prop/rust/prop_grid_rs/src/fetcher.rs
Graham McIntire efa3cc804d
perf(rust): disk-backed HRRR idx cache + scale workers off talos5
Shared NFS idx cache at /data/hrrr_idx lets all prop-grid-rs and
hrrr-point-rs replicas deduplicate redundant NOAA S3 idx fetches.
Opt-in via HRRR_IDX_CACHE_DIR; falls back to in-memory-only when unset.

Also scales prop-grid-rs to 3 replicas (no talos5 pinning),
hrrr-point-rs to 2 replicas, and drops hot pod replicas 4→3 so the
physical-host anti-affinity still fits with talos5 cordoned off for
Postgres.
2026-04-20 13:18:38 -05:00

599 lines
18 KiB
Rust

//! HRRR fetcher. 1:1 port of the grid-fetch path of
//! `lib/microwaveprop/weather/hrrr_client.ex`:
//! * URL derivation (`hrrr.tHHz.wrfsfcfFH.grib2` / `wrfprsfFH.grib2`)
//! * `.idx` parse + per-`(var, level)` byte-range lookup
//! * range merge + parallel `Range:` downloads
//! * in-process 1h TTL `.idx` memoization (idx files are immutable per run)
//! * retry on HTTP 429/5xx with jittered exponential backoff
use std::collections::HashMap;
use std::fs;
use std::io::Write;
use std::path::{Path, PathBuf};
use std::sync::Mutex;
use std::time::{Duration, Instant};
use chrono::{DateTime, NaiveDate, Timelike, Utc};
use futures::stream::{FuturesUnordered, StreamExt};
use reqwest::header::{HeaderMap, HeaderValue, RANGE};
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Product {
Surface,
Pressure,
}
#[derive(Debug, Clone)]
pub struct IdxEntry {
pub msg: u32,
pub offset: u64,
pub var: String,
pub level: String,
}
#[derive(Debug, Clone, Copy)]
pub struct ByteRange {
pub start: u64,
pub end: u64,
}
#[derive(Debug, thiserror::Error)]
pub enum FetchError {
#[error("http: {0}")]
Http(#[from] reqwest::Error),
#[error("idx HTTP {status}")]
IdxStatus { status: u16 },
#[error("grib HTTP {status}")]
GribStatus { status: u16 },
#[error("idx text was empty")]
IdxEmpty,
}
pub const HRRR_BASE_DEFAULT: &str = "https://noaa-hrrr-bdp-pds.s3.amazonaws.com";
pub const IDX_TTL: Duration = Duration::from_secs(3_600);
pub const MAX_PARALLEL_RANGES: usize = 8;
pub const REQUEST_TIMEOUT: Duration = Duration::from_secs(120);
pub const GRID_PRESSURE_LEVELS: &[u16] = &[
1000, 975, 950, 925, 900, 875, 850, 825, 800, 775, 750, 725, 700,
];
pub const SURFACE_MESSAGES: &[(&str, &str)] = &[
("TMP", "2 m above ground"),
("DPT", "2 m above ground"),
("PRES", "surface"),
("HPBL", "surface"),
("PWAT", "entire atmosphere (considered as a single layer)"),
("UGRD", "10 m above ground"),
("VGRD", "10 m above ground"),
("TCDC", "entire atmosphere"),
("APCP", "surface"),
];
pub fn hrrr_url(
base: &str,
date: NaiveDate,
hour: u8,
product: Product,
forecast_hour: u8,
) -> String {
let date_str = date.format("%Y%m%d").to_string();
let hh = format!("{hour:02}");
let fh = format!("{forecast_hour:02}");
let file = match product {
Product::Surface => format!("hrrr.t{hh}z.wrfsfcf{fh}.grib2"),
Product::Pressure => format!("hrrr.t{hh}z.wrfprsf{fh}.grib2"),
};
format!("{base}/hrrr.{date_str}/conus/{file}")
}
pub fn parse_idx(text: &str) -> Vec<IdxEntry> {
text.split('\n')
.filter(|l| !l.is_empty())
.filter_map(parse_idx_line)
.collect()
}
fn parse_idx_line(line: &str) -> Option<IdxEntry> {
let mut parts = line.splitn(8, ':');
let msg: u32 = parts.next()?.parse().ok()?;
let offset: u64 = parts.next()?.parse().ok()?;
parts.next()?; // date
let var = parts.next()?.to_string();
let level = parts.next()?.to_string();
Some(IdxEntry {
msg,
offset,
var,
level,
})
}
pub fn byte_ranges_for_messages(idx: &[IdxEntry], wanted: &[(String, String)]) -> Vec<ByteRange> {
let mut out = Vec::new();
for (var, level) in wanted {
for (i, entry) in idx.iter().enumerate() {
if entry.var == *var && entry.level == *level {
let end = idx
.get(i + 1)
.map(|nxt| nxt.offset.saturating_sub(1))
.unwrap_or(entry.offset + 10_000_000);
out.push(ByteRange {
start: entry.offset,
end,
});
}
}
}
out
}
pub fn merge_ranges(mut ranges: Vec<ByteRange>) -> Vec<ByteRange> {
if ranges.is_empty() {
return ranges;
}
ranges.sort_by_key(|r| r.start);
let mut out: Vec<ByteRange> = Vec::with_capacity(ranges.len());
for r in ranges {
match out.last_mut() {
Some(prev) if r.start <= prev.end.saturating_add(1) => {
prev.end = prev.end.max(r.end);
}
_ => out.push(r),
}
}
out
}
pub fn nearest_hrrr_hour(dt: DateTime<Utc>) -> DateTime<Utc> {
let total_seconds = dt.minute() as i64 * 60 + dt.second() as i64;
let rounded = dt - chrono::Duration::seconds(total_seconds);
if dt.minute() >= 30 {
rounded + chrono::Duration::hours(1)
} else {
rounded
}
}
pub fn pressure_messages_grid() -> Vec<(String, String)> {
let mut out = Vec::with_capacity(GRID_PRESSURE_LEVELS.len() * 3);
for &level in GRID_PRESSURE_LEVELS {
for var in ["TMP", "DPT", "HGT"] {
out.push((var.to_string(), format!("{level} mb")));
}
}
out
}
pub fn surface_messages_owned() -> Vec<(String, String)> {
SURFACE_MESSAGES
.iter()
.map(|(v, l)| ((*v).to_string(), (*l).to_string()))
.collect()
}
// ── Idx cache ────────────────────────────────────────────────────────
// Disk TTL is longer than the in-memory TTL because idx files are
// immutable per HRRR run — cross-replica sharing on NFS benefits from
// keeping hits alive across pod restarts and the next hourly cycle.
pub const IDX_DISK_TTL: Duration = Duration::from_secs(24 * 3_600);
#[derive(Debug, Clone)]
struct IdxCacheEntry {
body: String,
inserted: Instant,
}
#[derive(Debug)]
pub struct DiskIdxCache {
dir: PathBuf,
ttl: Duration,
}
impl DiskIdxCache {
pub fn new(dir: impl AsRef<Path>, ttl: Duration) -> Self {
let dir = dir.as_ref().to_path_buf();
let _ = fs::create_dir_all(&dir);
Self { dir, ttl }
}
fn path_for(&self, url: &str) -> PathBuf {
// Sanitize URL into a safe filename. Collisions would require
// two distinct URLs to differ only in unsafe characters mapped
// to '_' — not possible for HRRR S3 URLs.
let mut sanitized = String::with_capacity(url.len());
for c in url.chars() {
if c.is_ascii_alphanumeric() || c == '.' || c == '-' || c == '_' {
sanitized.push(c);
} else {
sanitized.push('_');
}
}
self.dir.join(sanitized)
}
pub fn get(&self, url: &str) -> Option<String> {
let path = self.path_for(url);
let meta = fs::metadata(&path).ok()?;
let mtime = meta.modified().ok()?;
if mtime.elapsed().ok()? > self.ttl {
return None;
}
fs::read_to_string(&path).ok()
}
pub fn put(&self, url: &str, body: &str) {
// Atomic write: tmp file + rename so concurrent readers across
// replicas never observe a partial body.
let path = self.path_for(url);
let tmp = path.with_extension("tmp");
if let Ok(mut f) = fs::File::create(&tmp) {
if f.write_all(body.as_bytes()).is_ok() {
drop(f);
let _ = fs::rename(&tmp, &path);
}
}
}
}
#[derive(Debug, Default)]
pub struct IdxCache {
inner: Mutex<HashMap<String, IdxCacheEntry>>,
disk: Option<DiskIdxCache>,
}
impl IdxCache {
pub fn new() -> Self {
Self::default()
}
pub fn with_disk(dir: impl AsRef<Path>) -> Self {
Self {
inner: Mutex::new(HashMap::new()),
disk: Some(DiskIdxCache::new(dir, IDX_DISK_TTL)),
}
}
pub fn get(&self, url: &str) -> Option<String> {
if let Ok(map) = self.inner.lock() {
if let Some(entry) = map.get(url) {
if entry.inserted.elapsed() < IDX_TTL {
return Some(entry.body.clone());
}
}
}
if let Some(disk) = &self.disk {
if let Some(body) = disk.get(url) {
if let Ok(mut map) = self.inner.lock() {
map.insert(
url.to_string(),
IdxCacheEntry {
body: body.clone(),
inserted: Instant::now(),
},
);
}
return Some(body);
}
}
None
}
pub fn put(&self, url: String, body: String) {
if let Some(disk) = &self.disk {
disk.put(&url, &body);
}
if let Ok(mut map) = self.inner.lock() {
map.insert(
url,
IdxCacheEntry {
body,
inserted: Instant::now(),
},
);
}
}
}
// ── HTTP client ──────────────────────────────────────────────────────
pub struct HrrrClient {
http: reqwest::Client,
base: String,
idx_cache: IdxCache,
}
impl HrrrClient {
pub fn new(base: impl Into<String>) -> Result<Self, FetchError> {
let http = reqwest::Client::builder()
.timeout(REQUEST_TIMEOUT)
.pool_max_idle_per_host(MAX_PARALLEL_RANGES)
.build()?;
// HRRR_IDX_CACHE_DIR opts into a disk-backed idx cache shared
// across replicas via NFS (e.g. `/data/hrrr_idx`). Unset leaves
// the in-memory-only behaviour intact.
let idx_cache = match std::env::var("HRRR_IDX_CACHE_DIR") {
Ok(dir) if !dir.is_empty() => IdxCache::with_disk(dir),
_ => IdxCache::new(),
};
Ok(Self {
http,
base: base.into(),
idx_cache,
})
}
pub fn default_base() -> Result<Self, FetchError> {
Self::new(HRRR_BASE_DEFAULT)
}
pub fn base(&self) -> &str {
&self.base
}
pub async fn fetch_idx(&self, url: &str) -> Result<String, FetchError> {
if let Some(body) = self.idx_cache.get(url) {
return Ok(body);
}
let body = retry_get_text(&self.http, url).await?;
if body.is_empty() {
return Err(FetchError::IdxEmpty);
}
self.idx_cache.put(url.to_string(), body.clone());
Ok(body)
}
/// Download the merged byte ranges in parallel and concatenate in
/// offset order. Mirrors `download_grib_ranges_remote/2` in Elixir.
pub async fn download_grib_ranges(
&self,
url: &str,
ranges: Vec<ByteRange>,
) -> Result<Vec<u8>, FetchError> {
if ranges.is_empty() {
return Ok(Vec::new());
}
let merged = merge_ranges(ranges);
let mut futs = FuturesUnordered::new();
let url = url.to_string();
for r in merged.iter().copied() {
let client = self.http.clone();
let url = url.clone();
futs.push(async move {
let bytes = retry_get_range(&client, &url, r).await?;
Ok::<(u64, Vec<u8>), FetchError>((r.start, bytes))
});
}
// Cap concurrent in-flight requests at MAX_PARALLEL_RANGES.
let mut results: Vec<(u64, Vec<u8>)> = Vec::with_capacity(merged.len());
while let Some(item) = futs.next().await {
results.push(item?);
}
results.sort_by_key(|(off, _)| *off);
let total: usize = results.iter().map(|(_, b)| b.len()).sum();
let mut out = Vec::with_capacity(total);
for (_, b) in results {
out.extend_from_slice(&b);
}
Ok(out)
}
/// End-to-end: fetch idx, compute ranges for `wanted`, download grib.
pub async fn fetch_product_blob(
&self,
date: NaiveDate,
hour: u8,
product: Product,
forecast_hour: u8,
wanted: &[(String, String)],
) -> Result<Vec<u8>, FetchError> {
let url = hrrr_url(&self.base, date, hour, product, forecast_hour);
let idx_url = format!("{url}.idx");
let idx_body = self.fetch_idx(&idx_url).await?;
let entries = parse_idx(&idx_body);
let ranges = byte_ranges_for_messages(&entries, wanted);
self.download_grib_ranges(&url, ranges).await
}
}
async fn retry_get_text(client: &reqwest::Client, url: &str) -> Result<String, FetchError> {
for attempt in 0..5 {
match client.get(url).send().await {
Ok(resp) => {
let status = resp.status().as_u16();
if status == 200 {
return Ok(resp.text().await?);
}
if !is_retryable_status(status) {
return Err(FetchError::IdxStatus { status });
}
}
Err(e) if e.is_connect() || e.is_timeout() => {
tracing::warn!(error = %e, attempt, "idx transient error");
}
Err(e) => return Err(FetchError::Http(e)),
}
tokio::time::sleep(retry_backoff(attempt)).await;
}
Err(FetchError::IdxStatus { status: 0 })
}
async fn retry_get_range(
client: &reqwest::Client,
url: &str,
range: ByteRange,
) -> Result<Vec<u8>, FetchError> {
let value = HeaderValue::from_str(&format!("bytes={}-{}", range.start, range.end))
.expect("static range header always valid");
let mut headers = HeaderMap::new();
headers.insert(RANGE, value);
for attempt in 0..5 {
match client.get(url).headers(headers.clone()).send().await {
Ok(resp) => {
let status = resp.status().as_u16();
if status == 206 || status == 200 {
return Ok(resp.bytes().await?.to_vec());
}
if !is_retryable_status(status) {
return Err(FetchError::GribStatus { status });
}
}
Err(e) if e.is_connect() || e.is_timeout() => {
tracing::warn!(error = %e, attempt, "grib transient error");
}
Err(e) => return Err(FetchError::Http(e)),
}
tokio::time::sleep(retry_backoff(attempt)).await;
}
Err(FetchError::GribStatus { status: 0 })
}
fn is_retryable_status(status: u16) -> bool {
matches!(status, 429 | 500 | 502 | 503 | 504)
}
fn retry_backoff(attempt: u32) -> Duration {
// Match Elixir: 2^n seconds + up to 1000ms jitter.
let base = 2u64.pow(attempt).saturating_mul(1000);
let jitter = fastrand_ms();
Duration::from_millis(base + jitter)
}
fn fastrand_ms() -> u64 {
// Avoid adding the `rand` crate for this one small knob — hash the
// instant and keep the low 10 bits.
let nanos = std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.map(|d| d.as_nanos() as u64)
.unwrap_or(0);
nanos & 0x3ff // 0..1023 ms
}
#[cfg(test)]
mod tests {
use super::*;
use chrono::TimeZone;
#[test]
fn url_matches_elixir() {
let date = NaiveDate::from_ymd_opt(2026, 4, 19).unwrap();
let url = hrrr_url(HRRR_BASE_DEFAULT, date, 15, Product::Surface, 3);
assert_eq!(
url,
"https://noaa-hrrr-bdp-pds.s3.amazonaws.com/hrrr.20260419/conus/hrrr.t15z.wrfsfcf03.grib2"
);
}
#[test]
fn idx_parsing_drops_bad_lines() {
let idx = "1:0:d=2024:TMP:2 m above ground:anl:\n<html>bad</html>\n2:12345:d=2024:DPT:2 m above ground:anl:";
let parsed = parse_idx(idx);
assert_eq!(parsed.len(), 2);
assert_eq!(parsed[0].var, "TMP");
assert_eq!(parsed[1].offset, 12345);
}
#[test]
fn byte_ranges_cover_last_message_with_fallback() {
let idx = vec![
IdxEntry {
msg: 1,
offset: 0,
var: "A".into(),
level: "x".into(),
},
IdxEntry {
msg: 2,
offset: 100,
var: "B".into(),
level: "y".into(),
},
];
let wanted = vec![("A".into(), "x".into()), ("B".into(), "y".into())];
let r = byte_ranges_for_messages(&idx, &wanted);
assert_eq!(r.len(), 2);
assert_eq!(r[0].start, 0);
assert_eq!(r[0].end, 99);
assert_eq!(r[1].start, 100);
assert_eq!(r[1].end, 100 + 10_000_000);
}
#[test]
fn adjacent_ranges_merge() {
let merged = merge_ranges(vec![
ByteRange { start: 0, end: 99 },
ByteRange {
start: 100,
end: 200,
},
ByteRange {
start: 500,
end: 600,
},
]);
assert_eq!(merged.len(), 2);
assert_eq!(merged[0].start, 0);
assert_eq!(merged[0].end, 200);
}
#[test]
fn nearest_hour_rounds_properly() {
let base = Utc.with_ymd_and_hms(2026, 4, 19, 15, 29, 0).unwrap();
assert_eq!(nearest_hrrr_hour(base).hour(), 15);
let up = Utc.with_ymd_and_hms(2026, 4, 19, 15, 30, 0).unwrap();
assert_eq!(nearest_hrrr_hour(up).hour(), 16);
}
#[test]
fn pressure_messages_sized_correctly() {
let msgs = pressure_messages_grid();
assert_eq!(msgs.len(), GRID_PRESSURE_LEVELS.len() * 3);
}
#[test]
fn idx_cache_returns_within_ttl() {
let cache = IdxCache::new();
cache.put("u".into(), "body".into());
assert_eq!(cache.get("u").as_deref(), Some("body"));
assert!(cache.get("missing").is_none());
}
#[test]
fn disk_idx_cache_round_trips() {
let tmp = tempfile::TempDir::new().unwrap();
let cache = DiskIdxCache::new(tmp.path(), Duration::from_secs(3_600));
cache.put("https://example.com/a.grib2.idx", "body1");
assert_eq!(
cache.get("https://example.com/a.grib2.idx").as_deref(),
Some("body1")
);
assert!(cache.get("https://example.com/missing.idx").is_none());
}
#[test]
fn disk_idx_cache_respects_ttl() {
let tmp = tempfile::TempDir::new().unwrap();
let cache = DiskIdxCache::new(tmp.path(), Duration::from_millis(10));
cache.put("u", "body");
std::thread::sleep(Duration::from_millis(30));
assert!(cache.get("u").is_none());
}
#[test]
fn idx_cache_disk_fallback_rewarms_memory() {
let tmp = tempfile::TempDir::new().unwrap();
// Seed disk via one cache instance...
let c1 = IdxCache::with_disk(tmp.path());
c1.put("u".into(), "body".into());
// A fresh cache simulates a new pod — in-memory map is empty
// but the disk entry must still resolve.
let c2 = IdxCache::with_disk(tmp.path());
assert_eq!(c2.get("u").as_deref(), Some("body"));
}
}