Is it planned to merge this PR?

@leokster Taking a look now – likely this PR will require some changes before we can merge

@szokeasaurusrex you can do the review, I will adjust accordingly

Might be interesting to review the PR for Langchain #2911

I am not familiar with Ray, but if it is a framework similar to Langchain, then following a similar model should be ideal

cc @colin-sentry if you have any thoughts on this integration

Hey @glowskir, thanks for reaching out again and for contributing this PR.

We are currently working on an initiative to support AI tools in the Sentry SDK, and we would be interested in learning how you are using Ray and what you hope to be able to achieve by integrating Ray with Sentry.

If you are interested, we can set up a customer call with you and anyone else interested in seeing a Ray integration in Sentry, so that we can learn how you are using Ray and how a Ray integration in the SDK could help you. You can reach me at [email protected] – please let me know when you would have time.

@szokeasaurusrex This PR introduces RayIntegration to enable out of the box distributed tracing support. Ray has high level task interface and we want to see chain of tasks as single transaction in performance tab. At the moment of creating this PR there was no Opentelemetry support, now in newer versions of ray one can use Opentelemetry facitlities provided by Ray. I don't know how Sentry integrations relate to Opentelemetry hooks tbh, whether you are planning to adapt them or ignore them but this is a thing to consider.

@glowskir, understood. In that case, we can help to get this PR merged.

Before I do an in depth review, I wanted to mention a few things that might be helpful here. First of all, since it looks like Ray is mostly used with AI, it might make sense to design this integration so it is compatible with our LLM Monitoring product. If you need some inspiration for how to do this, you can look at our existing Langchain integration, which works with the LLM Monitoring feature.

Additionally, I noticed while looking at your PR that you are using sentry_sdk.Hub in several places. sentry_sdk.Hub is deprecated as of 2.0.0, and although some of our existing code uses it, we are going to migrate all of these uses away from the Hub API and would like to avoid introducing any new Hub usages. Please see here for a guide of how to migrate Hub calls to the new Scope API.

Please let me know whether you have any questions! I am happy to help

Hi @glowskir , i tried to use your integration with a sample ray app (I have never used ray before)

Can you tell me how I can initialize sentry in the ray worker that is started when I call ray.init()? Is there an event/signal I can listen to?

Codecov Report

Attention: Patch coverage is 2.81690% with 69 lines in your changes missing coverage. Please review.

Project coverage is 79.33%. Comparing base (2c1e31c) to head (4c5b16d).
Report is 1 commits behind head on master.

✅ All tests successful. No failed tests found.

@@            Coverage Diff             @@
##           master    #2444      +/-   ##
==========================================
- Coverage   79.71%   79.33%   -0.38%     
==========================================
  Files         132      133       +1     
  Lines       14264    14335      +71     
  Branches     3003     3013      +10     
==========================================
+ Hits        11370    11373       +3     
- Misses       2071     2140      +69     
+ Partials      823      822       -1     

Ok, got it to work. Did some more cleanup and the transactions/spans look nice:

This is the ray app I used for testing: https://github.com/antonpirker/testing-sentry/tree/main/test-ray

I fixed the crash when using Actors. Performance data is still not captured when using Actors but at least we are not crashing users code anymore. I also mention this in the docs, so everyone knows that Actors are not yet supported.

Thanks for the work on this ya'll, excited to use this when merged.