Four optimisations targeting the 11.4s chain-step critical path:
1. Drop legacy .mp.gz dual-write (~2.5-2.8s, 24% of step)
The Elixir reader prefers .sgrid; the chunked gzip+msgpack write was
pure overhead on the Rust write path. Removed scalar_future from both
run_chain_step and run_analysis_step.
2. Parallelize wgrib2 decodes (~0.5-1s)
sfc and prs decode ran serially (await sfc, then await prs). Both now
spawn concurrently on the blocking pool.
3. Overlap sfc decode with prs fetch (~1-2s)
sfc (9 messages) resolves before prs (39 messages). Start the sfc
wgrib2 subprocess as soon as the sfc blob lands instead of waiting
for both blobs. The sfc decode hides behind the remaining prs fetch.
4. Cap grib range retries from 5 to 3
2^n exponential backoff meant 1+2+4+8+16 = 31s worst-case per failed
range. Three retries caps at 1+2+4 = 7s. 404s already fail fast.
Also: instrument .sgrid write (observe_stage) so it appears in the
prop_grid_rs_stage_duration_seconds histogram alongside fetch/decode/derive.
Expected mean chain step: ~6-7s (down from 11.4s).
Reworks the post-fetch half of the propagation pipeline. Fetch and GRIB2
decode were already cheap — measured against a live HRRR cycle, all 39
pressure messages decode via `wgrib2 -lola` in 0.29 s and the 31 MB
byte-range fetch takes ~3 s — so nothing here touches the decoder. All the
cost was downstream.
Also fixes a broken NOTIFY that made every chain step run up to 5 times.
pg_notify
`NOTIFY propagation_ready, $1` is a Postgres syntax error: NOTIFY is a
utility statement whose payload must be a literal, so a bind raises
42601. It shared a transaction with the `status='done'` UPDATE, so every
successful step rolled back, stayed 'running', and was requeued by
reclaim_stale_running up to @max_reclaim_attempts times. Elixir's
NotifyListener never fired either, so ScoreCache warm and the
"propagation:updated" fan-out were dead.
FieldGrid
A decoded grid was HashMap<(i32,i32), HashMap<Arc<str>, f32>> — a dense
rectangular grid stored as ~95k nested hash maps, costing ~4.6M inserts
on decode, ~3.7M on merge and ~14M lookups across three derivation
passes. wgrib2 -lola already emits one dense row-major f32 block per
message, so keep it: dense per-message planes, names hashed once per
grid into plane ids, NaN as the missing sentinel. This is what forced
PROP_GRID_RS_PARALLELISM=1 under a 3Gi limit.
Fused pass
Three 95k-cell derivation passes plus 23 band-major scoring passes over
a staged Vec<(f64,f64,Conditions,BandInvariants)> (~19MB re-streamed 23
times) collapse into one pass: levels extracted once per cell, all 23
bands scored while the cell is hot, scores accumulated cell-major so
rayon chunks own disjoint slices. Scores land straight in the dense
score-file body — no ScorePoint scatter.
.pgrid
The profile artifact was an rmpv tree plus gzip -9, written 30x an hour,
and ProfilesFile.read_point/3 gunzipped and unpacked the entire 95k-cell
file to return one cell on every map click and Skew-T load. Replaced
with a dense cell-major f32 record array carrying a self-describing
field table. Elixir reads it via :file.pread; .mp.gz and .etf.gz remain
readable so files written before this drain out of the 48h window.
Measured on a full CONUS grid (95,073 cells x 48 planes x 23 bands):
derive + score + build artifacts 0.022 s
profile write 3.957 s -> 0.006 s (22.0 MB -> 22.4 MB on disk)
single-cell read whole-file decode -> 0.5 us
23 score files 0.003 s
Also
- hrrr_points: batched UNNEST upsert replacing one awaited INSERT per
point. Keeps ON CONFLICT DO UPDATE — the PSKR sampler's two-pass loop
depends on it.
- fetcher: real semaphore capping in-flight ranges at
MAX_PARALLEL_RANGES, which the comment claimed but the code did not do
(it spawned all 27 while the connection pool was sized for 8).
- metrics: per-stage histogram. Only chain-step and decode durations
were instrumented, which is why the write cost stayed invisible.
- profiles_file: parse_valid_time anchors on the known extension set, so
sibling-suffixed names like <iso>.hrdps.prop no longer parse as
<iso>.hrdps and vanish from prune and list operations.
- PROP_GRID_RS_PARALLELISM 1 -> 3. Memory limit held at 3Gi until RSS is
observed at the new parallelism.
- cargo fmt over the crate; worker.rs, hrdps_fetcher.rs and nexrad.rs
were already unformatted at HEAD and the pre-commit hook gates on it.
HRDPS still runs at 0.5 degrees. wgrib2 -lola scales linearly in output
points on rotated lat/lon (12.5 s wall, 202 s CPU for one message at
0.125 degrees) because it has no inverse projection for those grids; a raw
native dump is 0.32 s. The fix is decode-once plus a closed-form
rotated-pole index, left for a follow-up.
Python pskr_mqtt_listen.py:
- Guard against malformed CONNACK/SUBACK/PUBLISH packets
- Fix _s() truthiness: is not None instead of if v (zero SNR was lost)
- EINTR-safe select loop, OOB data handling, VBInt overflow check
- Proper socket cleanup with try/finally + DISCONNECT on shutdown
Elixir:
- Log dropped spot errors in aggregator (was silently discarded)
- Wire retain_scores_window into NotifyListener chain completion
- Recurse sweep_tmp_dir into band/weather_scalars subdirectories
- Rescue recalibrator.run/0 to write failed status row on crash
- Handle nil in fmt_snr/fmt_callsigns (no more nil dB crashes)
- Replace Stream.run with Enum.reduce logging exits in poll_worker
- Handle File.stat race in ms_footprints prune, grid_center nil log
- Fix unused variables, stale comments, missing get_path!/1
Rust:
- Graceful JoinError handling in fetcher/hrdps_fetcher (no more panics)
- round_to_5min returns Option (leap-second safe)
- Acquire/Release ordering on shutdown flags (was Relaxed)
- Parameterized NOTIFY in db.rs, defensive scalar sweep, NaN guard
- OsString::push for tmp naming, clippy fixes
HRRR publishes f21-f48 at 3-hourly intervals in addition to the
hourly f01-f18 we already fetch. This extends the grid_tasks seed
from 19 rows (1 analysis + 18 forecast) to 29 rows (1 analysis +
18 hourly + 10 3-hourly), and bumps the UI forecast window
constant from 18h to 48h so the map timeline and path calculator
forecast chart automatically show the extended horizon.
No Rust logic changes needed — the pipeline is data-driven and
u8 forecast_hour already handles f48.
Rust cell_to_profile_entry now writes surface_refractivity and
min_refractivity_gradient at the cell level so Elixir consumers have a
guaranteed fallback when SoundingParams.derive hits levels without
dewpoint. The SkewT build_profile_assigns falls back to these pre-computed
cell keys when the derived map has nil for those fields.
Production observation 2026-04-29: HRDPS chain steps using
decoder::extract_grid (wgrib2 -lola) ran at 1000m CPU for 10+ minutes
per task without producing any output. The -lola full-grid
interpolation against HRDPS's rotated lat/lon source grid is far
slower than against HRRR's Lambert source — wgrib2 appears to rebuild
the projection table per record.
The probe (lib/mix/tasks/hrdps_probe.ex, deleted in 1da78a80) had
already proved that wgrib2 -lon at five Canadian cities completes in
under a second. Add decoder::extract_points that uses the same -lon
flag for the HRDPS-only point set (~57k cells), batched at 2000 points
per wgrib2 invocation to stay under Linux ARG_MAX.
run_chain_step_hrdps now calls extract_points instead of extract_grid.
The post-extract HashSet filter to hrdps_only_points is gone —
extraction is already restricted to those cells. Per-cell wall time
should drop from "never completes" to ~30-90 s for 57k cells.
extract_grid stays unchanged — it works correctly for HRRR's CONUS
Lambert source.
Tests: parse_lon_output exercises the wgrib2 -s -lon output shape and
proves multi-record / multi-point parsing. parse_lon_segment unit
tests cover the undefined-value sentinel and the snap-back rejection
that ignores cells wgrib2 returned that don't correspond to anything
in the requested batch.
CI runs cargo clippy with -D warnings. The yesterday's HRDPS commits
introduced two pedantic-but-blocked lints:
- grid.rs:97 in_conus_bbox used `>=`/`<=` chain — clippy's
manual_range_contains says use (LAT_MIN..=LAT_MAX).contains(&lat).
- pipeline.rs:426 backfill_dpt_from_depr iterated `cell.iter()` and
threw the values away — clippy's iter_kv_map says use
cell.keys().filter_map(...). Tightened the inline format strings
to `{depr_key}` instead of the deprecated `{}`-with-arg form to
avoid a follow-up useless_format complaint on the same touch.
cargo clippy --all-targets -- -D warnings passes; cargo test
--release stays at 158 passed.
The /weather map now shows HRRR-derived data wherever HRRR has
coverage and falls back to HRDPS-derived data for Canadian cells
outside HRRR's bbox. North of the 49°N HRRR-coverage line the map
stays populated instead of going blank.
Rust:
- weather_scalar_file::dir_for_hrdps + write_atomic_hrdps land HRDPS
scalar chunks at `<vt>.hrdps/` so HRRR's `<vt>/` write isn't
clobbered. The HRRR writer wipes its dir before each write — both
sources need their own dir to coexist.
- pipeline::run_chain_step_hrdps now derives a ScalarRow for each
scored cell (same derive_row that HRRR uses; same wire format) and
writes them via write_atomic_hrdps alongside the .hrdps.prop score
files.
Elixir:
- ScalarFile.dir_for_hrdps + read_bounds + read_point + exists? all
walk both `<vt>/` (HRRR) and `<vt>.hrdps/` (HRDPS). read_bounds
concatenates with HRRR-precedence on overlap (defensive — Rust
pipeline doesn't write HRRR-overlap cells, but cheap to enforce).
- Weather.weather_grid_at unchanged: it routes through ScalarFile,
which now transparently returns the union. GridCache stays
per-valid_time so a single cache entry holds both regions.
Tests cover four scenarios: HRRR-only, HRDPS-only, both-present
union, and HRRR-precedence on the same cell.
Stages 4-5 of the HRDPS plan. Lands the Rust forecast-hour pipeline
for source='hrdps' grid_tasks rows.
What ships:
- `hrdps_fetcher.rs` — sibling of fetcher.rs for ECCC's MSC Datamart
shape: date-prefixed URLs, one variable per GRIB2 file, concurrent
fetch + byte-concat into a single multi-record blob. wgrib2 reads
multi-record files natively so the downstream decoder path is
unchanged. Mirrors lib/microwaveprop/weather/hrdps_client.ex.
- `grid::hrdps_only_points()` + `grid::hrdps_grid_spec()` — Canadian
cells (49-60°N, -141 to -52°W) minus the HRRR overlap. Disjoint
from conus_points by construction.
- `db::TaskSource` enum + `source` column on ClaimedTask. claim_next
surfaces source for forecast lane dispatch; claim_next_analysis
filters to source='hrrr' only (HRDPS analysis path is not yet
ported — would need native-duct + NEXRAD + commercial merges that
Rust doesn't have for HRDPS).
- `pipeline::run_chain_step_hrdps` — fetch combined blob, decode with
HRDPS grid spec, drop HRRR-overlap cells via a HashSet of
hrdps_only_points, back-fill DPT from DEPR (HRDPS publishes DEPR
as primary), score every band, write `<band>/<iso>.hrdps.prop`.
Skips per-valid_time profile + scalar artifacts since those would
clobber HRRR's at the same valid_time — surfacing HRDPS on /weather
is a separate stage.
- worker.rs dispatches forecast lane on (kind, source) — Hrrr →
run_chain_step, Hrdps → run_chain_step_hrdps. Analysis stays
HRRR-only. Worker constructs an HrdpsClient alongside HrrrClient.
- `scores_file::write_atomic_hrdps` writes to `.hrdps.prop` with the
Canadian grid spec via the new encode_with_spec. Coexists with the
HRRR `.prop` file for the same (band, valid_time).
163 unit tests pass; backfill_dpt_from_depr handles surface and
pressure-level levels with no double-fill when DPT exists.
Move ScalarFile from a dual-writer dual-format setup (ETF on the Elixir
side, materialized via NotifyListener; nothing on the Rust side) to a
single chunked-MessagePack format that both languages produce and
consume.
Elixir side:
- ScalarFile now writes gzipped MessagePack `.mp.gz` per chunk via
Msgpax. Reader decodes with Msgpax + zlib. Atom keys are stripped to
strings on encode and re-atomized via a whitelist on read so callers
see the same shape regardless of which side wrote the bytes.
- DateTime values round-trip through ISO8601 strings.
- New cross-language wire-format test that decodes a hand-rolled
payload mirroring exactly what `rmp_serde::to_vec_named` emits, so a
format regression on either side fails the suite.
Rust side:
- New `prop_grid_rs::weather_scalar_file` module:
- 1:1 port of `Microwaveprop.Weather.WeatherLayers.derive/1` plus
the surface fields `Weather.build_grid_cache_row/4` adds on top
(lapse rate, mid lapse rate, inversion strength + base, 850/700 mb
interpolation, surface RH, surface refractivity, ducting flag).
- `write_atomic` chunks rows into 5°×5° buckets, writes one
`<lat_band>_<lon_band>.mp.gz` per chunk via tmp + rename, drops
pre-existing chunks first to keep writes canonical.
- 5 unit tests covering surface-only rows, upper-air derivation,
write/read round-trip, chunk band boundaries, and the
out-of-physical-range surface-temp filter.
- Pipeline integration: both `run_chain_step` (f01..f18) and the f00
analysis path now build `Vec<ScalarRow>` in the same merged-cell
pass that builds `CellEntry` and `Conditions`, then fire a parallel
`spawn_blocking` write that overlaps with scoring. Awaiting the
scalar future after the score-band writes keeps the failure-surface
symmetric with the existing profile_future.
Result: every new forecast hour lands with the scalar artifact already
on disk, so /weather reads never fall through to ProfilesFile decode +
per-cell SoundingParams + WeatherLayers derivation. The Elixir-side
`NotifyListener.handle_propagation_ready/1` materialization remains as
an idempotent safety net for hours that pre-date this commit.
The /skewt page was empty in prod because Phase 2 of the Rust cutover
moved f01..f18 grid scoring to Rust but left profile-file writing only
on the analysis (f00) path. Result: /data/scores/profiles/ accumulated
one file per chain run instead of nineteen, so SkewtLive only ever saw
the analysis hour — and that hour falls outside the 1 h past cutoff
within an hour of the chain finishing, leaving the page empty most of
the time.
Two surgical fixes:
* `pipeline::run_chain_step` now builds `profile_entries` alongside
`prepared` from the merged grid in a single iter pass, then spawns
`profiles_file::write_atomic` on the blocking pool to overlap the
NFS write with the score-band scoring/write fan-out — same pattern
`run_analysis_step` already uses. `ChainStepStats` gains
`profile_cells_written` so the worker log line is symmetric with
the analysis step's existing field and a future failure mode that
drops the profile would show up as a divergence.
* `SkewtLive.available_valid_times/0` widens the past cutoff from
1 h to 3 h. The chain's analysis hour can be ~70 min old by the
time the next chain finishes (and ~2 h on a missed cycle), so the
1 h cutoff was leaving the page blank during normal transients.
18 h forward window unchanged.
Cosmetic: cargo fmt picked up trivial rewrites in decoder/fetcher/
native_duct/hrrr_point_worker that had drifted; rolled in.
Verification: cargo build --release, cargo clippy --lib --bins -D
warnings, cargo test --lib (134/134), mix test
test/microwaveprop_web/live/skewt_live_test.exs (3/3) all green.
Each fix is covered by a regression test that fails on `main` and
passes on this commit.
Round 1 (initial review):
* propagation: thread `latitude` into the conditions map so
`score_season/4` actually picks up regional multipliers
* hrrr_client / fetcher.rs: `nearest_hrrr_hour` rounds DOWN, never at
a future cycle that NOAA hasn't published yet
* radio: spherical-vector great-circle midpoint replaces the
arithmetic mean — anti-meridian paths no longer fold to Greenwich
* weather: `reconcile_weather_statuses` scales the longitude band by
`1 / cos(lat)` so the bbox stays ~150 km wide at every latitude
* radio/maidenhead: clamp 90°/180° below the field-bucket overflow so
`from_latlon` never emits invalid characters like 'S'
* prop_grid_rs/pipeline: merge HRRR + NEXRAD-derived rain rates and
read `best_duct_freq_ghz` into `best_duct_band_ghz` so the Native
Duct Boost actually fires
* propagation/region (Elixir + Rust): inclusive upper bounds so points
exactly at lat_max get the regional multiplier
* weather/sounding_params (Elixir + Rust): drop the 10 m gradient
floor so HRRR's thin near-surface layers stop hiding sharp ducts
* weather/sounding_params: when the profile ends inside a duct,
finalize it with the highest sample as the top instead of throwing
it away (Rust port already correct)
Round 2 (post-fix sweep):
* radio + commercial: single canonical haversine in Radio (atan2
form); Commercial delegates instead of carrying a second copy that
could disagree at threshold distances
* prop_grid_rs/profiles_file: `snap_coords` matches Elixir's
step-aware snap (`round(coord/0.125) * 0.125`, then 3-dp round) so
Rust-keyed and Elixir-keyed profile maps land on the same cell
* weather/grib2/wgrib2: `parse_lon_val_segment` uses `Float.parse`
uniformly — wgrib2 dropping the trailing `.0` from a longitude no
longer crashes the whole chain step
Found by auditing the ProfilesFile, hrrr_profiles, and weather-layer
contracts against what the Rust writers actually emit.
Bug 1 (HIGH) — Rust `hrrr_points` worker upserted hrrr_profiles rows
with NULL `surface_refractivity`, `min_refractivity_gradient`, and
`ducting_detected`. The Elixir HrrrClient path always wrote them
via SoundingParams.derive. Per-QSO pages that rely on these
scalars ("N:", "dN/dh:", duct badge) showed "—" for any contact
enriched through Stream C. Fixed by porting
`surface_refractivity` + `ducting_detected` into Rust's
`sounding_params.rs` and deriving at upsert time.
Bug 2 (HIGH) — Rust f01..f18 `cell_to_profile_entry` never emitted
wind_u / wind_v / cloud_cover_pct / precip_mm into ProfilesFile
cells. `Propagation.factors_for` recomputes factor scores from the
profile cell on /map click, and these four fields are read
(wind_speed_kts, sky_cover_pct, precip_to_rate_mmhr). Result: every
f01..f18 point-detail popup showed 0 for wind/sky/rain factors.
Fixed by extracting the grib values at cell_to_profile_entry time
and whitelisting the keys for atomization in
profiles_file.ex `@mp_atom_keys`.
Bug 4 (MEDIUM) — `WeatherLayers.duct_field/2` only knew the
`SoundingParams.detect_ducts` shape (`d["base"]` / `d["strength"]`).
The Rust/HrrrNativeClient shape uses `:base_m` / `:thickness_m`
atom keys (post-ProfilesFile.read atomization). On /weather, the
Duct Base / Duct Strength layers returned nil for every Rust-origin
cell. Tolerant lookup now reads either shape; uses `thickness_m`
as the "strength" proxy for the native shape (thicker duct traps
a wider band).
Skipped (verified but lower impact):
- Rust ProfilesFile aggregates ducts as scalars, not per-layer — the
`ducts` array in the popup is always empty. Requires larger
DuctMetrics refactor in Rust; deferred.
- `complete_hrrr_task` emits no NOTIFY (vs `complete` which does).
Elixir has no matching listener yet, so adding NOTIFY alone is a
no-op; the existing cron reconciler catches it.
- `:best_duct_band_ghz` fallback in propagation.ex is dead code;
cleanup, not a bug.
133 Rust tests + 2842 Elixir tests + credo green.
- Add one-line moduledocs to 18 LiveViews that were carrying
@moduledoc false. Each line names the route and summarizes what
the page does so a future reader knows where to look before
opening the file.
- Add MapLive handle_event tests for toggle_radar, select_time /
set_selected_time, point_detail, map_bounds, retry_initial_scores.
Assertions focus on socket-state transitions and "didn't crash"
rather than raw HTML — map_bounds in particular has no visible
render-side effect (it push_events the new scores to the JS hook).
- Rust pipeline: integration-shaped test that hand-builds a 2-cell
surface + pressure grid, runs merge → cell_to_conditions →
precompute_band_invariants → composite_score_with → write_atomic
→ decode, asserting header + body length round-trip. Closes the
'pipeline's happy path is only covered by unit tests' gap.
Batched from a system-wide review pass.
**Rate-limit + transient-error log hygiene**
- WeatherFetchWorker: classify IEM HTTP 429 as {:snooze, 300} instead
of {:error}. Stops Oban.PerformError stack-trace spam on every
routine rate-limit during backfill, keeps the retry semantics.
- IonosphereFetchWorker: compress GIRO TLS :unknown_ca error from a
~400-char inspect blob to 'TLS unknown_ca (CA bundle missing)'.
- FreshnessMonitor: log only when an enqueue actually lands (the
Oban unique conflict case was silently dropping jobs but still
producing 'enqueuing grid worker' info lines every 5 minutes).
**Perf**
- Rust grid_level_keys(): pre-format 'TMP:{p} mb' / DPT / HGT keys
once via OnceLock instead of format!()'ing per-cell. Removes ~3.6M
String allocations per f01..f18 chain step across pipeline.rs,
hrrr_points.rs. Same for the wgrib2 :(TMP|DPT|HGT): pattern in
hrrr_points.process_batch (sfc_pattern / prs_pattern OnceLock).
- MapLive preload_forecast: Task.async_stream with max_concurrency=4
replaces the 18-wide Enum.map serial walk. Forecast cache warms
~4× faster after a band change, with ordering preserved.
- grid_tasks: partial composite index on (run_time DESC,
forecast_hour ASC) WHERE status='queued' AND kind='forecast',
matching the Rust claim query's ORDER BY. Drops the old
status-only partial that forced a sort per claim.
**Correctness**
- ContactImportWorker: add Oban unique:[keys: [:import_run_id,
:offset]]. Was missing on a worker whose perform() does a
non-idempotent atomic counter increment — a retry would double-
count imported rows.
- CommonVolumeRadarWorker: x_min..x_max default step is -1 when
x_min > x_max, triggering a runtime deprecation warning. Force
Range.new(_, _, 1) explicitly.
**Cleanup**
- Drop unused tmp_dir parameter from hrrr_point_worker + its
process_batch signature.
Score-grid files now land at `<band>/<iso>.prop` with a 4-byte "PROP"
magic header. Readers still accept the legacy `.ntms` extension and
"NTMS" magic so a rolling deploy doesn't invalidate any file already
on the shared NFS mount — legacy files age out through normal
retention.
Applied both sides of the seam: Elixir ScoresFile + regex parsers,
Rust scores_file writer + decoder. Updated the comments in all
modules that mention the extension (NotifyListener, Reconciler,
gefs_fetch_worker, map_live) and the runbook diagram. talos5 is a
DB host now, not a Rust worker — corrected the runbook accordingly.
New telemetry module wires tracing-subscriber to both a local JSON
fmt layer (keeps kubectl-logs output identical) and an
OpenTelemetry OTLP/gRPC exporter, activated only when
OTEL_EXPORTER_OTLP_ENDPOINT is set. The returned TelemetryGuard
holds the SdkTracerProvider until process shutdown so queued spans
flush before exit.
Both bin targets (worker, hrrr_point_worker) now call
telemetry::init(service_name) at startup; service_name becomes the
OTel service.name attribute so Tempo groups spans per binary.
Tracing instrumentation on the three main work units:
- pipeline::run_chain_step (forecast f01..f18)
- pipeline::run_analysis_step (analysis f00)
- hrrr_points::process_batch (per-QSO point drain)
k8s manifests set OTEL_EXPORTER_OTLP_ENDPOINT to the cluster
collector at otel-collector.observability.svc.cluster.local:4317.
Backend wiring lives in the vntx-infra repo.
run_analysis_step chains the Rust-side f00 work end-to-end:
- fetch surface + pressure GRIB2 + native duct in parallel
(tokio::try_join!) with per-leg error-lift to PipelineError
- merge pressure into surface, fold native duct metrics per cell
- NEXRAD composite reflectivity overlay via the IEM n0q PNG path
- commercial-link degradation via the Postgres per-link baseline
- score all 23 bands in parallel (rayon) then write band files
via parallel spawn_blocking (try_join_all)
- write the ProfilesFile as MessagePack alongside the score files
worker.rs main loop now claims kind='analysis' first, falls back to
kind='forecast'. Analysis tasks dispatch to run_analysis_step;
forecast tasks still go through run_chain_step unchanged.
ChainOutcome enum tags log events (kind=analysis|forecast) and
carries analysis-specific counters (profile_cells_written, duct_cells,
nexrad_cells_with_echo, commercial_cells_boosted) so Phase 3 cutover
observability is readable from kubectl logs alone. 118 Rust tests
green. Elixir seeder flip + f00 code deletion lands next.
Three independent improvements in one commit:
1. Parallel band scoring (pipeline.rs):
- rayon par_iter across 23 bands replaces the serial for loop.
- All band files now write via try_join_all over spawn_blocking
instead of serializing one-at-a-time on NFS. Chain per-step
drops from ~60s to ~15-25s on a 4-core box.
2. Prometheus metrics (new metrics.rs):
- axum server on METRICS_ADDR (default :9100) exposes /metrics and
/health. Scraped by existing Prometheus via annotation.
- Histograms: chain step duration (by outcome), decode duration.
- Gauge: tasks in flight (RAII guarded so panics don't leak).
- Counter: chain steps by outcome.
- worker.rs records step duration and wraps each run in an
InFlightGuard.
3. HA + ordering fixes:
- deployment-grid-rs.yaml: replicas 1→2, required podAntiAffinity
spreads them across hosts, nodeAffinity prefers talos5 for one.
PROP_GRID_RS_PARALLELISM 4→3 per pod (6 concurrent across
replicas; smaller secondary-node footprint).
- Readiness probe on /health so rollouts wait for the runtime to
be alive.
- claim_next ordering: run_time ASC → DESC so the newest hourly
run drains before any stragglers from a broken prior run. Within
a run, forecast_hour ASC keeps nearest-first.
- grid_task_enqueuer sets kind="forecast" explicitly and updates
conflict_target to the new 3-column unique index introduced by
migration 20260419222624.
Previous pipeline held three overlapping large structures in RAM:
1. `merged: HashMap<Key, HashMap<String, f32>>` — 92k outer × ~60 String
keys inner = ~200 MB of fragmented heap
2. `per_band: HashMap<u32, Vec<ScorePoint>>` — all 23 bands × 92k
ScorePoints before the first write
3. A `scores.clone()` per band during the write loop (doubled that
band's footprint transiently)
Consolidate into:
- Consume `merged` into a compact `Vec<(lat, lon, Conditions,
BandInvariants)>`; the huge String-keyed HashMaps drop as soon as
collection finishes.
- Loop over bands: score → write → vector drops at end of iteration.
One band's ~1.8 MB score vector lives at a time.
Peak heap goes from ~1.5 Gi to ~250 Mi on the chain step. 2 Gi limit
stays, but we now have 8× headroom instead of tight-fit.
Extracts the memory-hostile HRRR fetch → decode → score pipeline for
forecast hours 1–18 into a separate Rust service (`prop-grid-rs`).
Elixir retains f00 with its native-duct + NEXRAD + commercial-link
enrichment; Rust handles the other 18 steps per hourly chain.
Hand-off via a new `grid_tasks` table (`FOR UPDATE SKIP LOCKED` claim
from Rust, Postgrex `NOTIFY propagation_ready` back to Elixir). Rust
writes the same on-disk score-grid format Elixir already uses, to the
same `/data/scores` NFS tree. Phase 1 ships in shadow mode with
PROP_SCORES_DIR=/data/scores_shadow.
Rust crate layout at `rust/prop_grid_rs/` (1:1 module parity with the
Elixir source it ports):
- grid, region, band_config, scores_file, sounding_params
- scorer: all 10 factors + composite. Matches the Elixir scorer
byte-for-byte across 115 golden-fixture samples (5 scenarios
× 23 bands).
- decoder: wgrib2 subprocess + Fortran-record lola-binary parser
- fetcher: HRRR URL derivation + idx cache (1h TTL) + 8-way
parallel byte-range downloads with 429/5xx retry
- pipeline: end-to-end chain step, f00 rejected at the boundary
- db: sqlx grid_tasks claim/complete + propagation_ready NOTIFY
- bin/worker: tokio main loop, JSON logs, SIGTERM-safe
91 Rust tests + clippy -D warnings clean. 2,158 Elixir tests green.
Elixir additions:
- `GridTaskEnqueuer` seeds fh=1..18 rows from
`PropagationGridWorker.seed_chain/0`
- `NotifyListener` LISTEN → warm `ScoreCache` → PubSub fan-out
- `ShadowComparator` diffs prod vs shadow `.ntms` bodies
- `mix rust.golden` task writes the Rust-side golden fixture
k8s: new `deployment-grid-rs.yaml` pinned to talos5 (32 GB NUC) via
`prop-grid-rs=primary` nodeSelector + `workload=grid-rs:NoSchedule`
toleration, 512 Mi limit, sharing the existing NFS `/data` mount.
The plan document is at `plans/vivid-hatching-quail.md` (local to my
workstation); phases 2 (cutover) and 3 (talos5 concurrency tuning)
follow after 72h of shadow-mode parity.
