The docker buildx build step has failed twice since Stream C added the
hrrr_point_worker binary. Dropping the target/ BuildKit cache mount
eliminates the most likely cause (disk pressure inside a persistent
cache shared between the two bin builds). Full rebuild from scratch
adds ~3 min to the build but the failure mode is worth more than the
cache.
Also dropped the empty unit test scaffolding in hrrr_points.rs; the
module is thin glue over fetcher+decoder which already carry coverage.
Integration tests for process_batch/upsert_profile are gated on a live
DB + HRRR mirror and live in tests/ (when added).
Two bugs surfacing 13 h after Stream A cutover:
1. prop-grid-rs pods OOMKilled 16–18× overnight on the analysis step.
f00's native-level GRIB2 (~530 MB + wgrib2 working set) runs on top
of 2 concurrent forecast tasks, briefly reaching ~2 Gi. 2 Gi limit
was too tight — 3 Gi plus parallelism 3→2 gives the analysis step
room without eliminating forecast-lane headroom.
2. hrrr-point-rs pod is CrashLoopBackOff because its container is
running image main-1776640915-65f7963 (pre-Stream-C) which doesn't
contain the hrrr_point_worker binary. The grid-rs CI run for
commit 4fefb81 failed and no newer image got published. Touch the
Dockerfile to re-trigger the workflow so flux picks up a new tag
with both binaries. No functional Dockerfile change, just a
docstring update so the path-filter kicks.
Forecast + analysis both claim per FOR UPDATE SKIP LOCKED, so
reducing per-pod parallelism doesn't break the chain — the two
replicas cover 4 slots cluster-wide, which still drains f01..f18 in
~5 min.
Phase 3 Stream C Rust side. Completes the HrrrFetchWorker port.
Pipeline:
- db::claim_next_hrrr_task — FOR UPDATE SKIP LOCKED on hrrr_fetch_tasks,
newest valid_time first. Accepts the points JSONB directly.
- hrrr_points::process_batch — fetch surface + pressure GRIB2 once
per task (tokio::try_join), decode via the existing wgrib2 plumbing,
then for each requested point pull the cell and UPSERT INTO
hrrr_profiles (conflict on lat/lon/valid_time).
- db::complete_hrrr_task / fail_hrrr_task — status transitions; Elixir
backfill re-enqueues failed rows on next /30-min scan.
Shipping pieces:
- new bin src/bin/hrrr_point_worker.rs
- new module src/hrrr_points.rs (process_batch, upsert_profile)
- new Cargo [[bin]] entry; Dockerfile builds both binaries in one stage
and ships them in the runtime image so a single CI pipeline covers
the whole cluster
- k8s/deployment-hrrr-point-rs.yaml (1 replica, 1 Gi limit, anti-affinity
against prop-grid-rs so chain + point work don't fight for wgrib2
slots). Uses the same image; command: override picks the right binary.
- kustomization.yaml: include the new deployment so flux applies it
- deployment-grid-rs.yaml: bump readiness initialDelaySeconds 3→15 +
failureThreshold 3→6 so a slow DB connect during startup can't race
the first probe
119 Rust tests green.
Phase 3 Stream C Elixir-side: HrrrFetchWorker is deleted; per-QSO HRRR
enrichment now writes to the new hrrr_fetch_tasks table for the Rust
hrrr-point-worker to drain.
Table shape:
- one row per valid_time (primary key) with a JSONB array of
{lat, lon} points
- UPSERT on conflict: array-union of points, status flips back to
queued if previously done/failed so a backfill re-scan naturally
refills the queue for Rust
Elixir changes:
- new migration 20260419231502_create_hrrr_fetch_tasks
- new Microwaveprop.Weather.HrrrPointEnqueuer with enqueue/1 and
enqueue_for_contacts/1. Pre-2014 contacts (NARR's territory)
are skipped here so hrrr_status can pin them to :unavailable.
- ContactWeatherEnqueueWorker: build_hrrr_jobs/1 removed; single-
contact path and batch perform/1 both route through
HrrrPointEnqueuer.enqueue_for_contacts/1. A placeholder jobs-list
is kept just to feed mark_hrrr_status!.
- contact_live/show.ex retry button enqueues via the same path.
- :hrrr queue removed from dev/config/runtime.exs
- HrrrFetchWorker module + test deleted
BackfillEnqueueWorker scans continue to flow through
ContactWeatherEnqueueWorker.enqueue_for_contact (unchanged), so the
30-min reconcile refills hrrr_fetch_tasks automatically.
4 new tests cover the routing, pre-2014 skip, UPSERT-union, and
status-reset-on-reschedule behaviour. Rust-side hrrr-point-worker
binary + k8s deployment land in the next commits.
The hourly cron now only seeds grid_tasks. The chain step, native-duct
merge, NEXRAD merge, commercial-link merge, scoring, ProfilesFile
write, and band-score writes all moved to rust/prop_grid_rs.
Elixir changes:
- GridTaskEnqueuer.seed_with_analysis/1: inserts 1 kind='analysis' row
(f00) + 18 kind='forecast' rows (f01..f18).
- PropagationGridWorker: stripped from 423 LOC to a thin seeder.
perform(%{}) → GridTaskEnqueuer.seed_with_analysis.
Deleted: process_forecast_hour, merge_native_duct_data,
merge_nexrad_data, merge_commercial_link_data, compute_scores_*,
persist_profiles, record_run_timing (Rust emits spans to Prometheus
instead), apply_nexrad_observations, apply_duct_grid, timed helpers.
Test rewritten for the new shape: 0 Oban fan-out jobs, 19 grid_tasks
rows with the expected kind distribution.
HrrrNativeClient and NexradClient remain — they have other callers
(HrrrNativeGridWorker for per-QSO duct batch; NexradWorker and
CommonVolumeRadarWorker for per-contact radar). Only f00's direct
use moved.
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.
Phase 3 Stream A: Rust's prop-grid-rs writes f00 profile files as
MessagePack (c4f309c). Elixir needs to read both formats during the
cutover window and indefinitely afterward — .mp.gz wins when both
exist.
Reader:
- path_for/1 stays on .etf.gz (Elixir legacy writer)
- mp_path_for/1 returns the .mp.gz sibling Rust lands
- read/1 prefers .mp.gz, falls back to .etf.gz, :enoent if neither
- decode_mp_body handles the top-level {v, valid_time, cells} shape
- normalize_profile atomizes a whitelist of known keys (surface_*,
native_min_gradient, ducts[], profile[] levels, ...) and leaves
everything else as strings so no String.to_atom bomb is possible
- Directory scan picks up both extensions so retain_window and
prune_older_than also work uniformly; list_valid_times dedupes
by timestamp
Adds msgpax ~> 2.4 dependency. 4 new tests cover format preference,
key atomization, :enoent path, and dedupe.
Rust-side writer for the f00 ProfilesFile. Replaces the Elixir ETF
format (:erlang.term_to_binary) with MessagePack + gzip so Rust can
produce files without porting the Erlang serializer.
Wire layout: top-level map with 'v' (format version), 'valid_time'
(ISO string), and 'cells' (array of {lat, lon, profile}). Profile
values are an rmpv::Value so the writer stays agnostic about what
fields the f00 pipeline decides to include — native-duct metrics,
NEXRAD reflectivity, commercial-link degradation, and the sounding
profile list all serialize through the same path.
All keys are strings — MessagePack has no atom type. Elixir's reader
(follow-up commit) converts known keys to atoms via a whitelist.
Atomic writes via rename(2) mirror ScoresFile/ProfilesFile pattern:
write to tmp sibling, rename into place on the shared NFS mount.
4 unit tests cover round-trip, atomic rename, path layout, and
lat/lon snap.
native_duct.rs is the Rust side of HrrrNativeClient.fetch_native_duct_grid:
fetch native-level HRRR idx → select duct byte ranges (TMP/SPFH/HGT/PRES
× 50 hybrid levels = 200 messages, ~300 MB subset) → wgrib2 decode via
the existing decoder::extract_grid plumbing → per-cell build_native_profile
→ duct::analyze → HashMap<(lat, lon), DuctMetrics>.
build_native_profile keeps the Elixir semantics intact: drop cells that
don't reach 3 levels, sort by ascending height, tolerate missing SPFH/PRES
as zero (the duct math handles zeros gracefully). merge_duct_grid gives
the pipeline.rs caller a drop-in replacement for Elixir's apply_duct_grid.
claim_next_analysis is the analysis-lane companion to claim_next: selects
kind='analysis' rows only, same SKIP LOCKED + newest-first ordering. Kept
on a separate claim function so the forecast lane can keep running
unaffected during cutover; the analysis pipeline wiring lands next.
9 pure unit tests cover message generation, URL builder, profile-sort,
min-level filter, and merge-into-base behaviour. Network-live golden
against a real native GRIB2 deferred to the cutover commit.
Same bug as the propagation map had in 9db9e6a. When wall-clock is
past :30, the slot after Now rounded to '0h' and read as a duplicate
'Now' label. Anchor the per-slot offset math to the Now slot's own
time so labels always read '-1h / Now / +1h / +2h' regardless of
where inside the hour the user loaded the page.
Port of the subset of Microwaveprop.Weather.NexradClient used by
f00's chain step: 5-min-rounded IEM n0q PNG fetch, palette-indexed
decode via the png crate, and per-POI max_reflectivity_dbz extraction
from a 25-pixel half-box (matches Elixir's DEFAULT_BOX_HALF).
The palette↔dBZ math and (lat, lon)↔pixel projection are bit-for-bit
copies of the Elixir formulas. 6 pure-math unit tests cover pixel
mapping, box clipping, and timestamp rounding. Network-live golden
fixture deferred until the full Stream A wiring lands.
Port of Microwaveprop.Commercial.build_link_lookup/2 +
link_degradation_from_lookup/3. The Elixir context stays intact
(LiveView sensor + SNMP poller); this Rust side is the read-only
consumer used by the f00 chain step (Phase 3 Stream A).
Three pure-math unit tests cover: haversine distance (DFW→Austin and
same-point zero), degradation_at aggregation behaviour, and
two-decimal rounding convention that Elixir's Float.round(_, 2)
imposes. SQL queries mirror the Elixir Ecto queries one-for-one —
baseline window, current point-in-time read, and MIN_BASELINE_SAMPLES
quality gate.
Pure port of lib/microwaveprop/propagation/duct.ex — the ITU-R P.834
duct detector that computes the modified refractivity M-profile and
finds regions where dM/dh < 0, then uses Bean & Dutton to derive the
minimum trapped frequency per duct.
Part of Stream A (f00 port): compute_duct_metrics and the per-cell
native-level duct merge both build on analyze(). Landing it now with
standalone unit tests lets later tasks (fetch_native_duct_grid, the
wgrib2 subprocess for native levels) treat it as a known-good
dependency instead of a concurrent port.
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.
Replaces the serial claim→process loop with a JoinSet of N worker tasks,
each running the same claim→fetch→decode→score→complete flow against
grid_tasks. FOR UPDATE SKIP LOCKED already prevents contention between
workers so no additional coordination is needed.
talos5 is 4c/4t (i5-7260U); PROP_GRID_RS_PARALLELISM=4 turns the full
f01..f18 chain from ~18 min serial into roughly ~5 min. Per-task peak
memory after the streamed-bands refactor is ~250 Mi, so 4 concurrent is
~1 Gi — well under the 2 Gi container limit.
Postgres pool sized to parallelism + 2 so NOTIFY and an opportunistic
claim retry share the pool without blocking a worker transaction.
Introduces TaskKind { Forecast, Analysis } and threads it through the
claim flow. Forecast-only filter on the claim query keeps the claim
lane clean while Elixir still owns f00 — a half-deployed cluster won't
deadlock on an unfulfillable analysis row.
Prep for Stream A of Phase 3: once the Rust-side native-duct + NEXRAD +
commercial pipeline lands, a sibling claim_next_analysis function pulls
kind='analysis' rows and dispatches to the new pipeline.
Prep for Stream A of Phase 3 rust migration: f00's native-duct merge,
NEXRAD, and commercial-link enrichment move to Rust via the same
grid_tasks work queue. kind='analysis' (single row per run) vs
kind='forecast' (18 rows per run) lets one Rust worker dispatch on
payload type without a second table.
Forecast default keeps existing Rust claim_next behaviour unchanged
during cutover.
MrmsFetchWorker fired every 2 min on the :propagation queue, decoding an
MRMS PrecipRate GRIB2 into MrmsCache. Its only consumer,
AsosAdjustmentWorker, is disabled in all configs (dev / config / runtime).
Net effect: 30 GRIB2 decodes/hour of dead work on hot pods.
Removed:
- MrmsFetchWorker, MrmsClient, MrmsCache and their tests
- cron entries in config.exs, dev.exs, runtime.exs
- mrms_req_options stub in test.exs
- MrmsCache from supervision tree in application.ex
- stale MRMS references in PropagationGridWorker docstring and test
Also adds docs/plans/2026-04-19-rust-migration-phase3.md which tracks
the broader Stream A (f00 port) and Stream C (HrrrFetchWorker port)
follow-on work.
Mirrors the Oban Job Queue Status table with a new section sourced
from the grid_tasks hand-off table. Columns: Running, Queued,
Retrying, Done (1h).
Active rows render an f## badge per in-flight forecast hour and a
detail line below with run_time / fh / attempt so it's obvious which
HRRR cycle the worker is on.
Pipes through the existing 2s debounced refresh plus a new PubSub
subscription to "propagation:updated" — PropagationNotifyListener
fans out the Rust worker's NOTIFY propagation_ready, so the panel
transitions running → done the moment Rust writes the score file.
Both the hot (`prop`) Deployment and the `prop-backfill` Deployment
labelled their pods with `app: prop` so both joined the same libcluster
Erlang cluster. The `prop` Service selector was `app: prop` too — which
meant the backfill pod (PHX_SERVER=false, no listener on :5000) was in
the Service endpoint pool.
With sessionAffinity: ClientIP, Cloudflared's source IP got hashed to
the backfill endpoint after a backfill pod roll and every request
from that client returned 502 (connection refused on :5000).
Fix: add `tier: hot` to hot-pod template labels; narrow the Service
selector to `{app: prop, tier: hot}`. Backfill keeps `tier: backfill`
and is cleanly excluded. libcluster selector (still `app: prop`) is
unchanged so backfill stays in the Erlang cluster for Oban queue
leader election and Oban.Pro's Smart engine.
Rust `prop-grid-rs` has been validated end-to-end on talos5 on the
streamed-bands image (main-1776635096-6d91461): real chain steps
complete in ~7 s each and the peak-heap refactor landed.
Changes:
- `PropagationGridWorker.seed_chain` now enqueues a single f00 Oban
job instead of f00–f18. The f00 analysis-hour keeps its enrichment
(native-level duct, NEXRAD composite, commercial-link boost,
ProfilesFile write) and the GridCache broadcast.
- `GridTaskEnqueuer.seed/1` is called again so grid_tasks gets the
18 forecast-hour rows the Rust worker claims.
- Test asserts only one Oban job is enqueued and the matching 18
grid_tasks rows land in the DB.
- `deployment-grid-rs.yaml` flips `PROP_SCORES_DIR` from
`/data/scores_shadow` to `/data/scores` so Rust writes to the live
score tree. No file collision — Elixir only writes the f00 files.
The hot pod memory shrink (6 Gi → 1 Gi) is deferred until one full
hourly cycle runs clean under the new split.
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.
1 Gi OOM'd during a real chain step. Each forecast hour loads both a
surface and pressure-level HRRR GRIB2 (~40 MB combined), decodes via
wgrib2 subprocess (which allocates its own buffers), scores 92k grid
points × 23 bands (~2.1M intermediates), and writes 23 score files.
2 Gi absorbs the spike; still ~3× smaller than the Elixir-per-pod
budget we're replacing.
The real worker binary OOM-killed at 512 Mi on startup. The plan's
512 Mi assumption was speculative — actual Rust RSS at boot includes
sqlx pool buffers, reqwest connection pool, tokio runtime, and
tracing JSON subscriber. 1 Gi leaves room for all of that plus the
steady-state ~500 Mi during a chain step.
