fix: Unify the request pipeline and retry policy across HTTP clients by vdusek · Pull Request #1006 · apify/apify-client-python

vdusek · 2026-08-12T09:09:35Z

Moves call, the retry loop, and the per-attempt request handling from ImpitHttpClient into HttpClient and HttpClientAsync, leaving Impit as a thin adapter over the transport hooks: send_request, is_retryable_transport_error, is_timeout_error, and close() / aclose() plus context managers. A custom client that overrides call keeps working unchanged and opts out of the shared pipeline.

The same pass fixes:

Transport errors a retry cannot fix (UnsupportedProtocol, LocalProtocolError, TooManyRedirects) fail on the first attempt instead of burning the whole backoff. Everything else in the impit.HTTPError tree stays retryable, including a bare HTTPError from a body that ends mid-chunk and a ProxyError from a proxy answering CONNECT with a transient status.
A failure while reading a streamed error response is classified and retried like a send failure, and the response is closed instead of being left open.
HttpResponse comes from typing_extensions, whose runtime protocol check looks attributes up statically, so isinstance(response, HttpResponse) no longer buffers or rejects an unread streaming body.
try_import only substitutes failed imports for the optional dependency it names, so an ImportError from the importing module or another dependency is no longer masked, and __all__ omits what a missing extra never provided.

Split out of #1004, which adds the built-in HTTPX client on top of this contract.

✍️ Drafted by Claude Code

Move `call`, the retry loop, and the per-attempt request handling from `ImpitHttpClient` into `HttpClient` and `HttpClientAsync`, leaving Impit as a thin adapter over the `send_request`, error-classification, and lifecycle hooks. A custom client that overrides `call` keeps working unchanged and opts out of the shared pipeline. The same pass fixes: - Transport errors a retry cannot fix (`UnsupportedProtocol`, `LocalProtocolError`, `TooManyRedirects`) fail on the first attempt instead of burning the whole backoff. Every other error in the `impit.HTTPError` tree stays retryable, including a bare `HTTPError` from a body that ends mid-chunk and a `ProxyError` from a proxy answering CONNECT with a transient status. - A failure while reading a streamed error response is classified and retried like a send failure, and the response is closed instead of being left open. - `HttpResponse` comes from `typing_extensions`, whose runtime protocol check looks attributes up statically, so `isinstance(response, HttpResponse)` no longer buffers or rejects an unread streaming body. - `try_import` only substitutes failed imports for the optional dependency it names, so an `ImportError` from the importing module or another dependency is no longer masked, and `__all__` omits what a missing extra never provided.

codecov · 2026-08-12T09:11:30Z

Codecov Report

❌ Patch coverage is 97.51244% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 94.91%. Comparing base (4281530) to head (3d4e037).
⚠️ Report is 2 commits behind head on master.

Files with missing lines	Patch %	Lines
src/apify_client/http_clients/_base.py	97.90%	3 Missing ⚠️
src/apify_client/_streamed_log.py	75.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1006      +/-   ##
==========================================
+ Coverage   94.86%   94.91%   +0.04%     
==========================================
  Files          58       58              
  Lines        5357     5406      +49     
==========================================
+ Hits         5082     5131      +49     
  Misses        275      275

Flag	Coverage Δ
integration	`91.62% <68.65%> (-0.66%)`	⬇️
unit	`85.27% <97.51%> (+0.63%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

…rings

…dencies (#1009) `try_import` now swallows only the `ModuleNotFoundError` of the optional dependency it names, so an import error raised by the importing module itself or by another dependency propagates instead of being masked. `__all__` of `apify_client.http_compressors` omits `BrotliHttpCompressor` when the `brotli` extra is not installed, since the symbol was never provided. Split out of #1006. *✍️ Drafted by Claude Code*

…onses (#1010) `HttpResponse` now comes from `typing_extensions`, whose runtime protocol check looks attributes up statically. The `typing` implementation on Python 3.11 calls `hasattr`, which evaluates properties, so `isinstance(response, HttpResponse)` on an unread streaming response either raised or silently buffered the whole body. Adds streaming regression tests (unit and integration) and a parametrized built-in-client fixture they run on. Split out of #1006. Stacked on #1009. *✍️ Drafted by Claude Code*

vdusek · 2026-08-18T07:20:34Z

This will be split into several smaller PRs

`_utils/errors.py::is_retryable_error` is a stale copy of `http_clients/_impit.py::_is_retryable_error`. No production code calls it — only its own tests did — and it sits in a private module, so nothing depends on it. Once #1019 narrows the real policy, keeping a second copy that still claims every `impit.HTTPError` is transient is actively misleading. Removes the function, its two test cases, and the imports they alone needed. Split out of #1006. *✍️ Drafted by Claude Code*

Every `impit.HTTPError` counted as transient, so a transport failure a retry cannot fix burned the whole backoff before surfacing. The permanent classes now fail on the first attempt instead: - `LocalProtocolError` - a request Impit rejects before sending it, e.g. one carrying an invalid header value. - `TooManyRedirects` - a routing loop, which repeating the request cannot break. - `UnsupportedProtocol` - the class Impit declares for a scheme it refuses to speak, but does not raise today; an unsupported scheme arrives as `impit.InvalidURL`, outside the `impit.HTTPError` tree and non-retryable already. Listed so the classifier stays right if Impit switches over. - `HTTPStatusError` - only `Response.raise_for_status()` raises it and the client never calls it; `_make_request` decides on status codes from the response itself. Everything else in the `impit.HTTPError` tree stays retryable, including a bare `HTTPError` - Impit wraps a failure its internal HTTP library did not classify in one, e.g. a non-HTTP response or a connection reset - and a `ProxyError`, which covers a rejected CONNECT tunnel and a 407 alike, so a transient case cannot be told from a permanent one. Also drops `InvalidResponseBodyError` from the retry check. Only `key_value_store.py` raises it, once `call` has already returned, so it can never reach the classifier. Its docstring no longer claims the client retries such requests. The duplicate `is_retryable_error` in `_utils/errors.py`, which carried the old policy, is gone. Split out of #1006 and re-targeted at `master`, so the retry fixes can ship in a patch release ahead of the pipeline refactor in #1011. *✍️ Drafted by Claude Code*

…1020) Before raising `ApifyApiError`, `_make_request` buffers the response so a streamed error body reaches the exception. That read sat outside the `try` block that classifies an attempt, so a failure there escaped the retry policy entirely: a non-retryable failure was retried through the whole backoff, and the response was left open. The read now goes through the same classification as a failed send — retried when it is a transient transport failure, surfaced immediately when it is not — and the response is closed before the exception propagates. Split out of #1006 and re-targeted at `master`, so the retry fixes can ship in a patch release ahead of the pipeline refactor in #1011. *✍️ Drafted by Claude Code*

vdusek added adhoc Ad-hoc unplanned task added during the sprint. t-tooling Issues with this label are in the ownership of the tooling team. labels Aug 12, 2026

vdusek self-assigned this Aug 12, 2026

github-actions Bot added this to the 147th sprint - Tooling team milestone Aug 12, 2026

github-actions Bot added the tested Temporary label used only programatically for some analytics. label Aug 12, 2026

vdusek added 2 commits August 12, 2026 14:26

docs: Keep HttpClientBase in the API reference

5c06d0e

refactor: Order HTTP client members by convention and fix their docst…

3d4e037

…rings

vdusek closed this Aug 18, 2026

This was referenced Aug 18, 2026

fix: Fail fast on transport errors a retry cannot fix #1019

Merged

fix: Classify and retry failures while reading a streamed error body #1020

Merged

refactor: Remove the unused retry classification helper #1021

Merged

vdusek mentioned this pull request Aug 18, 2026

feat: Unify the request pipeline across HTTP clients #1022

Draft

Navigation Menu

Uh oh!

fix: Unify the request pipeline and retry policy across HTTP clients - #1006

fix: Unify the request pipeline and retry policy across HTTP clients#1006
vdusek wants to merge 3 commits into
masterfrom
unify-http-client-pipeline

vdusek commented Aug 12, 2026

Uh oh!

codecov Bot commented Aug 12, 2026 •

edited

Loading

Uh oh!

vdusek commented Aug 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

vdusek commented Aug 12, 2026

Uh oh!

codecov Bot commented Aug 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

vdusek commented Aug 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

codecov Bot commented Aug 12, 2026 •

edited

Loading