Python 3.15 adds a lazy import keyword that defers each module until first use

Until now, deferring an import meant moving the statement inside the function that used it and hoping the linter stayed quiet ^[3]. With the 3.15 keyword, imports stay at the top of the file and each module loads on first use ^[1]. In the walkthrough's example, `lazy import json` and `lazy from pathlib import Path` sit above a print call that loads neither module. json loads at the `json.loads` call and pathlib at `Path(".")` ^[4].

How much a project saves depends on which modules each run touches. If a run never reaches the code that needs json, json never loads ^[2]. A CLI whose `--help` path touches none of its heavy dependencies gets the full saving. One whose help path imports them anyway saves nothing. The walkthrough names CLI tools, serverless functions and test runners as the main beneficiaries ^[17], and it labels the load costs in its own diagram as illustrative ^[12]. We'd use the sampling profiler that ships in the same release ^[6] to find where a given tool's startup time goes before rewriting any imports. The walkthrough text available for this report lists the UTF-8 default and the removals without saying what they break.

The breakage it does document comes from laziness itself. A misspelled or missing module no longer raises at the import line. It raises at first use, though the traceback still includes the original import line ^[10]. The keyword is legal only at module scope. Inside a function, a class, or a try/except/finally block it is a SyntaxError, and star imports and `__future__` imports cannot be lazy ^[11].

For projects that still support older interpreters, the walkthrough points to a module-level list: `__lazy_modules__ = ["json", "pathlib"]` makes plain import statements for those modules lazy ^[9]. We'd expect libraries to adopt that form before the keyword. It is the one the walkthrough recommends for code that must keep running on older versions ^[9].

The second documented trap is in a new builtin. frozendict inherits from object, so `isinstance(x, dict)` returns False for it, and the walkthrough suggests testing against `collections.abc.Mapping` instead ^[14]. Its equality ignores insertion order, so two frozendicts with the same items compare equal however they were built ^[15]. The cleanest use is configuration. Passing `object_pairs_hook=frozendict, array_hook=tuple` to `json.loads` gives a parsed document that cannot be mutated after loading ^[16].