[chrikrah]

_validate_labelname in prometheus_client/validation.py:76 only checks that a name encodes as UTF-8 when legacy validation is off, which is the default, so an empty label name passes. The OpenMetrics 2.0 grammar allows no such key: label-key = label-name / DQUOTE escaped-string-non-empty DQUOTE. The library then writes the unquoted {="x"}, which its own parsers refuse.

# base ff9a2d3, python 3.12; repro.py builds Gauge('g', 'h', [''], registry=r), prints
# generate_latest(r) and feeds that text to parser.text_string_to_metric_families
$ python repro.py 2>&1 | grep -E '^(g|ValueError)'
g{="x"} 1.0
ValueError: unquoted UTF-8 metric name
ValueError: Invalid labels: ="x"
$ python repro.py 2>&1 | grep -E '^(g|ValueError)'   # this branch
ValueError: label name cannot be empty

On this branch the Gauge(...) call raises, worded like the error _validate_metric_name gives for an empty metric name. Exemplar keys go through the same function, so inc(exemplar={'': 'abc'}) is refused too. Legacy validation already rejected both through METRIC_LABEL_NAME_RE, and LabelName.IsValid in prometheus/common returns false for an empty name in UTF-8 mode.

parser.py:97 calls the same function, so both parsers now reject the quoted g{""="x"} 1.0 too, and a case in each parser's tests pins that. At base they returned labels={'': 'x'}. GaugeMetricFamily(labels=['']) in a custom collector still writes g{="x"} 1.0, because metric families skip every name check today, so I left them out of scope.

Verification

$ python -m pytest -q
431 passed, 14 skipped, 34 subtests passed   # this branch
428 passed, 14 skipped, 28 subtests passed   # base ff9a2d3
$ python -m pytest -q   # validation.py reverted, tests kept
9 failed, 428 passed, 14 skipped, 28 subtests passed   # six metric types, the exemplar, both parsers
$ flake8 prometheus_client/ tests/   # flake8 7.4.1
$ isort --check prometheus_client/ tests/   # isort 5.10.1
$ mypy prometheus_client tests/test_core.py tests/test_parser.py tests/openmetrics/test_parser.py   # mypy 2.4.0
Found 5 errors in 2 files   # identical at base: aiohttp stubs and metrics.py:184-210
# not run: tox on other interpreters, PyPy

There is no issue for this. #1236 edits two other lines of the same file and leaves _validate_labelname alone.

@csmarchbanks, this makes a call that used to succeed raise at construction. Its exposition already fails to parse, so I'd ship the raise without a deprecation warning first. Does that work for you?