Logging

[rozyczko]

Logging via the standard logging module

• Affected: easyscience.global_object.logger, all modules that
  previously used print() or warnings.warn() for runtime/deprecation
  messages.

Context

Until now, runtime information was presented in two ad-hoc ways:

• print() statements — which always write to stdout, cannot be
  silenced by the host application, and pollute notebooks, test reports,
  and downstream CLIs.
• warnings.warn(...) — which is better, but routes through Python's
  warning machinery. Warnings are easy to accidentally suppress
  globally, are de-duplicated by default (so a repeated message is shown
  once), and are awkward to redirect to a log file or structured sink.

Neither gives the host application real control over what is shown and
where it goes, which is the defining requirement for a well-behaved
library.

Decision

Route all informational, warning, and deprecation messages through
Python's standard logging module, under a package-root logger named
easyscience, exposed through a small controller
global_object.log (an instance of Logger).

Key rules:

1. A named logger hierarchy, not warning categories. Every subsystem
   logs under a dotted child of easyscience
   (easyscience.fitting.bumps, easyscience.legacy,
   easyscience.deprecated, easyscience.variable, …). Setting a level
   on a parent silences all of its children. This replaces what warning
   categories used to give us, but is filterable per-subsystem rather
   than per-category.

2. Library-safe by construction. The controller never calls
   logging.basicConfig() and never attaches a handler. EasyScience
   only creates log records; the host application decides where they
   go. With no handler configured, Python's logging.lastResort handler
   still prints WARNING and above to stderr, so a standalone user
   sees warnings out of the box, while INFO/DEBUG stay hidden until
   they opt in.

3. Child loggers inherit.
   global_object.log.getLogger('fitting.bumps') returns a logger left
   at NOTSET, so it inherits level and handlers from the easyscience
   root. Modules obtain their logger through this controller method
   (rather than logging.getLogger('easyscience.…') directly) so the
   controller remains the single point of control if the policy changes.
   The only exceptions are the modules that participate in
   GlobalObject's own construction (global_object/undo_redo.py,
   global_object/hugger/property.py), which call logging.getLogger
   directly to avoid a bootstrap cycle; they are already correctly
   namespaced.

4. Level control is a property. global_object.log.level is a
   property whose getter reads the live level off the package-root
   logger and whose setter applies it — there is no mirrored attribute
   that can drift out of sync. A bad level string raises ValueError;
   the EASYSCIENCE_LOG_LEVEL environment variable is parsed leniently
   (an unset or malformed value falls back to the default rather than
   breaking import).

How to use it

Emitting messages (library/internal code)

Get a child logger named for your subsystem and log on it:

from easyscience import global_object

log = global_object.log.getLogger('fitting.bumps')
log.warning('Fit did not converge within %d evaluations', n)

For top-level package messages (e.g. deprecation shims), the convenience
methods log on the root easyscience logger:

global_object.log.warning('module X is deprecated; import from Y instead')

Prefer global_object.log.getLogger(...) over
logging.getLogger('easyscience.…') so all output flows through the
controller.

Silencing / redirecting (consumers)

from easyscience import global_object

global_object.log.level = 'ERROR'          # suppress WARNING/INFO/DEBUG

Or via the standard library / environment, and the full set of recipes
(per-subsystem filtering, embedding in another library, routing to a
file, keeping test output clean), see the user guide:
docs/docs/user-guide/controlling-log-output.md.

Consequences

Positive

• Host applications and test frameworks get full control over output and
  its destination; nothing is hard-wired to stdout.
• Per-subsystem verbosity via the logger hierarchy.
• No global side effects on import (no basicConfig, no handlers).

Negative / trade-offs

• We lose Python's warning category machinery (-W flags,
  DeprecationWarning filtering, pytest filterwarnings). The
  hierarchy covers per-subsystem filtering, but for genuine API
  deprecations that external tooling should detect, emitting a real
  DeprecationWarning (in addition to, or instead of, a log message)
  remains a valid option to revisit.
• Code must obtain loggers through the controller, a small convention
  contributors need to follow.

[rozyczko]

Discussion in #252