[zzylol]

Stack: main → this PR → no-forward mode.

Why

After #830, the remaining per-sample work outside the summary state update was label-set handling. Remote Write decoding built a new BTreeMap<String, String> per series per request. Each raw-sample window's per-series cell cache and the per-sample group numbers then hashed and compared that whole map once per run of samples, and the metric was looked up in it for every sample. With one sample per series per request, that is per sample.

What

• Each label set is interned once, when Remote Write or a raw-sample fetch from Prometheus delivers it. An ahash (fixed seeds) of its label bytes finds or creates one shared Series, with a dense u32 id and its metric (series_ids).
• RawSample::labels is an Arc<Series>, which derefs to the label map:
  • the raw-sample window's per-series cells (SeriesWindow), the per-sample group numbers and the instant selector's latest-sample map key by the id;
  • the metric is read from the series, not looked up in the label map.
• Text-column dictionaries use ahash.

Review follow-ups from #830, in the second commit:

• Per-series state follows the active series. An open pane numbers its own groups, so it holds a summary state only for the series that took a sample in it (before, ensure_groups gave every open pane a state for every series ever numbered). A series unseen for longer than its definition's retention_ms plus one pane is forgotten, with its cached cells. A query-time RawSampleWindow forgets series unseen for the plan's longest retention past its width.
• One lookup per sample. A sample looks its series up once, by the interned u32 id, and a run of samples into one pane reuses the group number there; the label map is not hashed per sample.
• Resent samples. A sample not after its series' previous one in the pane (a Remote Write retry, for instance) is dropped and counted in the new samples_dropped_duplicate status key, as Prometheus drops it. Before, a counter's state refused it and the whole pane failed; any other summary state counted it twice.
• calibrate returns an error (the pane fails and the next pane retries) instead of defaulting the output timestamp offset to 0.
• Renames: Pane → SummaryInstance, SeriesWindow → RawSampleWindowBatcher (raw_sample_window.rs).

The SDS is unchanged: its panes are the plan's output batches, whose series identity column is dictionary-encoded text in the plan's schema.

How

series_ids::intern_pairs sorts the Remote Write pairs by name only when they are not already sorted. It hashes each name and value with its length, so a byte moving across a label boundary changes the hash. It looks the hash up under a read lock and compares labels only against that hash's entries; only a new label set takes the write lock. The interner is process-wide and never forgets a series, so its memory grows with every distinct series ever seen (a follow-up if series churn matters).

Usage

No new flags. /api/v1/status/asap gains precompute.samples_dropped_duplicate.

Before this PR

Per series per request: a new label map. Per run of samples in each per-series cache: a SipHash of the map plus a map comparison. Per sample: a __name__ lookup.

After this PR

Per series per request: one ahash over the label bytes and one comparison with the interned set. Everything downstream uses the u32 id.

Verification

cargo test --workspace --locked passes (fmt and clippy not run for this draft; the benefit harness was not rerun). Tests that build RawSamples now intern their label maps (series_ids::intern). New series_ids tests, which fail without an interner:

• one label set is one series whether it comes as a map or as pairs in any order;
• a byte moving across a label boundary makes another series;
• interning from 8 threads at once gives each label set one id.

Also new:

• per_sample::a_resent_sample_is_dropped_not_failing_or_doubling_the_pane: Example 1 with one sample resent. One duplicate is counted, no pane fails, and every pane equals the plan's build without the resent sample.
• precompute::tests::per_series_state_follows_the_active_series: Example 4b, 10 series for 25 min, then one series to 200 min. The open pane holds one state, and nine series are forgotten.

The parity test of #830 (per_sample_panes_equal_the_batch_build) and the query_time_reference comparisons still pass.