The producer-only path left older hours stuck at 0% HRRR coverage
forever once their fetch tasks drained: the rolling-window cron only
re-passes the prior hour, so manual lookback runs (or any
out-of-window hour) wrote NULL samples + enqueued tasks, then never
re-ran to UPSERT the now-landed HRRR data.
Fix: split the responsibility cleanly.
* CalibrationSampler.build_for_hour/1 now returns
%{upserted: int, missing_hrrr_cells: int} so callers know whether
the producer enqueued anything (i.e. whether a follow-up makes
sense). Spec-tightened with a new @type result.
* PskrCalibrationWorker, after each hour's sampler pass, schedules a
follow-up of itself for that exact hour 10 min later when missing
> 0. The follow-up carries args %{"hour_utc" => ..., "_follow_up"
=> true}; the sentinel prevents follow-ups from chaining further
follow-ups (rolling-window cron is the safety net).
Tests:
* Updated 4 existing sampler tests to the new map return shape.
* Moved the follow-up assertions from the sampler suite into the
worker suite where they belong (sampler is now scheduling-free).
* Added 3 worker tests: schedules-when-missing, no-schedule-when-
covered, follow-up-doesn't-chain.
3313 tests + 228 properties, 0 failures.
Two performance fixes the property tests pinned down:
* CalibrationSampler.build_for_hour/1 walked every cell through
nearest_hrrr/3 twice — once in the producer and again in
build_sample. Now computed once into a (band, lat, lon) → hrrr_or_nil
map, halving the O(cells × profiles) scan (~10M comparisons saved
per fire at typical sizes).
* PartitionManager.ensure_quarterly_partitions/2 issued one
pg_inherits scan per (parent × lookahead). Refactored to one scan
per parent with in-memory coverage check across all quarters.
Property tests for both modules:
* PartitionManager: tile-coverage invariant (no gaps, no overlaps),
exact 3-month bounds, name format, multi-call idempotence,
multi-parent independence — caught a real NaiveDateTime sort bug
in the original test helpers. Default Enum.sort_by/2 falls back to
term comparison on NaiveDateTime; now uses NaiveDateTime as the
comparator module.
* CalibrationSampler: enqueued points = unique missing midpoints,
empty enqueue when fully covered, sample row count invariant under
HRRR availability.
Saved operational gotchas to CLAUDE.md (Oban leader election spans
all app=prop pods, kubectl exec deploy/prop ambiguity, half-year
partition coexistence, HRRR f000 publish lag).
3310 tests + 228 properties, 0 failures.
Adds the test that verifies the full producer→drain→re-pass loop:
sample written with NULL HRRR fields + fetch task enqueued, then once
an HRRR profile lands, the next sampler pass UPSERTs the same
calibration_samples row (id preserved) with non-NULL weather data.
Also tightens dialyzer surface:
* PartitionManager defines a `result()` typedoc and uses it on both
public functions instead of inlined tuple types.
* PartitionMaintenanceWorker.perform/1 + the lookahead/1 helper get
explicit specs.
* CalibrationSampler.enqueue_missing_hrrr/3 gets a spec covering its
group_by-shape input map and HRRR profile list.
`mix precommit` clean (3310 tests, 0 failures). Local `mix dialyzer`
hits an OTP 28.4 vs 28.5 toolchain mismatch unrelated to this change.
The CalibrationSampler joined PSKR midpoints against hrrr_profiles via
nearest-within-window, but post-Phase-3 Stream C nothing bulk-populates
the grid — hrrr_profiles only sees per-QSO point fetches from the Rust
hrrr-point-worker. Result: 0 / 16,042 prod samples had HRRR data.
Producer pattern: when a cell has no profile in the lookahead window,
enqueue a row into hrrr_fetch_tasks at the cell's HRRR-grid-rounded
midpoint. The Rust worker drains it; the next 5-min rolling-window
pass UPSERTs the same sample with non-NULL weather fields. Idempotent
with the existing on-conflict design — no extra worker, no churn on
cells already covered, multiple bands sharing a midpoint dedupe to a
single fetch via Enum.uniq + jsonb point-union in HrrrPointEnqueuer.
PSK Reporter is now an always-on data feed, so the calibration
corpus that recalibration will eventually train against can grow
hourly from when the firehose started. One row per (hour, band,
0.125° midpoint cell) joins three sources:
* `pskr_spots_hourly` — observed spot density (truth signal)
* `hrrr_profiles` nearest match at the cell midpoint at hour
boundary (atmospheric features: T, Td, PWAT, P, dN/dh, HPBL,
ducting flag)
* `geomagnetic_observations` latest Kp at the hour (space weather)
Predicted scores are intentionally NOT stored — they're a function
of the algorithm version under evaluation. Storing only features
keeps the corpus stable across every weight refit; the recalibrator
computes predictions on demand.
Components:
* Migration `20260504210756_create_pskr_calibration_samples` —
new table + unique index on (hour, band, midpoint_lat, lon),
plus a midpoint spatial index on `pskr_spots_hourly` to keep
the join cheap.
* `Pskr.CalibrationSample` — schema mirror of the table with
the same `(hour, band, midpoint_lat, lon)` unique constraint.
* `Pskr.CalibrationSampler.build_for_hour/1` — pulls all spots
for the hour, snaps midpoints to the propagation grid (0.125°),
builds an in-memory HRRR index over a ±0.07°/±60 min window,
and bulk-upserts samples. Idempotent — re-runs upsert.
* `Workers.PskrCalibrationWorker` — Oban cron entry on the
`:backfill_enqueue` queue. Default args target the previous
full hour; explicit `hour_utc` arg reruns any hour.
* Cron `25 * * * *` — fires past HRRR analysis publish window
(~HH:50→HH:05) and PSKR aggregator's 60s flush.
Forward-only: PSKR has no historical archive, so the corpus only
grows from when the feed started. Recalibration weight refits
should wait for ~30 days / ~10k samples to cover diurnal and
synoptic variability.
Tests cover cell-snapping, HRRR feature joining, Kp stamping,
median-distance aggregation, mode dedup, idempotent reruns, and
the worker's default-hour and explicit-hour args (12 new tests,
all passing).
Backfill pipeline untouched — none of these changes feed into
contact enrichment.
