Build Out of Memory on DigitalOcean – Vectorization Engine Using Doc

[Samsuthin-Kaavian]

Hey team,

I’m currently working on a vectorization engine for document processing using Docling. While everything builds fine locally, I encountered a "Build Out Of Memory" error when deploying it to DigitalOcean.

Surprisingly, the Docker image size is 4.24 GB, which seems quite large. Here's my current Pipfile for reference:

toml
Copy
Edit
[[source]]
url = "https://pypi.org/simple"
verify_ssl = true
name = "pypi"

[scripts]
start = "fastapi dev index.py"

[packages]
fastapi = {extras = ["standard"], version = ""}
uvicorn = ""
openai = ""
python-dotenv = ""
qdrant-client = ""
docling = ""
boto3 = ""
requests = ""
jmespath = ""
tiktoken = ""

[dev-packages]

[requires]
python_version = "3.12"
If you need any additional details (logs, Dockerfile, etc.), feel free to ask. I’d appreciate any suggestions on optimizing the build or reducing the image size.

Thanks!

[dosubot[bot]]

Hi @Samsuthin-Kaavian! I'm Dosu and I’m helping the docling team.

The large Docker image and "Build Out Of Memory" error are likely caused by Docling's heavy dependencies (especially PyTorch) and automatic model downloads during the build, which can spike memory usage and bloat the image size—even if everything works locally. This is a common issue with Docling deployments on resource-constrained platforms like DigitalOcean [source].

To optimize:

• Defer model downloads: Instead of downloading models during the Docker build, configure your setup to download them at runtime or mount a pre-populated model directory as a volume. This can dramatically reduce both image size and build-time memory usage [source].
• Use a persistent model cache: Pre-load models into a persistent volume (outside the image) and set the DOCLING_SERVE_ARTIFACTS_PATH environment variable to point to this directory. This avoids repeated downloads and keeps your image slim [source].
• Tune concurrency and resource limits: Limit the number of worker threads and batch sizes in Docling’s pipeline options and via environment variables (like DOCLING_SERVE_ENG_LOC_NUM_WORKERS) to reduce memory spikes during processing [source].
• Review dependencies: In your Pipfile, check if you need all the listed packages in production. For example, boto3, openai, and tiktoken can be heavy—move anything not needed at runtime to [dev-packages] to keep the image lean.
• Reference optimized images: The official docling-serve images use multi-stage builds, minimal OS packages, and dependency grouping to avoid unnecessary bloat. Their Dockerfiles and deployment docs are a good reference for best practices [source].

If you can share your Dockerfile or more build logs, I can give more targeted advice. But the biggest wins will come from moving model downloads out of the build step and tuning your dependency list and concurrency settings.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}