perf: implement lazy imports for heavy dependencies

[dimitri-yatsenko]

Summary

Implement lazy imports to reduce import datajoint time, especially on macOS where import overhead is significantly higher (~6s vs ~1s on Linux).

Problem

Heavy dependencies are loaded eagerly at import time even if unused:

• diagram.py imports networkx and matplotlib (~3s on Mac)
• admin.py imports pymysql via connection (~2.5s on Mac)
• cli.py imports click

Solution

Use __getattr__ in __init__.py to defer loading until first access:

_lazy_modules = {
    "Diagram": (".diagram", "Diagram"),
    "Di": (".diagram", "Diagram"),
    "ERD": (".diagram", "Diagram"),
    "kill": (".admin", "kill"),
    "cli": (".cli", "cli"),
}

def __getattr__(name: str):
    if name in _lazy_modules:
        module_path, attr_name = _lazy_modules[name]
        module = importlib.import_module(module_path, __package__)
        return getattr(module, attr_name)
    raise AttributeError(...)

Results

Before (estimated from issue):

• Mac: ~6s for import datajoint
• Linux: ~1s

After:

• Basic import: ~0.75s (Mac)
• Diagram access adds: ~0.25s when needed

What's lazy vs eager

Eager (always loaded):

• Schema, Table, FreeTable
• Manual, Lookup, Computed, Imported, Part
• Connection, conn, config
• Expression classes (Not, AndList, Top, U)
• errors, DataJointError
• Codec API

Lazy (loaded on first access):

• dj.Diagram, dj.Di, dj.ERD → networkx, matplotlib
• dj.kill → admin module
• dj.cli → click

Test plan

• Unit tests for lazy import behavior
• Verify core functionality available immediately
• Verify lazy modules load on access
• Verify Diagram aliases work correctly

Closes #1220

---

🤖 Generated with Claude Code