prometheus: scrape-path perf — histogram labels, text batching, protobuf reuse (tracking)

Motivation

src/core/prometheus.cc's scrape path (write_body() and friends) does more per-scrape work than it needs to: histogram/summary text output re-walks the label map once per bucket, text-path series writes go through the output stream one at a time, and the protobuf path allocates a fresh message/string/label-string per family instead of reusing one across the request. Three independent, minimal-diff PRs fix these, verified individually with tests/perf/metrics_perf.cc.

PRs

• PR1: render histogram/summary labels once per metric, not per bucket (branch prom-perf/1-histogram-labels)
• PR2: batch text-path series writes, avoid per-family name concat (branch prom-perf/2-text-batching)
• PR3: reuse protobuf message, buffer and label storage across families (branch prom-perf/3-protobuf-reuse)

Each is independently mergeable and individually benchmarked in its own PR description. This issue tracks the combined effect of stacking all three together with #3672 (prometheus: cache per-series aggregation key instead of rebuilding it every scrape).

Combined stack results (PR1+PR2+PR3+#3672 vs master)

Median of 3 runs, tests/perf/metrics_perf, -c1:

Case	Master inst/op	Stack inst/op (%Δ)	Master allocs/op	Stack allocs/op (%Δ)
test_few_metrics	8577.95	8709.07 (+1.53%)	22.125	22.125 (+0.00%)
test_large_families	1481.73	1436.61 (-3.05%)	0.165	0.165 (+0.00%)
test_large_families_int	991.52	943.87 (-4.81%)	0.165	0.165 (+0.00%)
test_large_families_int_aggr	857.72	312.49 (-63.57%)	0.158	0.268 (+69.62%)
test_many_families_int	2994.58	3106.81 (+3.75%)	7.273	7.273 (+0.00%)
test_many_families_int_aggr	4668.88	4445.79 (-4.78%)	12.289	12.399 (+0.90%)
test_middle_ground	1678.29	1648.84 (-1.75%)	0.846	0.846 (+0.00%)
test_middle_ground_int	1188.18	1156.67 (-2.65%)	0.845	0.846 (+0.12%)
test_middle_ground_protobuf	5773.06	4259.27 (-26.22%)	19.843	4.861 (-75.50%)
test_histogram	1715.90	1320.12 (-23.07%)	0.092	0.092 (+0.00%)
test_histogram_protobuf	416.19	344.57 (-17.21%)	1.226	0.548 (-55.30%)
test_histogram_aggr	245.03	203.33 (-17.02%)	0.113	0.118 (+4.42%)
test_name_filter_exact_match	170.18	170.29 (+0.06%)	0.538	0.538 (+0.00%)
test_name_filter_many_no_match	310.21	310.32 (+0.04%)	0.548	0.548 (+0.00%)

The two test_name_filter_* cases are flat — the fourth optimization in the original set (a hash-based name-filter lookup, targeting exactly those two cases) was dropped after investigation found no evidence of the many-exact-filter scrape pattern it optimizes (federation-style match[]=) in ScyllaDB's own monitoring config.

The aggregation allocs/op rises (_int_aggr/histogram_aggr) are #3672's cache, a one-time per-metadata-rebuild cost, not per-scrape — see #3672 for detail.

Per-PR benchmark tables (PR alone vs master) are in each PR's own description.

Generated with Claude Code