prop/lib/microwaveprop/workers/propagation_grid_worker.ex
2026-06-16 12:38:08 -05:00

96 lines
3.3 KiB
Elixir

defmodule Microwaveprop.Workers.PropagationGridWorker do
@moduledoc """
Hourly seed worker for the Rust `prop-grid-rs` chain.
Post-Phase-3-cutover, Elixir no longer runs any fetch/decode/score
work for the propagation grid. The hourly cron fires this worker
with empty args, and it inserts grid_tasks rows (1 analysis
f00 + forecast f01..f18 hourly, f21..f48 3-hourly) for Rust to drain. Rust owns everything
from there: HRRR fetch, wgrib2 decode, native-level duct merge,
NEXRAD composite, commercial-link degradation, band scoring, and
both the ProfilesFile (MessagePack) and the per-band score files.
The old chain-step perform/2 clauses and the merge/score helpers
moved to `rust/prop_grid_rs/src/pipeline.rs`. `Propagation.record_run_timing`
is still called from this module when the Rust worker reports a
chain step, through the PropagationNotifyListener path.
"""
use Oban.Pro.Worker,
queue: :propagation,
priority: 0,
max_attempts: 5,
# Deduplicate seed jobs in a 1-hour window. The hourly cron fires
# `%{}` args; unique protects against FreshnessMonitor re-fires
# during an outage stacking multiple chains for the same run.
unique: [period: 3600, states: :incomplete]
alias Microwaveprop.Instrument
alias Microwaveprop.Propagation.GridTaskEnqueuer
alias Microwaveprop.Weather.HrrrClient
require Logger
@impl Oban.Pro.Worker
def process(%Oban.Job{args: args}) when args == %{} do
Instrument.span([:propagation, :grid_worker, :perform], %{}, fn ->
seed_chain()
end)
end
defp seed_chain do
# Reclaim any `running` grid_tasks rows orphaned by a Rust worker
# that died mid-claim (SIGKILL, OOM, node drain). Without this
# the status panel shows a permanent-looking spinner and the
# rows are never re-worked.
_ = GridTaskEnqueuer.reclaim_stale_running()
run_time = pick_run_time(DateTime.utc_now())
Logger.info(
"PropagationGrid: seeding chain run_time=#{run_time} (f00 analysis + f01-f18 hourly, f21-f48 3-hourly forecasts via grid_tasks)"
)
case GridTaskEnqueuer.seed_with_analysis(run_time) do
{:ok, count} ->
:telemetry.execute(
[:microwaveprop, :propagation, :grid_worker, :seeded],
%{queued_tasks: count},
%{run_time: run_time}
)
:ok
{:error, reason} ->
:telemetry.execute(
[:microwaveprop, :propagation, :grid_worker, :exception],
%{},
%{run_time: run_time, reason: inspect(reason)}
)
{:error, reason}
end
end
@doc """
Pick the freshest HRRR cycle that's published.
Probes `now-1h` first via `HrrrClient.cycle_available?/1`. If the
pressure-level f18 idx file is on the mirror/S3, the cycle is fully
uploaded and we use it. Otherwise we fall back to the conservative
`now-2h` slot — NOAA's stated max publishing latency for HRRR.
Public so the test suite can exercise the selection logic without
going through Oban.
"""
@spec pick_run_time(DateTime.t()) :: DateTime.t()
def pick_run_time(%DateTime{} = now) do
fresh = now |> DateTime.add(-1, :hour) |> HrrrClient.nearest_hrrr_hour() |> DateTime.truncate(:second)
if HrrrClient.cycle_available?(fresh) do
fresh
else
now |> DateTime.add(-2, :hour) |> HrrrClient.nearest_hrrr_hour() |> DateTime.truncate(:second)
end
end
end