prop/lib/microwaveprop/pskr/calibration_sampler.ex

313 lines
12 KiB
Elixir
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

defmodule Microwaveprop.Pskr.CalibrationSampler do
@moduledoc """
Builds `Pskr.CalibrationSample` rows by joining PSKR spot density
with HRRR atmospheric state and SWPC geomagnetic state at a given
hour.
Workflow:
1. Pull every `pskr_spots_hourly` row for the target hour.
2. Snap each row's midpoint to the propagation grid (0.125°) and
bucket by `(band, snapped_lat, snapped_lon)`.
3. For each bucket, aggregate the spot stats (sum count, count
distinct paths, dedup modes, median distance).
4. For each bucket, look up the nearest `hrrr_profiles` row to the
midpoint at the target hour and copy the surface/refractivity
fields onto the sample. HRRR is pulled in one window query and
joined in-memory to keep the per-bucket cost a hashtable lookup.
5. Look up SWPC Kp once per run.
6. Bulk-upsert the result with `(hour_utc, band, midpoint_lat,
midpoint_lon)` as the conflict target.
The sampler is intentionally idempotent: re-running for the same
hour is a no-op except for the `updated_at` bump, so the worker
can safely overlap with manual reruns and the cron retry on
failure.
"""
import Ecto.Query
alias Microwaveprop.Pskr.CalibrationSample
alias Microwaveprop.Pskr.SpotHourly
alias Microwaveprop.Repo
alias Microwaveprop.SpaceWeather
alias Microwaveprop.Weather
alias Microwaveprop.Weather.HrrrPointEnqueuer
alias Microwaveprop.Weather.HrrrProfile
require Logger
@grid_step_deg 0.125
# HRRR domain is ~3 km native, so a ±0.07° (~7 km) lat/lon window
# plus ±60 min reliably finds at least one profile for any CONUS
# midpoint at most hours.
@hrrr_lookahead_deg 0.07
@hrrr_lookahead_seconds 3600
@typep cell_key :: {band :: String.t(), lat :: float(), lon :: float()}
@typedoc """
Result of a single `build_for_hour/1` pass:
* `upserted` — number of `pskr_calibration_samples` rows
inserted-or-updated.
* `missing_hrrr_cells` — number of distinct cell midpoints whose
HRRR profile was unavailable. The caller (typically
`Microwaveprop.Workers.PskrCalibrationWorker`) uses this to
decide whether to schedule a follow-up sampler pass after the
Rust hrrr-point-worker has had time to drain the just-enqueued
fetch tasks.
"""
@type result :: %{upserted: non_neg_integer(), missing_hrrr_cells: non_neg_integer()}
@doc """
Build calibration samples for one hour. Returns
`%{upserted: count, missing_hrrr_cells: count}`. Both counts are 0
when no PSKR spots were captured for that hour — for example,
during PSKR client downtime.
The `missing_hrrr_cells` count tells the caller whether the Rust
hrrr-point-worker has work to drain after this fire. The
`PskrCalibrationWorker` uses it to decide whether to schedule a
follow-up sampler pass for the same hour ~10 min later, so the
newly-landed HRRR data gets UPSERTed onto the samples we just
wrote with NULL fields.
"""
@spec build_for_hour(DateTime.t()) :: result()
def build_for_hour(%DateTime{} = hour_utc) do
hour_utc = align_to_hour(hour_utc)
spots = fetch_spots(hour_utc)
if spots == [] do
Logger.info("Pskr.CalibrationSampler: no spots for #{hour_utc} — skipping")
%{upserted: 0, missing_hrrr_cells: 0}
else
hrrr_index = build_hrrr_index(hour_utc)
kp = fetch_kp(hour_utc)
cells = bucket_by_cell(spots)
# Compute (band, lat, lon) → nearest_hrrr_or_nil ONCE. Both the
# producer (enqueue_missing_hrrr) and the per-bucket sample
# assembly need it, and `nearest_hrrr/3` is an O(profiles) scan;
# walking the cell list twice was ~10M extra comparisons per
# fire at typical sizes (~1k cells × ~5k profiles).
cell_hrrr =
Map.new(cells, fn {{_band, lat, lon} = key, _spots} ->
{key, nearest_hrrr(lat, lon, hrrr_index)}
end)
missing = enqueue_missing_hrrr(cell_hrrr, hour_utc)
samples = Enum.map(cells, &build_sample(&1, hour_utc, cell_hrrr, kp))
%{upserted: upsert(samples), missing_hrrr_cells: missing}
end
end
# ── Spot fetch + cell bucketing ─────────────────────────────────
defp fetch_spots(hour_utc) do
Repo.all(
from(s in SpotHourly,
where: s.hour_utc == ^hour_utc and not is_nil(s.midpoint_lat) and not is_nil(s.midpoint_lon),
select: %{
band: s.band,
midpoint_lat: s.midpoint_lat,
midpoint_lon: s.midpoint_lon,
spot_count: s.spot_count,
distance_km: s.distance_km,
modes: s.modes
}
)
)
end
defp bucket_by_cell(spots) do
Enum.group_by(spots, fn s -> {s.band, snap(s.midpoint_lat), snap(s.midpoint_lon)} end)
end
defp snap(deg), do: Float.round(deg / @grid_step_deg) * @grid_step_deg
# ── HRRR feature lookup ─────────────────────────────────────────
defp build_hrrr_index(hour_utc) do
time_start = DateTime.add(hour_utc, -@hrrr_lookahead_seconds, :second)
time_end = DateTime.add(hour_utc, @hrrr_lookahead_seconds, :second)
Repo.all(
from(h in HrrrProfile,
where: h.valid_time >= ^time_start and h.valid_time <= ^time_end,
select: %{
lat: h.lat,
lon: h.lon,
surface_temp_c: h.surface_temp_c,
surface_dewpoint_c: h.surface_dewpoint_c,
pwat_mm: h.pwat_mm,
surface_pressure_mb: h.surface_pressure_mb,
min_refractivity_gradient: h.min_refractivity_gradient,
hpbl_m: h.hpbl_m,
ducting_detected: h.ducting_detected
}
)
)
end
# Pick the HRRR row whose lat/lon is closest to (lat, lon). Linear
# scan over the in-memory window — at the typical ~5k profile
# window size and ~1k cells per hour, this is faster than 1k DB
# round trips. Returns `nil` if no profile is in the window.
defp nearest_hrrr(_lat, _lon, []), do: nil
defp nearest_hrrr(lat, lon, profiles) do
candidates =
Enum.filter(profiles, fn p ->
abs(p.lat - lat) <= @hrrr_lookahead_deg and abs(p.lon - lon) <= @hrrr_lookahead_deg
end)
case candidates do
[] -> nil
list -> Enum.min_by(list, fn p -> abs(p.lat - lat) + abs(p.lon - lon) end)
end
end
# ── HRRR fetch enqueue (producer side) ──────────────────────────
# For every PSKR cell that has no HRRR profile in the lookahead
# window, queue a point fetch at the cell midpoint. Dedup on
# `(rounded_hrrr_lat, rounded_hrrr_lon)` so multiple bands sharing
# a midpoint enqueue once, and let `HrrrPointEnqueuer.enqueue/1`
# union the points into the existing `hrrr_fetch_tasks` row for
# this hour rather than racing to insert duplicates.
#
# Returns the number of unique missing-HRRR cells. The caller uses
# that to decide whether a follow-up sampler pass is needed (see
# `Microwaveprop.Workers.PskrCalibrationWorker`).
@spec enqueue_missing_hrrr(%{cell_key() => map() | nil}, DateTime.t()) :: non_neg_integer()
defp enqueue_missing_hrrr(cell_hrrr, hour_utc) do
points =
cell_hrrr
|> Enum.flat_map(fn
{_key, %{}} -> []
{{_band, lat, lon}, nil} -> [Weather.round_to_hrrr_grid(lat, lon)]
end)
|> Enum.uniq()
case points do
[] ->
0
pts ->
{:ok, _} = HrrrPointEnqueuer.enqueue(%{hour_utc => pts})
Logger.info("Pskr.CalibrationSampler: enqueued #{length(pts)} HRRR points for #{hour_utc}")
length(pts)
end
end
# ── Kp lookup ───────────────────────────────────────────────────
# Single Kp value applies cluster-wide for the hour. We prefer the
# integer `kp_index` (the official 3-hour Kp) and fall back to the
# truncated `estimated_kp` when SWPC hasn't published the official
# value yet — same precedence as `Propagation.current_kp_index/0`.
defp fetch_kp(_hour_utc) do
case SpaceWeather.latest_kp() do
%{kp_index: kp} when is_integer(kp) -> kp
%{estimated_kp: kp} when is_number(kp) -> trunc(kp)
_ -> nil
end
end
# ── Per-bucket sample assembly ──────────────────────────────────
defp build_sample({{band, lat, lon} = key, cell_spots}, hour_utc, cell_hrrr, kp) do
hrrr = Map.get(cell_hrrr, key)
now = DateTime.truncate(DateTime.utc_now(), :second)
%{
id: Ecto.UUID.generate(),
hour_utc: hour_utc,
band: band,
midpoint_lat: lat,
midpoint_lon: lon,
spot_count: Enum.sum(Enum.map(cell_spots, & &1.spot_count)),
distinct_paths: length(cell_spots),
median_distance_km: median_distance(cell_spots),
modes: dedup_modes(cell_spots),
surface_temp_c: hrrr && hrrr.surface_temp_c,
surface_dewpoint_c: hrrr && hrrr.surface_dewpoint_c,
pwat_mm: hrrr && hrrr.pwat_mm,
surface_pressure_mb: hrrr && hrrr.surface_pressure_mb,
min_refractivity_gradient: hrrr && hrrr.min_refractivity_gradient,
hpbl_m: hrrr && hrrr.hpbl_m,
hrrr_ducting_detected: hrrr && hrrr.ducting_detected,
kp_index: kp,
inserted_at: now,
updated_at: now
}
end
defp median_distance(cell_spots) do
cell_spots
|> Enum.map(& &1.distance_km)
|> Enum.reject(&is_nil/1)
|> case do
[] ->
nil
list ->
sorted = Enum.sort(list)
n = length(sorted)
mid = div(n, 2)
if rem(n, 2) == 0, do: (Enum.at(sorted, mid - 1) + Enum.at(sorted, mid)) / 2, else: Enum.at(sorted, mid)
end
end
defp dedup_modes(cell_spots) do
cell_spots |> Enum.flat_map(& &1.modes) |> Enum.uniq()
end
# ── Upsert ──────────────────────────────────────────────────────
# Each sample has ~20 fields. PostgreSQL's parameter limit is 65,535,
# so a single insert_all with > ~3,200 rows will fail. Batch at 1,000
# to stay well under the limit even at peak hours.
@upsert_batch_size 1_000
defp upsert([]), do: 0
defp upsert(samples) do
samples
|> Enum.chunk_every(@upsert_batch_size)
|> Enum.reduce(0, fn batch, total ->
{count, _} =
Repo.insert_all(CalibrationSample, batch,
on_conflict:
from(c in CalibrationSample,
update: [
set: [
spot_count: fragment("EXCLUDED.spot_count"),
distinct_paths: fragment("EXCLUDED.distinct_paths"),
median_distance_km: fragment("EXCLUDED.median_distance_km"),
modes: fragment("EXCLUDED.modes"),
surface_temp_c: fragment("EXCLUDED.surface_temp_c"),
surface_dewpoint_c: fragment("EXCLUDED.surface_dewpoint_c"),
pwat_mm: fragment("EXCLUDED.pwat_mm"),
surface_pressure_mb: fragment("EXCLUDED.surface_pressure_mb"),
min_refractivity_gradient: fragment("EXCLUDED.min_refractivity_gradient"),
hpbl_m: fragment("EXCLUDED.hpbl_m"),
hrrr_ducting_detected: fragment("EXCLUDED.hrrr_ducting_detected"),
kp_index: fragment("EXCLUDED.kp_index"),
updated_at: fragment("EXCLUDED.updated_at")
]
]
),
conflict_target: [:hour_utc, :band, :midpoint_lat, :midpoint_lon]
)
total + count
end)
end
defp align_to_hour(%DateTime{} = dt) do
%{dt | minute: 0, second: 0, microsecond: {0, 0}}
end
end