Skip to content

BEP-1001-app-injection.md

Latest commit

 

History

History
64 lines (45 loc) · 2.51 KB

BEP-1001-app-injection.md

File metadata and controls

64 lines (45 loc) · 2.51 KB
Author Joongi Kim (joongi@lablup.com)
Status Draft
Created
Created-Version
Target-Version
Implemented-Version

App Injection

Motivation

  • Decouple the update cycle of container apps (e.g., Jupyter) and the container images (e.g., NGC PyTorch)
  • Allow users to pin a desired version of the apps (e.g., Jupyter Notebook v7 has many breaking changes and most nbextensions do not work)

Proposed structure

App Package

  • service-def.json (mounted as /opt/kernel/service-defs/{app-name}.json)
  • app/ (mounted as /opt/apps/{app-name}/)

Updates to Manager API

  • List the available app packages
  • Auto-populate required app packages into agent nodes when creating sessions and kernels
    (or, we need to assume that we have consistent app package files in all agent nodes...)
  • Specify the app packages when creating sessions
  • Perform a compatibility check by comparing some metadata
    (e.g., is it a static build or does it depend on a specific glibc version?)

Updates to Agent API

  • TODO

Migration of Existing Apps

Overall technical issues

  • If we mount the extracted app directories as read-only, some apps may not work as expected.
  • If the app depends on directories that only root can write to, it may not work as expected.
  • The delay required to load and extract the app packages
  • Managing the cached extracted app packages within the disk size limit of agents
  • Should we ship PBS (python-build-standalone) individually by each app?

Jupyter

Recently Jupyter has changed its configuration format. (TODO: issue link)

We need to generate version-specific .cfg file, by replacing the service definition to execute an (injected) script instead of directly writing the configuration file from the embedded template string.

Design discussion:

  • Make a wheelhouse directory and install them when first-launching the app
    vs. Make a pre-installed site-packages directory and mount subdirs into /opt/backend.ai?
  • We must install ipykernel at least in the target Python (the Python already-installed inside the container image) upon the first launch and generate the metakernel spec (like what we does in backend.ai-kernels). We need an explicit error reporting when this fails.

VSCode

  • Could we make a statically built app package?
  • Could we allow users to install their own extensions in container while the app package is read-only?