The hourly :35 fire was reliably being missed in production: the prop
deployment is pinned to a single replica, and frequent ArgoCD deploys
restart the pod often enough that some fires land in the swap window.
Between 13:35Z and 17:25Z today the cron fired exactly zero times.
Switch to :05/:25/:45. Each fire calls seed_current_hour which is
idempotent on the (run_time, forecast_hour, kind, source) unique
index, so duplicate inserts in the same hour are no-ops at the DB
level and don't add cost on the happy path. Three shots per hour
gives two retry windows when a deploy kills the pod mid-cron-tick.
The 1-hour valid_time filter inherited from /weather is wrong for
HRDPS: the model publishes 4x/day with 3-4h publication latency and
the chain step adds ~13 min on top, so the freshest data on disk is
routinely 4-6h old. With the 1h cutoff in place, /weather-ca's
timeline was empty whenever there wasn't an active chain step running,
which left the JS hook with nothing to fetch. The sidebar already
shows the selected valid_time prominently — users can see for
themselves how stale the data is and don't need the page to hide it.
Recalibrator and FrontalAnalysis use Nx for tensor math, but Nx is
declared `only: [:dev, :test]` in mix.exs because neither module has a
prod entry point — Recalibrator is invoked by hand from IEx and
FrontalAnalysis is research scaffolding. The prod compile would emit
"Nx.to_number/1 is undefined" warnings on every Nx call. Add a
`@compile {:no_warn_undefined, Nx}` directive to both modules so the
prod build is clean without dragging Nx/Axon/EXLA (~hundreds of MB) into
the runtime image.
New LiveView at /weather-ca duplicates the /weather UI but defaults the
viewport to central Canada (55N, -100W, z=4) and renders the map div
with data-source="hrdps". The JS hook reads that attribute and appends
&source=hrdps to every cell fetch; the WeatherTileController routes
those reads through Weather.weather_grid_hrdps_at/2, which only reads
the <vt>.hrdps scalar dir and skips HRRR merging entirely. Timeline is
sourced from ScalarFile.list_valid_times_hrdps/0 so HRDPS-only hours
show up even when no HRRR file exists for the same valid_time.
Also drops "— Unified" from the algo.md H1.
The HRDPS pipeline as written was unusable in production: each chain
step took ~5 hours (not 30-90s as the dev-bench claim) because every
`wgrib2 -lon` batch redundantly JPEG2000-decodes ~150 records. With 57
batches × 18 forecast hours, a single cycle would take days to drain.
The /weather Canadian view only needs the current hour, so the cheapest
fix is to stop seeding the rest of the chain entirely:
* Rust grid: HRDPS_STEP=0.5 (was 0.125) drops point count 16× to ~3.5k.
Coarser cells than HRRR, but /weather gets visible Canadian coverage
inside one chain step (~20 min) instead of never.
* GridTaskEnqueuer.seed_current_hour/2 inserts exactly one forecast row
whose valid_time is closest to `now`. fh is clamped to 1..18 so the
Rust F00Reserved branch never fires.
* HrdpsGridWorker switches to seed_current_hour and drops the
6-hour-cycle cron in favor of HH:35 hourly fires. Reseeds are
idempotent on the 4-column unique key.
Also fixes a pre-existing credo "nested too deep" in
ScalarFile.read_point by extracting lookup_in_chunk + hit_or_false
helpers — happened to be on the read path that surfaces this work, so
folding it into the same commit.
The prior Docker build re-compiled all 546 transitive Rust deps from
scratch on every push because target/ is intentionally not cache-mounted
(Stream C disk-pressure failures). Two changes drop the cold-build cost
substantially:
* cargo-chef splits dep compilation into a separate Docker layer keyed
on recipe.json (derived from Cargo.toml/Cargo.lock). Source-only
pushes — the common case — now reuse the cooked deps as a normal
layer instead of re-running cargo for every crate.
* clippy now runs in --release profile so it shares artifacts with
`cargo test --release` and `cargo build --release`. Previously clippy
ran in dev profile, forcing a full second compile of every dep before
the release build could start.
Verified `cargo clippy --release --all-targets -- -D warnings` is clean
locally — no new lints surface under the release profile.
wgrib2 3.8.0's CMakeLists.txt:93 calls find_package(g2c 2.3.0), so the
previous 2.1.0 pin makes both Dockerfiles error out with "Could not find
a configuration file for package g2c that is compatible with requested
version 2.3.0". Bumping to 2.3.0 across both the Elixir and Rust images
since their wgrib2-builder stages stay in lockstep.
Latest-version sweep across the Oban stack. Final state:
oban 2.22.1 (hexpm)
oban_pro 1.7.0 (vendored, latest in licensed repo)
oban_met 1.1.0 (vendored, latest in licensed repo and hexpm)
oban_web 2.12.3 (vendored from public hexpm)
oban_web was re-vendored via `mix hex.package fetch oban_web 2.12.3
--unpack` and dropped into vendor/oban_web. No local mods to
preserve — vendor commit history shows the dir was a vanilla
import.
Aligning vendoring style with the sibling towerops-web repo. They
have the same vendor/oban_pro + oban_met + oban_web setup but no
`override: true` on oban_web — because they don't have live_table.
We do (live_table requires `oban_web ~> 2.11` from hexpm), so the
override is load-bearing here. Comment captures the reason so a
future deps cleanup doesn't assume it's vestigial and remove it.
Previous commit (3c988a5f) switched oban_pro to the licensed hex
repo at deploy time. Reverting that — keep all three Pro packages
vendored so prod image builds don't depend on oban.pro reachability
or auth on every CI run.
oban_pro 1.7.0 dropped into vendor/oban_pro via
`mix hex.package fetch oban_pro 1.7.0 --repo=oban --unpack`. mix.exs
goes back to `path: "vendor/oban_pro"` (matching oban_met / oban_web,
which were already vendored — and stay vendored since the licensed
repo only has older versions of those: 0.1.11 / 2.10.6 vs the
1.1.0 / 2.12.1 we vendor).
Schema migration from 3c988a5f stays — 1.7.0 tables/indexes are
already applied. No code changes.
Oban Pro 1.7.0 ships database-backed workflow tracking, replaces the
generated `uniq_key` / `partition_key` columns with expression
indexes, and adds new partial indexes for query performance.
Switches oban_pro from the vendor/ tree to the licensed `oban` hex
repo (auth already configured via `mix hex.repo`). vendor/oban_pro
deleted — the vendoring was a 1.6.x stopgap, 1.7 is the first
version we land directly from hex.
oban_met (1.1.0) and oban_web (2.12.1) stay vendored — they're
not published to the licensed `oban` repo (verified by 404 on
`mix hex.package fetch`).
Migration 20260430142341 runs `Oban.Pro.Migration.up(version: "1.7.0")`
which creates the workflow tables and the new indexes. Verified on
dev + test DBs. Production is well below the size threshold the v1.7
guide flags for split-migration handling, so the inline default
suffices.
No code changes for the breaking-change list — the codebase only
references Oban.Pro.Engines.Smart and Oban.Pro.Plugins.DynamicLifeline,
both unchanged in 1.7. We don't use Workflow.after_cancelled/2 or
Oban.Pro.Workers.Chunk.
Test suite stays at 3132/3132.
Adds lat/lon/zoom URL params to /weather so reload / share-link
restores the exact map state. The existing ?layer= param keeps
working alongside.
LiveView side:
- mount() reads lat/lon/zoom from query params, clamps to valid
ranges (lat -90..90, lon -180..180, zoom 4..10), and falls back
to DFW@z7 when missing or invalid. Initial values flow to the
hook via three new data attributes.
- handle_event("viewport_changed", ...) accepts {lat, lon, zoom}
pushed by the hook on debounced moveend, clamps + rounds, and
push_patches the URL with replace: true so browser history
doesn't grow an entry per pan tick.
- build_path/2 helper centralises URL construction so select_layer
and viewport_changed both produce a deterministic param order
(layer, lat, lon, zoom). Tests assert exact URLs.
JS hook side:
- mounted() reads data-initial-lat/lon/zoom from the el dataset
with parseFloat/parseInt + Number.isFinite gates, then seeds
Leaflet's center/zoom from those instead of the previous
hardcoded (32.897, -97.038, 7).
- moveend handler now also schedules a 400 ms-debounced
pushEvent("viewport_changed", {lat, lon, zoom}) so the URL
catches up with whatever the user panned/zoomed to. Separate
debounce timer from the cells fetch since the URL push and the
network fetch have different cadence sweet spots.
Tests cover: initial seeding from URL params, clamping of bad
values, viewport_changed event producing the right patched URL,
and the original layer-shareability test (now expecting the full
viewport in the URL).
Production wgrib2 3.8.0 still hit `*** FATAL ERROR: packing type 40
not supported ***` on HRDPS files. ldd confirmed wgrib2 was linking
only libc/libm — no OpenJPEG, no Jasper, no g2c.
wgrib2 3.8.0 cmake doesn't have USE_OPENJPEG or USE_JASPER options.
The legacy flag names we passed were silently ignored. The actual
knob is USE_G2CLIB_LOW, which links wgrib2 to the external g2c lib
(already built in the previous stage with OpenJPEG support). Dropped
the bogus USE_PNG / USE_OPENJPEG flags and added USE_G2CLIB_LOW=ON
in both Dockerfiles.
Verified by reading wgrib2 v3.8.0's CMakeLists.txt: the documented
JPEG2000 path is "USE_G2CLIB_LOW: Use g2c low-level decoders
(png,jpeg2000)?".
Reproduced in production pod 2026-04-30: wgrib2 3.6.0 crashes with
`free(): invalid size` when -lon point extraction is run against any
HRDPS rotated lat/lon GRIB2 file, regardless of batch size (failed at
N=200, N=1000, N=2000). Output before the abort is denormal f32 garbage
(val=4.93944e-41) — wgrib2's internal record buffer is being corrupted.
Same files extract cleanly with wgrib2 3.8.0 locally. Bumping the build
arg in both Dockerfile (Elixir image, used by hrrr-point-rs/per-QSO
HRDPS lookups) and rust/prop_grid_rs/Dockerfile (Rust grid worker) so
both consumers get the fix.
Also bumped POINT_BATCH 200 → 1000 in decoder.rs. With wgrib2 3.8.0
stable at 1000+ point batches, this drops the chain-step invocation
count from 286 → 57 per step. Comment captures the reproduction
context so a future bisection finds it quickly.
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.
HPA: prop-grid-rs and hrrr-point-rs were sitting at 4 pods for
~16 min after the hourly chain spike — 10-min scaleDown stabilization
window plus a 1-pod-per-2-min drain policy. Both deployments stayed
visibly over-provisioned even when CPU dropped to 1m/pod.
Tighten scale-down to: 2-min stabilization (long enough that a real
multi-cycle surge doesn't immediately collapse) + Percent: 100 per
60s drain (back to min=1 within 1 minute of stabilization clearing).
Net: spike → scale up to handle queue (unchanged) → back to 1 pod
within ~3 min of spike ending. scaleUp left untouched.
CLAUDE.md updates so this doesn't bite again:
- Document `cargo clippy --all-targets -- -D warnings` as part of
the local Rust precommit ritual. The HRDPS commits in this branch
passed `cargo build` locally but failed CI on lints (manual_range_
contains, iter_kv_map). Adding the recipe under Commands plus a
Rust subsection in Key Conventions so future Rust touches don't
ship clippy regressions.
- Added a Testing note about not anchoring relative-time fixtures on
utc_now-truncated-to-hour — the weather_map_live one-hour-cutoff
test (fix in 2768fc68) was a textbook example: 50% flake rate
driven entirely by which minute of the hour the test ran in.
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 6, 7, 10 of the HRDPS plan, plus the Elixir read-side merge
that makes Rust's `.hrdps.prop` writes show up on /map.
Read side:
- `ScoresFile.path_for_hrdps/2` + `read_hrdps/2` — companions to the
HRRR equivalents. Same decode path; different filename.
- `ScoresFile.read_bounds/3` now reads HRRR + HRDPS files and
concatenates the cells. Files are disjoint by construction
(Grid.hrdps_only_points excludes CONUS) so no de-dup needed.
- `ScoresFile.read_point/4` checks `.prop`, then `.hrdps.prop`, then
legacy `.ntms` — first hit wins. Cells outside their owning region
read as no-data in the other file's grid, so the order doesn't
matter for correctness.
- `Propagation.warm_cache_and_broadcast/2` reads both files and
merges before broadcasting to the cluster ScoreCache. A single
missing file (HRDPS pre-cycle, HRRR briefly absent) is now OK —
the cache warms with whichever side is available.
Activation:
- runtime.exs cron: HRDPS seed worker fires at HH:35 of each cycle
hour (05:35Z, 11:35Z, 17:35Z, 23:35Z) — 5h after cycle, 30 min
staggered from HRRR's */15 schedule.
- HrdpsGridWorker moduledoc updated: no longer dormant; Rust
prop-grid-rs handles source='hrdps' rows.
- map_live North America bbox extended to (24.5, 60.0, -141.0, -52.0)
for the out-of-coverage banner check; visitors anywhere in CONUS
or the Canadian extent now stay banner-free.
Cleanup:
- mix hrdps.probe deleted — its risks are retired and the production
HrdpsClient covers the same surface.
What's still in motion (not blocking the Canadian /map):
- Per-valid_time profile + scalar artifacts for HRDPS (would let
/weather show Canadian temperature/refractivity layers). Rust
pipeline currently skips those so the HRRR-written companion files
aren't clobbered — separate stage when /weather UX is decided.
- FreshnessMonitor HRDPS staleness tracking — HRDPS has its own cron
and a different cadence than HRRR; HRRR-stale logic doesn't apply
cleanly. Defer until prod tells us what alarms we actually want.
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.
CLAUDE.md adds HRDPS to Key Data Sources and the Background Job
Queues table with explicit "dormant until Rust ships" framing.
The 2026-04-29 plan gets a Status snapshot at the top: which stages
shipped (0-3 Elixir + 8) vs which are deferred to the Rust port
sessions (4-7, 9). Activation path is one-paragraph: land Rust HRDPS,
add cron entry, widen map bounds, extend FreshnessMonitor.
Stage 3 (Elixir foundation) of the HRDPS plan. Lands the scaffolding
that lets the Rust prop-grid-rs HRDPS branch be wired up cleanly when
it ships, without further DB or Elixir-side changes.
What ships:
- `grid_tasks.source` column (default 'hrrr', backfill is a no-op).
Replaces the (run_time, forecast_hour, kind) unique index with a
4-column key (..., source) so HRRR and HRDPS analysis/forecast rows
for the same cycle coexist.
- `Grid.hrdps_only_points/0` — Canadian cells inside HRDPS bbox
(49-60°N, -141 to -52°W) but outside HRRR's CONUS bbox. Disjoint
from `conus_points/0` by construction so the two grids never
double-write the same (lat, lon).
- `Grid.point_source/1` — `:hrrr | :hrdps | :neither` classification
for a (lat, lon) pair. Routes per-QSO enrichment to the right NWP
source.
- `GridTaskEnqueuer.seed_with_analysis/2` and `seed/2` accept a
`:source` option; defaults to "hrrr" so existing callers are
unchanged.
- `HrdpsGridWorker` cycle seeder mirroring PropagationGridWorker's
shape but for HRDPS's 4×/day cadence. `:hrdps` queue added to base
Oban config.
What's deferred:
- The Rust `prop-grid-rs` worker doesn't yet branch on
`task.source`. Until that ships, `HrdpsGridWorker` stays out of
runtime.exs cron — it would just produce rows that the Rust worker
mishandles as HRRR. Module doc explicitly notes the dormant state.
- Map overlay extension and FreshnessMonitor HRDPS staleness deferred
to follow-up sessions once real data flows.
Activation path documented in the updated plan file.
Coverage cap recorded as Stage 8 decision: 60°N for v1 (SRTM stops
there). Arctic CDEM deferred to a follow-up plan.
Stage 2 of the HRDPS plan. Mirror of hrrr_profiles for HRDPS-derived
rows so HRDPS can be dropped/reprocessed without disturbing the 4500+
HRRR profiles already in prod. Same column set, same indexes,
PARTITION BY RANGE (valid_time) with quarterly partitions starting at
the deploy quarter.
Schema follows the @primary_key {:id, :binary_id, autogenerate: true}
convention and the same set of derived fields HrrrProfile carries
(surface_refractivity, min_refractivity_gradient, ducting_detected,
duct_characteristics, etc.) so the rest of the pipeline can consume
HRDPS-derived rows interchangeably with HRRR-derived rows.
Unique constraint on (lat, lon, valid_time) is enforced at the DB
level via the partitioned index; schema-level unique_constraint/2
matching against changeset errors doesn't apply because the
partition's index name varies per partition. Production upserts go
through `on_conflict: {:replace_all_except, ...}` per project
convention.
The test computed recent_t = (truncated_to_top_of_hour now) - 30min,
which is 30-90 min before real utc_now() depending on what minute the
test runs at. The LiveView's `cutoff = utc_now() - 1h` then dropped
recent_t whenever the real-time minute exceeded ~30, so the test
failed half the time.
Anchor recent_t at the top of the current hour — that's always within
the 1-hour window regardless of minute. The test still exercises the
"keep within last hour" path; just stops being clock-flaky.
Stage 1 of the HRDPS plan. Mirrors HrrrClient.fetch_grid/3's signature
so PropagationGridWorker can call either polymorphically.
Architectural shape per the probe wedge findings:
- ECCC's date-prefixed datamart URL structure
(dd.weather.gc.ca/{YYYYMMDD}/WXO-DD/model_hrdps/...)
- One-variable-per-file model — fetch ~14 small GRIB2s concurrently,
byte-concatenate into a single multi-record binary, hand to
Wgrib2.extract_points_from_file. wgrib2 reads multi-record files
natively so no -merge step is needed.
- wgrib2's -lon flag handles the rotated-lat/lon reprojection
internally — no manual rotation math.
Variable catalog covers the surface set the propagation scorer reads
plus the 1000-700 mb pressure-level set used by the refractivity
gradient. PWAT is unavailable in HRDPS f000 inventory; the scorer's
PWAT factor will fall back to its default for HRDPS-derived points.
DEPR (T-Td depression) is treated as an alternate primary; when DPT
is absent build_profile_from_extracted derives dewpoint as
temp - depression so downstream consumers see the same shape as HRRR.
Plan task 1.6 (CYYR projection sanity check vs UWYO sounding) is
covered by the live probe at lib/mix/tasks/hrdps_probe.ex which
already validated wgrib2's rotated-grid handling against five
Canadian cities; the production implementation routes through the
same Wgrib2.extract_points_from_file path that the probe used.
Adds mix hrdps.probe — throwaway wedge that retires URL/projection/decode
risks against MSC Datamart. Verified against five Canadian cities with
plausible 2m temperatures via wgrib2's native rotated-lat/lon support.
Findings update what we knew:
- ECCC reorganized to date-prefixed paths; the old /model_hrdps/ root
404s now. Real URL is dd.weather.gc.ca/{YYYYMMDD}/WXO-DD/model_hrdps/...
- HRDPS publishes one variable per GRIB2 file (~3.5 MB each), not
HRRR's bundled wrfsfcf format. ~14 files per cycle/forecast hour.
- wgrib2 -lon LON LAT handles the rotated grid natively — no manual
reprojection math needed.
New plan at 2026-04-29-hrdps-canadian-prop-grid.md supersedes the
2026-04-13 one (which assumed HRRR-style idx + byte-range fetches that
don't apply). Probe deletion is a Stage 10 step in the new plan.
Bucket each cache entry into 5°×5° spatial chunks matching the layout
used by Weather.GridCache. fetch_bounds/3 now walks only the chunks
intersecting the viewport instead of the full 92k-cell CONUS map; a
typical zoomed viewport hits ~1500 cells per chunk × the chunks the
viewport overlaps instead of scanning all 92k.
Public API and observable behavior unchanged. Add cross-chunk-boundary
regression tests so future refactors can't silently flatten the layout.
Apply the same overlay refactor we did for /weather to the propagation
map. Score data used to flow as JSON via LiveView push_event
(update_scores + the 18-hour preload_forecast batch); switching bands
or scrubbing time meant a synchronous server roundtrip + ~MB JSON push
per change.
Now:
1. New /scores/cells endpoint returns a "PSCR" binary cell-pack —
columnar f32 lats/lons + u8 scores, ~3-4× smaller than JSON and
parsed via DataView + typed-array views with no JSON.parse.
2. JS hook owns score fetching. mount/moveend/band/time changes
trigger HTTP fetches against /scores/cells. After the initial
paint, every other forecast hour for the current band is fetched
in parallel in the background — timeline scrubs after the prefetch
finishes are pure cache hits with no network at all.
3. LiveView no longer pushes scores. select_band emits update_band_info
with band_mhz so the client knows what to refetch; select_time and
set_selected_time only update server-side state (URL, point detail
bbox). map_bounds, propagation_updated, and advance_now_cursor all
skip the score push — pack_scores, schedule_bounds_update,
schedule_preload_forecast, the :flush_bounds and :preload_forecast
handle_info clauses, and the start_async :initial_scores lifecycle
(incl. retry_initial_scores event) are all deleted.
Net result: instant band toggles after the first paint, no LiveView
WebSocket payloads larger than the timeline metadata, and ~370 fewer
lines in MapLive.
The ingress is cloudflared (Cloudflare Tunnel), which keeps an HTTP/2
connection pool to upstream Service IPs and reuses connections
aggressively. After a pod is removed from Service endpoints,
cloudflared can still hold in-flight HTTP/2 streams to the dying pod
for a noticeable window — 10 s wasn't enough, deploys still bounced
502 on LiveView reconnect.
Bump preStop sleep 10s → 25s and terminationGracePeriodSeconds 40s →
60s so cloudflared has time to turn over its upstream connections
before Phoenix begins draining.
Deploys were causing 502 Bad Gateway on LiveView reconnect. The kubelet
sent SIGTERM to the old pod and Phoenix immediately stopped accepting
connections, but the Service endpoint controller and ingress hadn't
yet propagated the pod's removal from endpoints — the browser's
reconnect raced the endpoint update and bounced 502 against the dead
pod, then gave up.
Add a 10s preStop sleep on the container and bump
terminationGracePeriodSeconds to 40 so the load balancer has time to
drain this pod before Phoenix begins shutting down.
The HRRR scalar cache can hold a previous run's forecast hours past
their useful life — the timeline was offering "current conditions"
points that were 2-3 hours old, which is more misleading than helpful
on a map labeled "Now".
Filter Weather.available_weather_valid_times() through a `now - 1h`
cutoff in mount, :weather_updated, and :propagation_pipeline_progress
so the timeline only ever shows the most recent analysis hour plus
forecast hours.
maxUnavailable: 1 was killing the only pod before the new one was
ready, taking the site down for the entire image pull + boot window
on every deploy. Switch to maxSurge: 1 / maxUnavailable: 0 so K8s
spins up the replacement and waits for its readinessProbe to pass
before tearing down the old pod.
The client moved to the binary cell-pack and canvas overlay last
commit, so the per-tile SVG endpoint, the TileRenderer module, and the
JSON shape of /weather/cells are all dead code. Remove them.
/weather/cells now always returns the binary WCEL pack (the ?format=bin
toggle is gone — there's only one format). The data-weather-tile-url
attribute and the weather_tile_url assign are dropped from the LiveView.
Also fix a latent bug surfaced once tests covered the default-layers
path: nil row values were writing the atom :nan into the binary segment
at runtime. Replaced with the explicit f32 quiet-NaN bit pattern so
Number.isNaN flags them correctly on the JS side.
Three changes that shrink first-paint and make layer toggles instant:
1. Server: /weather/cells now supports ?format=bin and ?layers=a,b,c. The
binary "WCEL" pack is ~3× smaller than JSON (Float32 columns, no key
overhead) and parses with DataView + typed-array views — essentially
memcpy on the client.
2. Client: replaced JSON parse + per-cell SVG <rect> with a custom
Leaflet canvas layer. The pack is columnar (lats/lons + Float32Array
per layer) and each draw is one fillRect per cell against a
precomputed 256-step color LUT. At low zoom (CONUS, ~80k cells) SVG
paint cost dominated; canvas keeps it under one frame.
3. Lazy prefetch: the initial fetch only requests the *current* layer
(~1/19 the bytes). Right after paint, a single background fetch
pulls every other layer into the same pack. Layer switches that land
after the prefetch finishes are pure recolor — no network, no
server CPU.
The /weather/tiles endpoint stays for backward compat. The JSON path of
/weather/cells stays too — the binary path is opt-in via ?format=bin.
Previously every layer toggle triggered 21 fresh tile requests with the
new layer in the URL — ~21 server-side SVG renders, plus the round-trip
overhead. Even with the GridCache fix the server still spent real CPU
generating thousands of <rect>s per tile, and the user felt it.
Add /weather/cells: a single JSON endpoint returning every layer field
for every cell in the requested viewport. The hook fetches once on
mount, time change, or pan/zoom, caches the cells in memory, and paints
them via L.svgOverlay. Layer switches now re-color the same cached
cells locally — zero network, zero server CPU.
The single SVG uses lat/lon coords with preserveAspectRatio="none";
Leaflet stretches it linearly to the bbox. There is mild equirect→
mercator distortion at low zoom over CONUS but it's imperceptible at
typical use (z6+) and the layer-toggle UX win is large.
Tile endpoint kept for back-compat. svgOverlay opacity matches the
prior tile renderer's 0.55 fill-opacity.
Each /weather/tiles request was running the full ScalarFile decode path
(File.read → gunzip → Msgpax.unpack → normalize) for any non-analysis
valid_time, because GridCache deliberately skipped forecast hours to keep
memory low. With 21 simultaneous viewport tiles all hitting the same few
chunk files, each tile took 1.5-3.2s of contended IO + CPU.
On a GridCache miss when a ScalarFile exists, hydrate GridCache from the
file once (deduped via the existing claim_fill primitive) and serve all
subsequent reads from ETS. Bound memory with prune_keep_latest/1 — keep
the 24 most-recently-touched valid_times, which covers a full HRRR run
(analysis + 18 forecasts) plus a few stragglers from the previous run.
In the steady state, the first viewport read for a given forecast hour
takes one full ScalarFile decode (~50-100 ms for 92k cells); every
subsequent tile, point lookup, and viewport read for the same hour
serves from ETS in <1 ms.
renderLayer was called during mounted() before selectedTime was seeded
from the dataset, so buildWeatherTileUrl returned null and no tile
layer was added. The overlay only appeared when the user clicked a
forecast hour and update_weather_overlay fired. Seed selectedTime
(and timelineData) from data attributes before the first renderLayer.
Also pulls db_connection 2.9.0 → 2.10.0 and spitfire 0.3.10 → 0.3.11 transitively.
oban_pro 1.6.14 adds suspended state support and conditional __opts__ @impl
to align with Oban v2.21+'s new Worker callback.
Set minReplicas == maxReplicas == 1 on the `prop` HPA so the cluster
runs exactly one app pod. HPA still owns the replica count (Flux
ignores spec.replicas on the Deployment), so this is the right knob —
no need to touch deployment.yaml. Restore autoscaling by raising both
bounds together.
Pre-existing CI failure (`tests/scorer_golden.rs`) — the Rust scorer
was still on the pre-recalibration weight vector, while the Elixir
side picked up new gradient-descent-fit weights in f98ee153.
- Update `DEFAULT_WEIGHTS` in band_config.rs to match the new Elixir
values (humidity 0.1262, time_of_day 0.0380, td_depression 0.1010,
refractivity 0.0986, sky 0.0841, season 0.1134, wind 0.0841,
rain 0.1431, pressure 0.0967, pwat 0.1147).
- Loosen the sum-to-one tolerance in `default_weights_sum_to_one`
from 1e-9 to 1e-3 to match Elixir's `assert_in_delta total, 1.0,
0.001` — gradient-descent fits land near 1.0 but not exact.
- Regenerate `tests/scores.bincode` from `mix rust.golden` so the
golden fixture reflects the recalibrated scorer outputs.
`cargo clippy --all-targets -- -D warnings` clean.
139 Rust unit tests + golden fixture pass.
3072 Elixir tests pass.
Clippy's `neg_cmp_op_on_partial_ord` lint fails on `-D warnings`,
which CI enforces. Using `<=` directly is clearer and equivalent for
the f64 inputs we pass (both are caller-supplied target pressures
that come from the level-key constants — neither is ever NaN).
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.
Mirror ScalarFile's 5°×5° chunk layout inside GridCache. Each ETS entry
is now `%{{lat_band, lon_band} => %{{lat, lon} => row}}` instead of a
flat `%{{lat,lon} => row}`. fetch_bounds/2 walks only the chunks that
intersect the requested viewport, so latest-hour pans become
proportional to visible chunks (~2-4) rather than O(92k cells).
fetch/1, fetch_point/3, put/2 and broadcast_put/2 keep their
pre-existing contracts.
Also closes the open items in docs/weather.md: Phase 3 is now done end
to end. The only remaining future-work item is moving WeatherLayers.derive
itself into the Rust pipeline so the BEAM doesn't pay the one-time
materialization pass per new valid_time — small runtime savings vs.
significant port effort, so left as documented future work.
Move the /weather render path off raw HRRR profile decoding + per-cell
SoundingParams + WeatherLayers derivation. The dominant cost was
re-deriving the same scalar fields on every viewport pan and timeline
scrub.
Server side:
- New Microwaveprop.Weather.ScalarFile persists pre-derived scalar
rows on NFS, bucketed into 5°×5° chunk files under
<base>/weather_scalars/<iso>/<lat_band>_<lon_band>.etf.gz. Viewport
reads only decode the chunks that overlap the requested bounds;
point-detail clicks read exactly one chunk.
- weather_grid_at/2 and weather_point_detail/3 prefer ScalarFile;
ProfilesFile is now the cold-start fallback only.
- warm_grid_cache_and_broadcast/1 and warm_grid_cache_from_latest_profile/0
also persist the scalar artifact, and the cold-derive path kicks
off a per-valid_time-locked async materialization so the next
reader gets the cheap path.
- NotifyListener.handle_propagation_ready/1 fires
Weather.materialize_scalar_file/1 in a detached Task on the Rust
pipeline's NOTIFY propagation_ready, so steady-state forecast hours
arrive with their scalar artifact already on disk.
- available_weather_valid_times/0 unions ScalarFile + ProfilesFile so
the timeline survives an aggressive retention sweep on either side.
Browser side (finding #8):
- Mount the legend Leaflet control once and patch its inner content
on layer changes instead of remove + re-add on every renderLayer.
- renderTimeline now keys off the rendered button set: selection-only
updates take an applyTimelineSelection patch path that restyles
existing buttons in place. Full innerHTML rebuild only fires when
timelineData itself changes.
Also fixes a credo nesting-depth flag in tile_renderer.ex by extracting
interpolate_pair/2 from the inner anonymous function.
Refit on the current contact corpus via mix recalibrate_scorer (val loss
0.140 vs 0.146 train, initial 0.434). Mostly small adjustments — the
notable shifts are time_of_day −0.0116 (now the smallest factor) and
rain +0.0069 (continues to be the largest single weight).
Sums to 1.0 (within rounding); per-band overrides not affected.
`mix prop.compare` runs both scorers against a recent contact sample
and measures both against achieved contact distance. Each run appends
to priv/calibration/history.jsonl, so trends in alg-ML divergence and
algorithm-distance correlation surface over weeks of running.
When drift exceeds threshold (alg-ML RMSE > 8 score points, or current
correlation > 0.10 below the history median), the task writes a
recommendation pointing at the right remediation:
mix recalibrate_scorer # algorithm drift → refit weights
mix propagation_train # ML drift → retrain on the new algorithm
That is the feedback loop: measure → recommend → recalibrate/retrain →
loop. Operational rather than automatic, so a bad data window can't
silently corrupt the model.
Pure analysis lives in Microwaveprop.Propagation.Calibration with its
own unit tests; the Mix task only handles data loading, file I/O, and
console formatting.
- Unify haversine_km / deg_to_rad: Geo is the canonical home (atan2
form for antipodal stability); Radio, common_volume, rain_scatter,
ionosphere, terrain/srtm, terrain/elevation_client now delegate.
- Drop safe_min/safe_max wrappers in favor of Enum.min/max defaults.
- Move parse_float / parse_int to MicrowavepropWeb.LiveHelpers; eme,
path, and rover LiveViews import from there.
- Extract MicrowavepropWeb.LocationResolver from the eme/path
resolve_source / resolve_location duplicate. Standardize the kind
key and "Invalid grid square" error message.
- Convert Scorer cond chains (humidity, td_depression, sky, wind,
rain, pwat, pressure) to multi-clause guarded defs, matching the
existing classify_time_period style.
- extract_profile_fields uses a named map accumulator instead of a
6-tuple; build_path_conditions guards on %{temps: []} / %{dewpoints: []}.
- Collapse scores_file clamp/3 to max |> min.
- humidity_effect_label lives in BandConfig; map_live and path_live
both use it.
