mirror of
https://github.com/MBombeck/HealthLog.git
synced 2026-11-08 20:42:43 +00:00
Marc-André Bombeck
a451a114a2
The profile section in docs/contributing/benchmarks.md says how to run scripts/bench/boot-profile.ts, what the profile of the v1.44 runtime found (the worker graph as one 753 ms block, a live heap of about 165 MB of which the query compiler holds 50 and the catalogs 33) and what changed. scripts/bench/results/v1.45.0-*.md hold the before and after numbers per area, from back-to-back runs on the same machine. The size baselines take the lowered numbers.
2.7 KiB
2.7 KiB
v1.45.0: read path
Request benchmark on the seeded account (1,265,561 rows), production build,
web process, heap ceiling 792 MB, 20 requests per row, one at a time.
v1.45.0-before is the v1.44 runtime (1915f9152), v1.45.0-final this
branch; both on the same machine within the hour, load average 7 to 30
(other work on the host), so latency compares within a factor of noise and
statements exactly. Cold means every request follows a one-row manual sync.
| request | before p50 / p95 ms | after p50 / p95 ms | statements before / after |
|---|---|---|---|
| dashboard snapshot, cold | 1482 / 1526 | 545 / 593 | 158 / 153 |
| dashboard summary, cold | 170 / 221 | 87 / 101 | 31 / 30 |
| series pulse 3650 d, cold | 1627 / 1665 | 335 / 349 | 8 / 7 |
| series pulse 3650 d, warm | 2091 / 2435 | 328 / 346 | 8 / 7 |
| series pulse 365 d, cold | 184 / 193 | 41 / 56 | 8 / 7 |
| series pulse 30 d, cold | 24 / 31 | 19 / 27 | 8 / 8 |
| insights health status, cold | 245 / 401 | 156 / 190 | 68 / 67 |
What moved them:
- JIT off for app sessions. Postgres compiled every fold whose estimated
cost passed
jit_above_cost; on the ten-year pulse pick that was 100 to 140 ms of a 220 ms statement (EXPLAIN ANALYZE: 395 ms with JIT, 220 ms without). - The canonical-source pick groups first.
DISTINCT ONnow sorts the distinct(type, day, source)triples instead of every raw reading (the ten-year pulse pick sorted 260,000 rows and spilled to disk); 220 ms to 175 ms on its own. Parity:tests/integration/canonical-source-pick-parity.test.tsruns the old statement beside the new one over five reader shapes and two ladders on a fixture with overlapping and off-ladder sources, tombstones, empty days and readings around midnight; identical rows. - The dense series reads its rows once. Buckets and strip figures come
from one statement over a materialised window instead of two statements
that each picked the canonical source and scanned the readings. Parity:
tests/integration/dense-series-read-parity.test.ts, the two old statements as reference, pulse (two ranked sources), glucose and oxygen saturation, day and hour grains, UTC, Europe/Berlin, Asia/Kolkata and America/New_York; identical buckets and figures (relative 1e-9). - Closed blocks survive a sync. The slice's pre-fold remainder (every
reading older than five years, the snapshot's largest statement at 357 ms)
and the day view's past history blocks are cached under a key that names
the instant their readings end before; a sync of newer readings no longer
drops them (
src/lib/cache/closed-block.ts). The values are the same statement's; only how long they are kept changed.