy-loading behavior for experiment and datasets, caching/idempotence, and that deprecation warnings are emitted for datasets.

_{✏️ Tip: You can customize this high-level summary in your review settings.}

[CLAassistant]

All committers have signed the CLA.

[ellipsis-dev]

Important

Looks good to me! 👍

Reviewed everything up to 2f53649 in 10 seconds. Click for details.

• Reviewed 125 lines of code in 2 files
• Skipped 0 files when reviewing.
• Skipped posting 0 draft comments. View those below.
• Modify your settings and rules to customize what types of comments Ellipsis leaves. And don't forget to react with 👍 or 👎 to teach Ellipsis.

Workflow ID: wflow_PbsjYa6y5uOLuSYK

^{You can customize by changing your verbosity settings, reacting with 👍 or 👎, replying to comments, or adding code review rules.}

[coderabbitai]

📝 Walkthrough

Walkthrough

Client class attributes datasets and experiment were converted from eager public attributes to lazy-loaded properties; datasets is marked deprecated with setter/deleter. Tests were added to verify lazy-loading, caching/idempotence, and deprecation warnings.

Changes

Cohort / File(s)	Summary
Client lazy-loading implementation packages/traceloop-sdk/traceloop/sdk/client/client.py	Replaced eager public attributes datasets and experiment with lazy-loaded properties. Added private _datasets and _experiment state, deprecation decorator for datasets, and setter/deleter for datasets.
Tests for lazy-loading & deprecation packages/traceloop-sdk/tests/test_client.py	Added imports for Datasets and Experiment. Introduced tests verifying lazy instantiation, type checks, stateful caching/idempotence for client.datasets and client.experiment, and deprecation warnings on access/assignment/deletion of datasets.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~22 minutes

Poem

🐇 I nibble code where properties sleep,
I wake them gently when callers peep.
Deprecated carrots left on the lawn,
New lazy sprouts greet the dawn. ✨

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 60.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately summarizes the main changes: converting datasets and experiment attributes to lazy-loading and marking datasets as deprecated.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🤖 Fix all issues with AI agents

In `@packages/traceloop-sdk/traceloop/sdk/client/client.py`:
- Around line 81-90: Update the deprecation messages that incorrectly say
"Dataclass traceloop.sdk.datasets" to correctly reference the Datasets
class/module by changing the reason text in the `@deprecated` decorators for the
datasets accessor and the two other deprecated items in this file (the datasets
method and the other two decorators around lines 93-99 and 101-104); ensure all
three reasons consistently say something like "Datasets traceloop.sdk.datasets"
or "the Datasets class in traceloop.sdk.datasets" so they uniformly point to the
Datasets class/module.
- Around line 1-3: Ruff F821 arises because the name Datasets is used in a
setter type annotation but not defined at import time; add "from typing import
TYPE_CHECKING" to the imports and inside an if TYPE_CHECKING: block import the
Datasets symbol (e.g., "from .datasets import Datasets" or the correct module
path) so the type name is available for type checking only and doesn't trigger
runtime imports—update the setter/property annotation to remain unchanged,
relying on the forward-import under TYPE_CHECKING to satisfy the linter.

[nina-kollman]

Can you pls explain why you need this change ?

[tassadarius]

As laid out already in my initial PR description (see section Reasoning)

You let everybody who has pandas installed pay a 35 MB tax, when loading this library, where it is not clear why the experimental and datasets should be loaded alongside the main client.

If it is experimental code, it should not be loaded automatically into the environment of consumers of the library.

[nina-kollman]

Thanks for the detailed explanation @tassadarius , this is a valid concern.
We're currently working on some architectural changes to the SDK, and as part of that effort, we're planning to introduce package extras to properly address this.
This is on our roadmap but will take some time to implement properly without breaking existing users.
In the meantime, if this is blocking you, you could apply a local patch to defer the import — for example, by setting an environment variable check before the pandas-dependent imports in your local installation.