perf(infra): read the container pivot from a freshness slice, not the whole window

Over a 24-hour window the containers query did not complete at all — it hit the
statement timeout, so `containersInWindowCached` never got a value to store, the
300s TTL never applied, and every load paid the full timeout and rendered
nothing. Measured on prod TF against the demo project: 1.77s at 1h, 42.33s at
6h, timeout at 24h.

The window was the wrong input. The values shown are by definition the newest
datapoint per series, and a series that still exists reports every 15-60s, so
its newest datapoint is always within the last few minutes. Scanning further
back can only turn up older points that `rn = 1` then discards — while paying to
sort every one of them, with ~15 variant_get lookups per row along the way.

So the pivot now reads `[max(fromTime, toTime - 15m), toTime]`. Anchored on the
window's end rather than on `now`, so a historical from/to range reports the
state at the end of the range it selected, and a window shorter than the slice
still only reads itself. Result: 0.70s flat at every window — 63x at 6h, and 24h
goes from not completing to sub-second.

That drops a container which stopped reporting hours ago off a wide window,
instead of showing it with hours-old numbers sorted among the live ones by CPU.
That is the intended reading of the table, and keeping the old semantics is not
affordable: a presence-only query over the window costs 8.77s at 6h and 25.12s
at 24h, because the scan is the expense, not the pivot. The result summary now
says "reporting in the last 15m" rather than implying the picker's window.

Also fixes the two usage charts above the table, which rendered as empty panels.
They sized themselves purely by `aspect-ratio: 4/1`, so side by side on a 1600px
viewport each canvas got 56px — no plot area for a multi-series line chart with a
legend, even though both /chart_data requests returned full datasets. A height
floor takes the canvas to 140px; the ratio still governs growth on wide screens.

plans/infra-sessions-rum-performance.md has the measurements for all three
pages, including why the sessions query is deliberately left alone (every
rewrite that stays correct on both backends is only ~1.3x) and why the real fix
for RUM's 31x read amplification is the ingest-time category column, not SQL
shape. It also records the one UI bug found and not fixed: the containers table
is clipped rather than scrollable, and the obvious fix breaks sticky headers
app-wide.
This commit is contained in:
Anthony Alaribe 2026-09-01 01:50:15 +02:00
commit c2ffd32e85
8 changed files with 365 additions and 8 deletions