I've poked around both the internal URL that Jupyter kernel is running on and some of the files in /sagemaker/.jupyter but no luck so far - I can find plenty of kernel info, but not session

environ{'PYTHONNOUSERSITE': '0',
        'HOSTNAME': 'gfp-science-ml-t3-medium-d579233e8c4b53bc5ad626f2b385',
        'AWS_CONTAINER_CREDENTIALS_RELATIVE_URI': '/_sagemaker-instance-credentials/xxx',
        'JUPYTER_PATH': '/usr/share/jupyter/',
        'SAGEMAKER_LOG_FILE': '/var/log/studio/kernel_gateway.log',
        'PATH': '/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/miniconda3/condabin:/tmp/anaconda3/condabin:/tmp/miniconda2/condabin:/tmp/anaconda2/condabin'...

nice! Just tested it on my end as well, looks like it works!

lots of things like {"__timestamp__": "2023-02-23T23:49:23.285946Z", "__schema__": "sagemaker.kg.request.schema", "__schema_version__": 1, "__metadata_version__": 1, "account_id": "", "duration": 0.0007679462432861328, "method": "GET", "uri": "/api/kernels/6ba227af-ff2c-4b20-89ac-86dcac95e2b2", "status": 200}

As another test I ran Jupyter Lab locally using the same custom Docker container that we're using for Sagemaker Studio, and it works great there, just like the native local Jupyter Lab. So it's seemingly not the image, but maybe something to do with how Studio runs it as a kernel.

I additionally tried using a Sagemaker Notebook instance, to see if it was the kernel dockerization that Studio uses that was messing things up. But it seems to actually log less information from a Notebook instance vs Studio .


![image](https://clearml-web-assets.s3.amazonaws.com/scoold/im...

if there are any tests/debugging you'd like me to try, just let me know

sh-4.2$ cat /var/log/studio/kernel_gateway.log | head -n10
{"__timestamp__": "2023-02-23T21:48:28.036559Z", "__schema__": "sagemaker.kg.request.schema", "__schema_version__": 1, "__metadata_version__": 1, "account_id": "", "duration": 0.0012829303741455078, "method": "GET", "uri": "/api", "status": 200}
{"__timestamp__": "2023-02-23T21:48:39.111068Z", "__schema__": "sagemaker.kg.request.schema", "__schema_version__": 1, "__metadata_version__": 1, "account_id": "", "duration": 0.00128793...

and cat /var/log/studio/kernel_gateway.log | grep ipynb comes up empty

done: None

cool, thanks! the first one was what I had thought of but seemed unpythonic, so I'll give the second a shot

I think it just ends up in /home/sagemaker-user/{notebook}.ipynb every time

if I change it to 0.0.0.0 it works

local Jupyter Lab:

if I instead change the request url to f"http://{server_info['hostname']}:{server_info['port']}/api/sessions" then it gets a 200 response... however , the response is an empty list

the server_info is

[{'base_url': '/jupyter/default/',
  'hostname': '0.0.0.0',
  'password': False,
  'pid': 9,
  'port': 8888,
  'root_dir': '/home/sagemaker-user',
  'secure': False,
  'sock': '',
  'token': '',
  'url': '

',
  'version': '1.23.2'}]

Yeah, I think that'd work

I'll give it a shot and see! Just setting up a test server now, so it's still a hypothetical question just from reading the docs so far

I'm not sure if Subprojects will work for that - can you use the Web UI to compare the artifacts from two separate subprojects?

So I'm thinking maybe a Project for each thing we're forecasting, and then new Tasks for each time we run it

awesome, I'll test it out - thanks for the tips!

And then we want to compare backtests or just this week's estimates across multiple of those models/branches

None

cool, if I call clearml.binding.hydra_bind.PatchHydra._register_omegaconf(cfg) directly inside the notebook I can manually cause it to log the Hydra config, it just doesn't seem to autodetect if you're doing a manual call to Hydra Compose

My use case is running forecasting models in production across multiple businesses

I could just loop through and create separate pipelines with different parameters, but seems sort of inefficient. the hyperparameter optimization might actually work in this case utilizing grid search, but seems like kind of a hack

those look like linear DAGs to me, but maybe I'm missing something. I'm thinking something like the map operator in Prefect where I can provide an array of ["A", "B", "C"] and run the steps outlined with dotted lines independently for each of those are arguments

But we're also testing out new models all the time, which are typically implemented as git branches - they run on the same set of inputs but don't output their results into production