[haeganm]

Histogram leaves out _sum when its lowest bucket is negative, since OpenMetrics 1.0 says a histogram with negative buckets MUST NOT contain a sum. MultiProcessCollector never did that check, so in multiprocess mode the same histogram still had a _sum, OM 1.0 output included:

# Histogram('temp_delta', 'help', buckets=[-1, 0, 1]).observe(-0.5), multiprocess
temp_delta_sum -0.5
temp_delta_bucket{le="-1.0"} 0.0
temp_delta_bucket{le="0.0"} 1.0
temp_delta_bucket{le="1.0"} 1.0
temp_delta_bucket{le="+Inf"} 1.0
temp_delta_count 1.0

OpenMetrics 2.0 goes the other way (every histogram has count and sum), so per review this keeps _sum in the samples everywhere and filters it at exposition instead:

• Histogram always emits _sum
• openmetrics.exposition.generate_latest drops _sum for histograms with a negative bucket when version is below 2.0.0
• the histogram docs say which output leaves it out

side effect: the classic text format now shows _sum for these histograms in single process too.

test_histogram_negative_buckets in tests/test_multiprocess.py fails on master (b'h_sum' unexpectedly found in the OM 1.0 output) and passes here, and the OM exposition test now checks that 1.0 leaves _sum out and 2.0 keeps it. full suite passes on 3.12 (Windows), and flake8, isort and mypy are clean.

@csmarchbanks

[csmarchbanks]

Since OM 2.0 is adding sum back in everywhere I would learn towards at lest keeping it in the samples. That said, we should probably filter it out in exposition for at least OM 1.0 users.

[haeganm]

makes sense, flipped it in ca715c8. Histogram now always keeps _sum (multiprocess already did, so that part of the diff is gone), and openmetrics.exposition.generate_latest drops _sum for histograms with a negative bucket when version is below 2.0.0. so 1.0 output has no sum for these and 2.0 gets it.

one side effect, the classic text format now shows _sum for these in single process too, since that format never had the rule. happy to filter it there as well if you'd rather.

one more thing on 1.0: its text format says _count is there if and only if there's a sum, so a fully valid 1.0 output would drop _count for these too. that's what the OM parser trips on (_sum/_gsum must be present if _count is present), same on master. want me to drop it here? the +Inf bucket keeps the same value. the parser also rejects a negative _sum in 2.0 output (Counter-like samples cannot be negative), but 2.0 only treats sum as a counter when nothing negative was observed. left that one out of this PR.