"""Exception taxonomy for ``dataretrieval``.
Every service module (``nwis``, ``wqp``, ``nldi``, ``waterdata``,
``streamstats``) raises a subclass of :class:`DataRetrievalError` when a request
fails, so one ``except dataretrieval.DataRetrievalError`` catches them all. That
includes connection-level failures (timeouts, DNS, refused connections), which
remain inside this taxonomy rather than leaking ``httpx`` exceptions. A
deterministic failure is :class:`NetworkError`; a recoverable failure that
exhausts retries during fan-out is a resumable ``ServiceInterrupted``.
Most failures are an :class:`HTTPError` carrying the response ``.status_code``,
of which :class:`TransientError` (429 / 5xx) is the retryable subset. The rest
aren't a plain status: :class:`RequestTooLarge` (with :class:`URLTooLong` /
:class:`Unchunkable`), :class:`NetworkError` (a failed connection, per above),
:class:`NoSitesError`, and :class:`ConfigurationError` for an unusable setting.
:func:`error_for_status` maps a status to its type.
This module has no third-party runtime dependencies -- ``httpx`` is imported only
for type checking. Any module can therefore import it without pulling in pandas
or httpx, and without risking an import cycle.
"""
from __future__ import annotations
import math
from datetime import datetime, timezone
from email.utils import parsedate_to_datetime
from typing import TYPE_CHECKING, Any, ClassVar
if TYPE_CHECKING:
import httpx
__all__ = [
"DataRetrievalError",
"HTTPError",
"TransientError",
"RateLimited",
"ServiceUnavailable",
"RequestTooLarge",
"URLTooLong",
"Unchunkable",
"NetworkError",
"NoSitesError",
"ConfigurationError",
"error_for_status",
"parse_retry_after",
]
[docs]
class DataRetrievalError(Exception):
"""Base class for every failed-request error in ``dataretrieval``.
Catch it to handle any USGS or EPA service failure uniformly, and branch on
the read-anywhere fields below without needing the concrete subclass::
try:
df, md = dataretrieval.waterdata.get_daily(...)
except dataretrieval.DataRetrievalError as e:
if e.retryable: # 429 / 5xx / connection failure
time.sleep(e.retry_after or backoff)
... # re-issue the request
elif e.status_code == 404: # ``None`` unless an HTTP status error
...
else:
raise
Connection-level failures (timeouts, DNS) remain subclasses of this base:
:class:`NetworkError` when deterministic, or a resumable
``ServiceInterrupted`` when recoverable fan-out retries are exhausted.
"""
#: HTTP status that triggered the error, or ``None`` for errors without one
#: (connection failure, too-long URL, no data). Set by :class:`HTTPError`.
status_code: int | None = None
#: Seconds the server asked us to wait before retrying (its ``Retry-After``
#: header), or ``None`` when it gave no hint. Set by :class:`TransientError`.
retry_after: float | None = None
#: Whether re-issuing the same request might succeed -- ``True`` for the
#: transient HTTP statuses (429 / 5xx, :class:`TransientError`) and for
#: connection failures (:class:`NetworkError`); ``False`` otherwise.
retryable: ClassVar[bool] = False
# These errors get pickled back across process boundaries (a lithops /
# multiprocessing worker returns whatever it raises). Default ``BaseException``
# pickling rebuilds via ``cls(*args)``, which these subclasses can't survive --
# keyword-only constructor fields, and ``ChunkInterrupted`` builds its message
# internally. So reconstruct via ``__new__`` + the standard getstate/setstate
# protocol, bypassing ``__init__``; a subclass drops unpicklable state by
# overriding ``__getstate__`` (see ``ChunkInterrupted``).
def __reduce__(self) -> tuple[Any, ...]:
return (_new_error, (self.__class__,), self.__getstate__())
def __getstate__(self) -> dict[str, Any]:
return {"args": self.args, **self.__dict__}
def __setstate__(self, state: dict[str, Any] | None) -> None:
state = state or {}
self.args = state.pop("args", ())
self.__dict__.update(state)
def _new_error(cls: type[DataRetrievalError]) -> DataRetrievalError:
"""Build a blank :class:`DataRetrievalError` for unpickling.
Bypasses ``__init__``; pickle then calls ``__setstate__`` to restore state.
"""
return cls.__new__(cls)
# --- HTTP status errors --------------------------------------------------
[docs]
class HTTPError(DataRetrievalError):
"""The service returned an error HTTP status.
The numeric status is on :attr:`status_code`; branch on it, e.g.
``except HTTPError as e: ... if e.status_code == 404``. :class:`TransientError`
(429 / 5xx) is the retryable subset, and is itself an ``HTTPError``. The one
exception to "a status is an ``HTTPError``" is a request the service rejects
as too long: it surfaces as :class:`URLTooLong` (a :class:`RequestTooLarge`),
*not* an ``HTTPError``. Catch :class:`DataRetrievalError` to be certain of
spanning every failure. See :func:`error_for_status` for the full mapping.
Parameters
----------
message : str
Human-readable error message.
status_code : int
The HTTP status the service returned.
"""
def __init__(self, message: str, *, status_code: int) -> None:
super().__init__(message)
self.status_code = status_code
[docs]
class TransientError(HTTPError):
"""A 429 or 5xx the server may serve on a later try.
:class:`RateLimited` covers 429 and :class:`ServiceUnavailable` covers 5xx.
This only classifies the condition; it does not itself retry. Whether to
retry is up to the calling path: a single-shot request raises it for the
caller to handle (e.g. wait :attr:`retry_after` seconds, then re-issue),
while the Water Data chunker retries and resumes automatically.
Parameters
----------
message : str
Human-readable error message.
status_code : int, optional
The HTTP status the service returned. Defaults to the leaf's canonical
code (429 / 503) when omitted; :func:`error_for_status` always passes the
real status.
retry_after : float, optional
Seconds to wait before retrying, parsed from the ``Retry-After`` response
header; ``None`` when the header is absent or unparseable.
"""
retryable: ClassVar[bool] = True
#: Canonical status a concrete transient stamps when built without an
#: explicit ``status_code`` (:class:`RateLimited` = 429,
#: :class:`ServiceUnavailable` = 503). ``TransientError`` itself is abstract
#: and sets none, so constructing it bare requires ``status_code``.
_DEFAULT_STATUS: ClassVar[int]
def __init__(
self,
message: str,
*,
status_code: int | None = None,
retry_after: float | None = None,
) -> None:
if status_code is None:
status_code = getattr(self, "_DEFAULT_STATUS", None)
if status_code is None:
raise TypeError(
f"{type(self).__name__} requires status_code "
"(only the RateLimited / ServiceUnavailable leaves default it)"
)
super().__init__(message, status_code=status_code)
self.retry_after = retry_after
[docs]
class RateLimited(TransientError):
"""A request was rejected with HTTP 429 (too many requests)."""
_DEFAULT_STATUS = 429
[docs]
class ServiceUnavailable(TransientError):
"""A request was rejected with a server error (HTTP 5xx).
Raised by both the legacy ``query`` path and the Water Data path, so a 5xx
surfaces as one type whichever subsystem issued the request. ``.status_code``
holds the actual 5xx; it falls back to 503 only on a bare hand-construction.
"""
_DEFAULT_STATUS = 503
# --- Request can't fit (not necessarily an HTTP status) ------------------
[docs]
class Unchunkable(RequestTooLarge):
"""No chunking plan fits the URL byte limit.
Raised by the Water Data chunker when even the smallest reducible plan
(every list axis at one atom per chunk, the filter at one clause per
chunk) still exceeds the server's byte limit. Unlike
:class:`URLTooLong`, then, automatic splitting has already been tried and
exhausted. Shrink the input lists, simplify the filter, or split the call
manually.
"""
# --- Connection failure (no HTTP response) -------------------------------
[docs]
class NetworkError(DataRetrievalError):
"""The request never completed a round-trip to the service.
A DNS failure, refused connection, or timeout stopped it, so no HTTP
response arrived to classify.
Wraps the underlying ``httpx`` transport exception, preserved on
``__cause__``. Worth retrying (:attr:`~DataRetrievalError.retryable` is
``True``), but carries no ``.status_code`` because no response came back.
"""
retryable: ClassVar[bool] = True
# --- Bad configuration ---------------------------------------------------
[docs]
class ConfigurationError(DataRetrievalError, ValueError):
"""A ``dataretrieval`` setting holds a value that can't be used.
The setting may be an environment variable or a policy field; either way,
no request was issued.
It is a :class:`DataRetrievalError` so ``except`` around a retrieval catches
it rather than letting a bare ``ValueError`` escape a request path, and a
:class:`ValueError` so code that already treats a bad setting as one keeps
working.
"""
# --- Empty result --------------------------------------------------------
[docs]
class NoSitesError(DataRetrievalError):
"""A request succeeded (HTTP 200) but matched no sites/data.
A no-data result is normally **not** an error: the modern getters
(``waterdata``, ``wqp``, ``nldi``) return an empty ``DataFrame``. Only the
deprecated ``nwis`` (waterservices) path still raises this.
"""
def __init__(self, url: httpx.URL) -> None:
self.url = url
def __str__(self) -> str:
return (
"No sites/data found using the selection criteria specified in "
f"url: {self.url}"
)
[docs]
def error_for_status(
status: int, message: str, *, retry_after: float | None = None
) -> DataRetrievalError:
"""Return the typed :class:`DataRetrievalError` for an HTTP error *status*.
The one status-to-type mapping every request path shares (the legacy
``query`` path, ``waterdata``, ``streamstats``), so a given status
becomes the same type everywhere:
* **413, 414** -> :class:`URLTooLong` (a :class:`RequestTooLarge`) -- the
"too long" semantic is more actionable than a bare status, and it matches
the client-side over-long-URL case
* **429** -> :class:`RateLimited`
* **5xx** -> :class:`ServiceUnavailable`
* **anything else** -> :class:`HTTPError`
``message`` is used verbatim; ``retry_after`` is attached only to the
transient (:class:`TransientError`) types. *status* must be an error status
(``>= 400``) -- classifying a success or redirect is a usage error and raises
:class:`ValueError`.
"""
if status < 400:
raise ValueError(
f"error_for_status expects an HTTP error status (>= 400), got {status}"
)
if status in (413, 414):
return URLTooLong(message)
if status == 429:
return RateLimited(message, status_code=status, retry_after=retry_after)
if 500 <= status < 600:
return ServiceUnavailable(message, status_code=status, retry_after=retry_after)
return HTTPError(message, status_code=status)
[docs]
def parse_retry_after(value: str | None) -> float | None:
"""Parse a ``Retry-After`` header into seconds, or ``None`` for no usable hint.
Both header forms mean the same thing and are treated the same way: the
seconds are returned as given, however large. A value past what a caller will
wait out inline stops the retry and surfaces a transient carrying the hint on
``.retry_after``, so a long wait becomes the caller's decision (and, for a
chunked call, a resumable interruption) instead of being ignored.
An over-long hint is honored rather than discarded. Dropping it would make
the client retry *harder* against a service that just asked for a long
pause, and would deny the caller the number it needs on ``.retry_after``.
Clock skew can inflate a date-form hint, but trusting one costs a
recoverable escalation while ignoring it costs hammering a service that is
already asking for room.
A date that has *already* passed yields no hint at all rather than ``0.0``.
Read literally it says "retry now", but the likelier reading is that our
clock runs ahead of the server's -- and acting on it would re-send almost
immediately against a service that just asked for a pause. Falling back to
our own bounded backoff is right under either reading. (Delta-seconds is
clock-independent, so a literal ``Retry-After: 0`` is still honored as the
instruction it is, floored by
:meth:`~dataretrieval.transport.retry.RetryPolicy.backoff`'s jitter.)
"""
if not value:
return None
raw = value.strip()
try:
seconds = float(raw)
except ValueError:
pass
else:
# ``inf``/``nan`` parse cleanly but poison every later comparison: an
# infinite hint would refuse retry forever and travel to the caller on
# ``.retry_after``. Treat them as no hint at all.
return max(0.0, seconds) if math.isfinite(seconds) else None
try:
retry_at = parsedate_to_datetime(raw)
except (TypeError, ValueError, OverflowError):
return None
if retry_at.tzinfo is None:
retry_at = retry_at.replace(tzinfo=timezone.utc)
delay = (retry_at - datetime.now(timezone.utc)).total_seconds()
return delay if delay > 0 else None