recipe/http.py

"""Fetching, with a cache that knows which documents can never change.

An archived news release is immutable by definition: it is what the agency
published on a particular morning, and the agency's own archive page is the
promise that it stays that way. A schedule page is the opposite -- it changes
whenever a release moves, which is the whole reason the calendar is a table
here. The cache treats the two differently, and nothing else in the pipeline
has to think about it.
"""

from __future__ import annotations

import asyncio
import hashlib
import json
import logging
import time
from dataclasses import dataclass
from datetime import UTC, datetime
from pathlib import Path
from urllib.parse import urlsplit

import httpx

LOGGER = logging.getLogger(__name__)

# Agencies publish no numeric rate limit. This is deliberately slower than what
# the hosts would tolerate: a backfill is a few thousand documents, it runs
# once, and nothing downstream is latency sensitive.
DEFAULT_RPS = 4.0
# Census answers 429 well below the rate the other agencies tolerate, and a
# backfill that ignores it loses whole years -- fifty-nine housing releases in
# one run, recorded as unreadable when they were simply never served. These are
# the rates each host is asked at, and they are per host, so one slow agency
# does not hold up the others.
HOST_RPS = {
    # Each yield-curve CSV is generated on request and takes about seventeen
    # seconds. Spacing the starts closely is what lets several be in flight;
    # the concurrency limit in the pipeline is what keeps that polite.
    "home.treasury.gov": 4.0,
    "www.treasurydirect.gov": 4.0,
}
TIMEOUT = httpx.Timeout(60.0, connect=30.0)
RETRY_STATUS = frozenset({429, 500, 502, 503, 504})
# A 429 is the host asking for a pause, not a transient error, so it is waited
# out for much longer than a 503.
THROTTLED_BACKOFF = 20.0
MAX_ATTEMPTS = 5
# A request that is allowed to come back empty is a probe -- "is there a
# release for this week?" -- and ETA answers such a probe with 502 rather than
# 404. Retrying that five times with exponential backoff costs a minute per
# URL and buys nothing, so a probe gives up quickly.
MAX_ATTEMPTS_WHEN_OPTIONAL = 2
# How long a "there is nothing here" answer is believed. Long enough that a
# rebuild does not re-probe two thousand URLs, short enough that a release
# published today is found tomorrow.
MISSING_MAX_AGE = 2 * 24 * 3600


@dataclass(frozen=True, slots=True)
class Artifact:
    """One downloaded document and everything needed to prove where it came from."""

    url: str
    content: bytes
    sha256: str
    content_type: str
    downloaded_at: datetime
    from_cache: bool

    @property
    def text(self) -> str:
        # Agency releases are ASCII with the occasional Windows-1252 dash; a
        # strict decode would fail on those and lose a whole release over a
        # punctuation mark.
        return self.content.decode("utf-8", errors="replace")

    @property
    def length(self) -> int:
        return len(self.content)


def cache_key(url: str) -> str:
    return hashlib.sha1(url.encode("utf-8")).hexdigest()


class Fetcher:
    """Rate-limited, cached, retrying HTTP client for the agency hosts."""

    def __init__(
        self,
        *,
        user_agent: str,
        cache_dir: Path,
        requests_per_second: float = DEFAULT_RPS,
        client: httpx.AsyncClient | None = None,
        force_revalidate: bool = False,
    ) -> None:
        self.user_agent = user_agent
        self.cache_dir = Path(cache_dir)
        self.cache_dir.mkdir(parents=True, exist_ok=True)
        self._interval = 1.0 / max(requests_per_second, 0.1)
        self._host_intervals = {
            host: 1.0 / min(rate, max(requests_per_second, 0.1))
            for host, rate in HOST_RPS.items()
        }
        # `--rebuild` re-reads the pages that change -- the archive indexes and
        # the schedules -- and still serves the archived releases from cache.
        # An archived release cannot change, so refetching a decade of them to
        # pick up a new index entry would be pure waste asked of the agency.
        self.force_revalidate = force_revalidate
        self._client = client
        self._owned = client is None
        # One gate per host: BLS being slow is no reason to stall Census.
        self._gates: dict[str, asyncio.Lock] = {}
        self._last: dict[str, float] = {}

    async def __aenter__(self) -> Fetcher:
        if self._client is None:
            self._client = httpx.AsyncClient(
                timeout=TIMEOUT,
                follow_redirects=True,
                headers={"User-Agent": self.user_agent, "Accept-Encoding": "gzip, deflate"},
            )
        return self

    async def __aexit__(self, *_exc) -> None:
        if self._owned and self._client is not None:
            await self._client.aclose()
            self._client = None

    def _paths(self, url: str) -> tuple[Path, Path]:
        key = cache_key(url)
        return self.cache_dir / f"{key}.bin", self.cache_dir / f"{key}.json"

    def known_missing(self, url: str, *, max_age_seconds: float = MISSING_MAX_AGE) -> bool:
        """Whether this URL was recently found not to exist.

        Discovery probes thousands of URLs that will never exist -- a week with
        no release, a month Census never archived -- and without this every
        rebuild asks the agencies for all of them again.
        """

        _, meta = self._paths(url)
        if not meta.is_file():
            return False
        try:
            record = json.loads(meta.read_text(encoding="utf-8"))
        except json.JSONDecodeError:
            return False
        if not record.get("missing"):
            return False
        age = (datetime.now(UTC) - datetime.fromisoformat(record["checked_at"])).total_seconds()
        return age <= max_age_seconds

    def _store_missing(self, url: str) -> None:
        _, meta = self._paths(url)
        meta.write_text(
            json.dumps(
                {"url": url, "missing": True, "checked_at": datetime.now(UTC).isoformat()},
                indent=2,
            ),
            encoding="utf-8",
        )

    def cached(self, url: str, *, max_age_seconds: float | None = None) -> Artifact | None:
        blob, meta = self._paths(url)
        if not (blob.is_file() and meta.is_file()):
            return None
        try:
            record = json.loads(meta.read_text(encoding="utf-8"))
        except json.JSONDecodeError:
            return None
        downloaded_at = datetime.fromisoformat(record["downloaded_at"])
        if max_age_seconds is not None:
            age = (datetime.now(UTC) - downloaded_at).total_seconds()
            if age > max_age_seconds:
                return None
        content = blob.read_bytes()
        digest = hashlib.sha256(content).hexdigest()
        if digest != record.get("sha256"):
            # A cache entry that does not hash to what it claims is not a cache
            # entry. Silently trusting it would put an unprovable value into a
            # dataset whose whole promise is provenance.
            LOGGER.warning("cache entry for %s does not match its digest; refetching", url)
            return None
        return Artifact(
            url=url,
            content=content,
            sha256=digest,
            content_type=record.get("content_type", ""),
            downloaded_at=downloaded_at,
            from_cache=True,
        )

    def _store(self, url: str, content: bytes, content_type: str) -> Artifact:
        blob, meta = self._paths(url)
        digest = hashlib.sha256(content).hexdigest()
        now = datetime.now(UTC)
        blob.write_bytes(content)
        meta.write_text(
            json.dumps(
                {
                    "url": url,
                    "sha256": digest,
                    "content_type": content_type,
                    "content_length": len(content),
                    "downloaded_at": now.isoformat(),
                },
                indent=2,
            ),
            encoding="utf-8",
        )
        return Artifact(
            url=url,
            content=content,
            sha256=digest,
            content_type=content_type,
            downloaded_at=now,
            from_cache=False,
        )

    async def _throttle(self, host: str) -> None:
        interval = self._host_intervals.get(host, self._interval)
        gate = self._gates.setdefault(host, asyncio.Lock())
        async with gate:
            elapsed = time.monotonic() - self._last.get(host, 0.0)
            if elapsed < interval:
                await asyncio.sleep(interval - elapsed)
            self._last[host] = time.monotonic()

    async def get(
        self,
        url: str,
        *,
        max_age_seconds: float | None = None,
        allow_missing: bool = False,
        missing_max_age: float = MISSING_MAX_AGE,
    ) -> Artifact | None:
        """Fetch a document, or return the cached copy.

        `max_age_seconds=None` means "this document is immutable" -- an
        archived release. Pass a number for anything the agency rewrites.
        """

        if self.force_revalidate and max_age_seconds is not None:
            max_age_seconds = 0
        hit = self.cached(url, max_age_seconds=max_age_seconds)
        if hit is not None:
            return hit
        if allow_missing and self.known_missing(url, max_age_seconds=missing_max_age):
            return None
        if self._client is None:
            raise RuntimeError("use the Fetcher as an async context manager")
        host = urlsplit(url).netloc
        last_error: Exception | None = None
        attempts = MAX_ATTEMPTS_WHEN_OPTIONAL if allow_missing else MAX_ATTEMPTS
        for attempt in range(1, attempts + 1):
            await self._throttle(host)
            try:
                response = await self._client.get(url)
            except httpx.HTTPError as error:  # pragma: no cover - network dependent
                last_error = error
                await asyncio.sleep(min(2**attempt, 30))
                continue
            if response.status_code == 404 and allow_missing:
                self._store_missing(url)
                return None
            if response.status_code in RETRY_STATUS:
                await asyncio.sleep(
                    THROTTLED_BACKOFF * attempt
                    if response.status_code == 429
                    else min(2**attempt, 30)
                )
                last_error = httpx.HTTPStatusError(
                    f"{response.status_code} for {url}", request=response.request,
                    response=response,
                )
                continue
            response.raise_for_status()
            return self._store(
                url, response.content, response.headers.get("content-type", "").split(";")[0]
            )
        if allow_missing:
            LOGGER.info("nothing at %s (%s)", url, last_error)
            self._store_missing(url)
            return None
        raise RuntimeError(f"could not fetch {url}: {last_error}")


__all__ = ["Artifact", "Fetcher", "cache_key"]