Compute the rate family's windows by an ASOF join, not a frame list (T-632)

[chasers]

TL;DR: A pushed rate, irate, increase, delta (and the rest of their family) now costs the same at any window. rate(m[1h]) over 7 d on 2.8 M samples: 22.3 s → 1.9 s, same answers.

Tracker: T-632.

Why

• The rate family gave every sample a list of every sample in its window (list(...) OVER a RANGE frame).
• Cost was samples × window. Grafana's $__rate_interval grows with the range, so long ranges paid twice.
• On the sandbox, sum(rate(m[1h])) over 7 d took 30.7 s, then 134 s. It pushed the query pod to 10 GiB and the liveness probe killed it.
• The scan was not the cost: count(*) over the same 2.8 M rows takes 0.4 s.

What changed

• Pushdown.Windows has a second descriptor for the edge rollups: rate, deriv_fast, increase, increase_pure, delta, idelta, irate, ideriv, first_over_time, present_over_time.
• Each sample unnests into the points where it is the window's last. That part is unchanged.
• An ASOF join to the first row of each timestamp group finds the window's first sample.
• Everything else a point needs is a lag/lead on those two rows: the sample before the window, the one after the first, the one before the last, and irate's earlier sample.
• Cost: samples + series × points, whatever the window.
• The edge rollups are now pushed past the 32-step gate (rate(m[1d]) at 15 s no longer falls back to Elixir).
• The whole-window rollups (stddev_over_time, changes, …) keep the list descriptor and the gate.
• Upgrade note in docs/deployment.md.

Measured

2.8 M synthetic counter samples, 7 d, step 600 s, sum by (op):

query	before	after	same answer
rate(m[615s])	7.0 s	1.8 s	✅
rate(m[30m])	11.9 s	1.6 s	✅
rate(m[1h])	22.3 s	1.9 s	✅
irate(m[1h])	22.1 s	2.6 s	✅

Tests

• ✅ Over-the-stack test: the pushed answer equals the Elixir evaluator's, byte for byte. Range, off-grid and instant grids, with duplicate timestamps and counter resets.
• ✅ That test adds windows of 9 m to 1 h at a 15 s step (more than 32 steps), including a window longer than the data.
• ✅ windows_test: SQL shape of the ASOF path and of the list path.
• ✅ plan/2: edge rollups are planned at 1d; whole-window ones are still not.
• ✅ mix precommit, reach.check --arch --smells.

Review

• A Fable-model review checked every descriptor column against the old list path, including duplicate timestamps, empty windows, per-series windows, instant queries and a series' first sample. Nothing blocking.
• Its small findings are applied: an unused window column is gone, the doc numbers match the measurements, and a longer-than-data window was added to the tests.

🤖 Generated with Claude Code