尚未翻译

本页面尚未翻译成中文，因此显示英文原文。帮助翻译

gettext Module Complexity¶

The gettext module looks up translated messages in GNU .mo catalogs. Loading a catalog reads the whole file and decodes every message into a dictionary, once; after that a lookup is a dictionary probe. translation() keeps each parsed catalog for the life of the process, but the module-level gettext() family calls translation() again for every message, so each call repeats the search for catalog files.

k is the characters in the message id being looked up, the context included for the p variants; b is the bytes in the .mo files a call parses; f is the catalogs in a fallback chain; d is the candidates find() considers, which is every requested language expanded into its locale variants (up to eight per language), each at most one existence check. A filesystem existence check and one evaluation of a catalog's plural formula are treated as O(1).

Complexity Reference¶

Module-level functions¶

Operation	Time	Space	Notes
`gettext.bindtextdomain(domain, localedir=None)`	O(1)	O(1)	Records where the module-level functions look for `domain`, and returns the current binding
`gettext.textdomain(domain=None)`	O(1)	O(1)	Sets or returns the domain `gettext()`, `ngettext()`, `pgettext()` and `npgettext()` use
`gettext.gettext(message)`, `gettext.dgettext(domain, message)`, `gettext.ngettext(msgid1, msgid2, n)`, `gettext.dngettext(domain, msgid1, msgid2, n)`	O(d² + b + f·k)	O(d + b)	Each call runs `translation()` again, so the file search repeats; b is zero once the catalogs are cached. With no catalog found, the message comes back unchanged
`gettext.pgettext(context, message)`, `gettext.dpgettext(domain, context, message)`, `gettext.npgettext(context, msgid1, msgid2, n)`, `gettext.dnpgettext(domain, context, msgid1, msgid2, n)`	O(d² + b + f·k)	O(d + b + f·k)	As above, and the context key is built again on every call

find and translation¶

Operation	Time	Space	Notes
`gettext.find(domain, localedir=None, languages=None, all=False)`	O(d²)	O(d)	At most one existence check per candidate, stopping at the first hit unless `all=True`; candidates are deduplicated by list scan. Languages default to `LANGUAGE`, `LC_ALL`, `LC_MESSAGES` or `LANG`, and a `C` entry ends the search
`gettext.translation(domain, localedir=None, languages=None, class_=None, fallback=False)`	O(d² + b)	O(d + b)	b counts only files not parsed before: catalogs are cached per class and absolute path for the life of the process, and a changed file is not reread. Returns a copy chaining every catalog found; raises `FileNotFoundError` when there is none, unless `fallback=True`
`gettext.install(domain, localedir=None, *, names=None)`	O(d² + b + p)	O(d + b + p)	`translation(..., fallback=True)` followed by its `install(names)`

NullTranslations¶

Operation	Time	Space	Notes
`gettext.NullTranslations(fp=None)`	O(1)	O(1)	Does not read `fp`: its `_parse()` does nothing
`NullTranslations._parse(fp)`	O(1)	O(1)	The hook a subclass overrides to read another catalog format; the constructor calls it when `fp` is given
`NullTranslations.gettext(message)`, `NullTranslations.ngettext(msgid1, msgid2, n)`, `NullTranslations.pgettext(context, message)`, `NullTranslations.npgettext(context, msgid1, msgid2, n)`	O(1)	O(1)	Returns the message unchanged, `msgid1` or `msgid2` by whether `n == 1`; with a fallback added, the call is the fallback's
`NullTranslations.add_fallback(fallback)`	O(f)	O(f)	Walks to the end of the chain and appends there
`NullTranslations.info()`	O(1)	O(1)	The catalog's header dictionary itself, not a copy
`NullTranslations.charset()`	O(1)	O(1)	From the catalog's `Content-Type` header; `None` without a catalog
`NullTranslations.install(names=None)`	O(1 + p)	O(1 + p)	p = names passed; binds `_` in `builtins`, plus whichever of `gettext`, `ngettext`, `pgettext` and `npgettext` are named

GNUTranslations¶

Operation	Time	Space	Notes
`gettext.GNUTranslations(fp)`, `GNUTranslations._parse(fp)`	O(b)	O(b)	Reads the whole file and decodes every message; raises `OSError` for a bad magic number or version, or a message that runs past the end
`GNUTranslations.gettext(message)`, `GNUTranslations.ngettext(msgid1, msgid2, n)`	O(k)	O(1)	A hit compares the message id with the stored one. A miss repeats the call on the fallback, so a chain of f catalogs costs O(f·k) time and O(f) space
`GNUTranslations.pgettext(context, message)`, `GNUTranslations.npgettext(context, msgid1, msgid2, n)`	O(k)	O(k)	The context key is built on every call; a miss repeats that on the fallback, O(f·k) time and space

Loading a Catalog¶

Parsing Reads the Whole File¶

GNUTranslations decodes every message in the file when it is built, whether or not any of them is ever asked for. It needs only an object with read(), so a catalog can come from memory.

import gettext
import io
import struct

def make_mo(messages):
    """Encode {msgid: msgstr} as a little-endian GNU .mo file."""
    entries = sorted({"": "Content-Type: text/plain; charset=UTF-8\n", **messages}.items())
    count = len(entries)
    tables, strings = [b"", b""], b""
    for column in (0, 1):
        for entry in entries:
            data = entry[column].encode()
            tables[column] += struct.pack("<2I", len(data), 28 + 16 * count + len(strings))
            strings += data + b"\0"
    header = struct.pack("<7I", 0x950412DE, 0, count, 28, 28 + 8 * count, 0, 0)
    return header + tables[0] + tables[1] + strings

catalog = gettext.GNUTranslations(io.BytesIO(make_mo({"Hello": "Hei"})))  # O(b)
assert catalog.gettext("Hello") == "Hei"  # O(k)
assert catalog.gettext("Goodbye") == "Goodbye"  # a miss returns the message
assert catalog.charset() == "UTF-8"  # O(1)
assert catalog.info() is catalog.info()  # O(1) - the header dictionary itself

try:
    gettext.GNUTranslations(io.BytesIO(b"not a catalog"))
except OSError as error:
    assert "Bad magic number" in str(error)
else:
    raise AssertionError("a file without the magic number was parsed")

The Catalog Cache¶

translation() searches for files every time, but parses each one only once per process: later calls get a copy that shares the parsed messages. The cache is never checked against the file, so a catalog rewritten while the program runs is not seen until it restarts.

import gettext
import pathlib
import struct
import tempfile

def make_mo(messages):
    """Encode {msgid: msgstr} as a little-endian GNU .mo file."""
    entries = sorted({"": "Content-Type: text/plain; charset=UTF-8\n", **messages}.items())
    count = len(entries)
    tables, strings = [b"", b""], b""
    for column in (0, 1):
        for entry in entries:
            data = entry[column].encode()
            tables[column] += struct.pack("<2I", len(data), 28 + 16 * count + len(strings))
            strings += data + b"\0"
    header = struct.pack("<7I", 0x950412DE, 0, count, 28, 28 + 8 * count, 0, 0)
    return header + tables[0] + tables[1] + strings

with tempfile.TemporaryDirectory() as localedir:
    mofile = pathlib.Path(localedir, "fi", "LC_MESSAGES", "app.mo")
    mofile.parent.mkdir(parents=True)
    mofile.write_bytes(make_mo({"Hello": "Hei"}))

    assert gettext.find("app", localedir, ["fi"]) == str(mofile)  # O(d²)

    first = gettext.translation("app", localedir, ["fi"])  # O(d² + b) - parses
    second = gettext.translation("app", localedir, ["fi"])  # O(d²) - cached
    assert first is not second and first.info() is second.info()

    mofile.write_bytes(make_mo({"Hello": "Moi"}))
    assert gettext.translation("app", localedir, ["fi"]).gettext("Hello") == "Hei"

    # No catalog: an error, or a NullTranslations when asked for one
    try:
        gettext.translation("app", localedir, ["de"])
    except FileNotFoundError:
        pass
    else:
        raise AssertionError("a missing catalog was found")
    null = gettext.translation("app", localedir, ["de"], fallback=True)
    assert type(null) is gettext.NullTranslations
    assert null.gettext("Hello") == "Hello"

Looking Up Messages¶

Bind Once, Not Per Message¶

gettext.gettext() and its siblings take no catalog: each call runs translation() for the current domain, searching the candidate directories again before the dictionary probe. Binding the lookup once pays for the search once, and every hit after that costs O(k).

import gettext
import os
import pathlib
import struct
import tempfile

def make_mo(messages):
    """Encode {msgid: msgstr} as a little-endian GNU .mo file."""
    entries = sorted({"": "Content-Type: text/plain; charset=UTF-8\n", **messages}.items())
    count = len(entries)
    tables, strings = [b"", b""], b""
    for column in (0, 1):
        for entry in entries:
            data = entry[column].encode()
            tables[column] += struct.pack("<2I", len(data), 28 + 16 * count + len(strings))
            strings += data + b"\0"
    header = struct.pack("<7I", 0x950412DE, 0, count, 28, 28 + 8 * count, 0, 0)
    return header + tables[0] + tables[1] + strings

with tempfile.TemporaryDirectory() as localedir:
    mofile = pathlib.Path(localedir, "fi", "LC_MESSAGES", "app.mo")
    mofile.parent.mkdir(parents=True)
    mofile.write_bytes(make_mo({"Hello": "Hei"}))
    os.environ["LANGUAGE"] = "fi"

    gettext.bindtextdomain("app", localedir)  # O(1)
    gettext.textdomain("app")  # O(1)

    # PER CALL: O(d² + f·k), plus O(b) the first time - the search runs for every message
    assert gettext.gettext("Hello") == "Hei"

    # ONCE: O(d² + b), then O(k) per message
    _ = gettext.translation("app", localedir).gettext
    assert _("Hello") == "Hei"

Plural Forms and Context¶

ngettext() asks the catalog's plural formula which form a count takes, then looks that form up; pgettext() joins the context to the message to make the key, which it does again on every call.

import gettext
import io
import struct

def make_mo(messages):
    """Encode {msgid: msgstr} as a little-endian GNU .mo file."""
    entries = sorted({"": "Content-Type: text/plain; charset=UTF-8\n", **messages}.items())
    count = len(entries)
    tables, strings = [b"", b""], b""
    for column in (0, 1):
        for entry in entries:
            data = entry[column].encode()
            tables[column] += struct.pack("<2I", len(data), 28 + 16 * count + len(strings))
            strings += data + b"\0"
    header = struct.pack("<7I", 0x950412DE, 0, count, 28, 28 + 8 * count, 0, 0)
    return header + tables[0] + tables[1] + strings

catalog = gettext.GNUTranslations(io.BytesIO(make_mo({
    "": "Content-Type: text/plain; charset=UTF-8\nPlural-Forms: nplurals=2; plural=n != 1;\n",
    "file\0files": "tiedosto\0tiedostoa",  # msgid1 NUL msgid2: form 0 NUL form 1
    "menu\x04Open": "Avaa",  # context EOT message
})))

assert catalog.ngettext("file", "files", 1) == "tiedosto"  # O(k)
assert catalog.ngettext("file", "files", 3) == "tiedostoa"  # O(k)
assert catalog.pgettext("menu", "Open") == "Avaa"  # O(k) - builds "menu\x04Open"
assert catalog.pgettext("door", "Open") == "Open"  # another context misses

Fallback Chains¶

Every catalog translation() finds joins one chain, in the order the languages were given. A message the first catalog lacks is asked of the next, so a miss costs a probe per catalog.

import gettext
import pathlib
import struct
import tempfile

def make_mo(messages):
    """Encode {msgid: msgstr} as a little-endian GNU .mo file."""
    entries = sorted({"": "Content-Type: text/plain; charset=UTF-8\n", **messages}.items())
    count = len(entries)
    tables, strings = [b"", b""], b""
    for column in (0, 1):
        for entry in entries:
            data = entry[column].encode()
            tables[column] += struct.pack("<2I", len(data), 28 + 16 * count + len(strings))
            strings += data + b"\0"
    header = struct.pack("<7I", 0x950412DE, 0, count, 28, 28 + 8 * count, 0, 0)
    return header + tables[0] + tables[1] + strings

with tempfile.TemporaryDirectory() as localedir:
    for language, messages in [("fi", {"Hello": "Hei"}), ("sv", {"Hello": "Hej", "Bye": "Hej då"})]:
        mofile = pathlib.Path(localedir, language, "LC_MESSAGES", "app.mo")
        mofile.parent.mkdir(parents=True)
        mofile.write_bytes(make_mo(messages))

    chain = gettext.translation("app", localedir, ["fi", "sv"])  # f = 2
    assert chain.gettext("Hello") == "Hei"  # O(k) - the first catalog has it
    assert chain.gettext("Bye") == "Hej då"  # O(f·k) - asked of the second
    assert chain.gettext("Thanks") == "Thanks"  # O(f·k) - nobody has it

    chain.add_fallback(gettext.NullTranslations())  # O(f) - appended at the end

Common Patterns¶

Installing `_` for a Whole Program¶

install() binds _ in builtins, so every module can call it without importing anything. With fallback=True behind it, a missing catalog is not an error: _ returns messages unchanged.

import builtins
import gettext

gettext.install("app", localedir="/nonexistent")  # O(d²) - no catalog, NullTranslations
assert builtins._("Hello") == "Hello"  # O(1)

gettext.NullTranslations().install(names=["ngettext"])  # O(1 + p)
assert builtins.ngettext("file", "files", 2) == "files"

Choosing a Language per Request¶

A server answering in each user's language needs one catalog chain per language, not per request. Build each once and keep it: every later request is a dictionary lookup.

import gettext

catalogs = {}

def catalog_for(language):
    if language not in catalogs:  # O(1)
        catalogs[language] = gettext.translation(  # O(d² + b), once per language
            "app", "/nonexistent", [language], fallback=True
        )
    return catalogs[language]

assert catalog_for("fi") is catalog_for("fi")
assert catalog_for("fi").gettext("Hello") == "Hello"  # no catalog installed

Performance Best Practices¶

✅ Do:

Bind translation(...).gettext once and call that, so the file search runs once, not per message
Keep one translation object per language when serving several, rather than calling translation() per request
Build a GNUTranslations directly for a catalog that changes while the program runs; translation()'s cache never rereads a file

❌ Avoid:

gettext.gettext() and its siblings in a hot loop - each call searches the locale directories again
Long languages lists - every entry expands into several candidates, each an existence check
Long fallback chains for messages that usually miss - every miss is asked of each catalog in turn

Version Notes¶

Python 3.11+: lgettext(), ldgettext(), lngettext(), ldngettext(), bind_textdomain_codeset(), NullTranslations.output_charset() and NullTranslations.set_output_charset() are removed, as is the codeset parameter of translation() and install(); install()'s names is keyword-only

locale - the locale names find() expands, and the environment it reads
builtins - where install() puts _
struct - the binary layout GNUTranslations unpacks