Docker image crashes upon querying

[r4881t]

Describe the bug

Short Issue: The docker container which runs a webserver running haystack, crashes while running a pipeline.

Long Text: I have a webapp which exposes query API. When someone makes a POST Call, the controller runs a RAG Pipeline. This keeps on crashing everytime the pipeline is run. I have been unable to identify any error log. This does not happen if i run the webserver natively on my OS

Error message No error message is thrown

Expected behavior The RAG Pipeline should run.

Additional context I have enabled DEBUG in haystack & my app logs and attaching the log file docker_crash.log. It crashes exactly after outputting Batches: 0%| | 0/1 [00:00<?, ?it/s]% text. My Dockerfile is pasted below

FROM  python:3.11 AS base

WORKDIR /app

COPY . .

RUN pip install --upgrade pip

RUN pip install --no-cache-dir -r ./requirements.txt

EXPOSE 8080

CMD ["uvicorn", "friday.server:app", "--host", "0.0.0.0", "--port", "8080"]

To Reproduce

FAQ Check

• [x] Have you had a look at our new FAQ page?

System:

• OS: Alpine
• GPU/CPU: CPU
• Haystack version (commit or version number): 1.22.1
• DocumentStore: ElasticSearch
• Reader: None. Reranker present
• Retriever: BM25Retriever, EmbeddingRetriever

[masci]

In the logs I see:

DEBUG - haystack.telemetry -  Telemetry couldn't make a POST request to PostHog.
Traceback (most recent call last):
  File "/usr/local/lib/python3.11/site-packages/haystack/telemetry.py", line 98, in send_event
    json.dumps({**self.event_properties, **dynamic_specs, **event_properties}, sort_keys=True)
  File "/usr/local/lib/python3.11/json/__init__.py", line 238, in dumps
    **kw).encode(obj)
          ^^^^^^^^^^^
  File "/usr/local/lib/python3.11/json/encoder.py", line 200, in encode
    chunks = self.iterencode(o, _one_shot=True)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/json/encoder.py", line 258, in iterencode
    return _iterencode(o, 0)
           ^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/json/encoder.py", line 180, in default
    raise TypeError(f'Object of type {o.__class__.__name__} '
TypeError: Object of type PromptTemplate is not JSON serializable

Can you try disabling telemetry by setting an environment variable HAYSTACK_TELEMETRY_ENABLED=False in your docker container?

[r4881t]

Tried with telemetry disabled. No change in the behaviour. I am on Apple M1. Here's additional details I did

1. Instead of docker run -p 8080:8080 --env-file .env friday, I went into an interactive terminal in the container and started the uvicorn from there. I notice that while pipe run, there's a segmentation fault that happens.
2. I noticed that I have use_gpu=True in my EmbeddingRetriever, so I turned it to False and rebuilt the image and repeated step 1
3. That didn't change anything, so I attached a debugger
4. I btd the debugger and attaching logs here. I am not much of an expert in low level debugging but it seems like something to do with pytorch libs. docker_crash_with_gdb.log

I will continue to fiddle more.

[masci]

Can you try setting also HAYSTACK_MPS_ENABLED : false?

[r4881t]

I tried with both HAYSTACK_MPS_ENABLED=False and HAYSTACK_MPS_ENABLED=false in the env, and it still crashes. Attached is the gdb log when HAYSTACK_MPS_ENABLED is set to false. docker_crash2.log

[masci]

I tried to reproduce on my Mac using the docker-compose file from here https://github.com/deepset-ai/haystack-demos/tree/main/explore_the_world pulling 1.21.1 but it works. If you can post a minimal setup I can reproduce locally I'll try again.

[alizingly]

I faced same issue but then realized that the problem happens due to OpenSearch wrong credentials, I don't know why the error message is so off the reason of the error