[NUAAwqy]

Which issue(s) this PR fixes:

None — self-found while auditing NumericTimeParser#parse_float.

What this PR does / why we need it:

NumericTimeParser#parse_float (lib/fluent/time.rb) kept ten characters of the fractional part (nsec_s[0..9]), so a timestamp with ten or more decimal digits stored an nsec of 1_000_000_000 or more. EventTime assumes nsec fits inside a second: to_f/to_r report the wrong instant, to_time normalises into the next second, %N prints ten digits, and to_msgpack_ext forwards the oversized field as-is, so other fluentd nodes inherit it. Of fraction lengths 1..20, eleven produce an out-of-range nsec:

Fluent::NumericTimeParser.new(:float).parse("1700000000.1234567890").to_f
# => 1700000001.2345679   (should be 1700000000.123456789)

The fix: keep the first nine fraction digits (nsec_s[0, 9], sub-nanosecond digits dropped, which is what Time#nsec does). Padding and caching behaviour are untouched, so the hot path keeps its shape.

Per the review, the negative-timestamp adjustment introduced earlier in this PR was dropped again: it changed how negative values interact with existing formatting and MessagePack serialization, and EventTime is also defined in the Forward protocol, so changing its encoding needs a separate discussion. Negative values now parse exactly as before this PR.

Validation (Ruby 3.3.6, TZ=UTC):

check	result
new test (test/test_time_parser.rb)	31 tests, 0 failures, 0 errors
regression slice: test_time_formatter 77, test_event_time 22, test_formatter 36, test_mixin 21, plugin/test_parser 23, plugin/test_parser_json 66, plugin/test_parser_apache 3, plugin_helper/test_compat_parameters 10, plugin/test_out_exec_filter 22	all 0 failures, 0 errors
rubocop 1.92.1 + rubocop-performance 1.27.0	no offenses

plugin/test_in_exec reports 2 failures (can handle non-ascii characters) that reproduce identically at f984d65 with the fix reverted — the flaky encoding tests pinned in #5562/#5572, unrelated to time parsing.

Docs Changes:

None needed.

Release Note:

time: fix parse_float with ten or more fractional digits

• With time_type float, a time string with ten or more fractional digits was parsed into a wrong time. Digits beyond nanoseconds are now dropped.

[Watson1978]

Thanks for finding this. I reviewed 606851a on Ruby 4.0.9 (Linux).

Keeping nine digits of the fraction looks good. It fixes the out-of-range nsec.

You are right that master parses -0.5 as +0.5. However, the negative-timestamp adjustment changes behavior in ways that interact badly with existing formatting and MessagePack serialization. Details are in the inline comments. Let's keep this PR focused on the nine-digit limit and address negative timestamps separately.

Negative timestamps require a separate discussion of formatting and serialization compatibility. EventTime is also defined in the Forward Protocol, so changing its encoding affects other implementations too.

• MUST: Drop the negative-timestamp adjustment and its test. Both inline comments have suggestions for this.
• MUST: Update the PR description and the release note to reflect the narrower scope.

With the two suggestions applied, test/test_time_parser.rb (31 tests) and test/test_time_formatter.rb (77 tests) pass.

[NUAAwqy]

Thanks for the thorough review — both MUSTs are applied in 5173c74:

• Dropped the negative-timestamp adjustment and its test; negative values parse exactly as before. Agreed the formatting/serialization compatibility (and the Forward protocol EventTime encoding) needs a separate discussion.
• Rebased the PR description and the release note onto the narrower nine-digit-fraction scope.

test/test_time_parser.rb (31 tests) and test/test_time_formatter.rb (77 tests) pass with the change.

[Watson1978]

Thanks for the quick update.
LGTM