* Re-tier HTTP request timeouts around per-source memory
Replaces the single global "recent timeout" slot in `HTTPRequestTimeoutManager`
with a per-host map (`host -> lastTimeoutDate`, 10-minute per-entry expiry,
pruned on access) and re-tiers the request timeouts:
- Main-source, endpoint with no fallback-URL support: 15s base / 5s reduced
- Main-source, endpoint with fallback-URL support: 5s base / 2s reduced
- Fallback-host and proxied requests: flat 30s (never consult the memory)
A main-source request to host H uses the reduced tier iff H has a non-expired
timeout entry. Timeouts are now recorded for any non-fallback request (dropping
the previous fallback-support-only condition), keyed by attempt host; a
successful main-source response clears only that host's entry. Renames the
`timeoutOnMainBackendForFallbackSupportedEndpoint` result to `mainSourceTimedOut`.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Define flat timeout tier as a Timeout constant
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add exhaustive multi-host tests for timeout manager
Co-authored-by: Cursor <cursoragent@cursor.com>
* Honor developer networkTimeout via internal NetworkTimeout enum
Introduce an internal NetworkTimeout enum (.default/.custom) threaded from
Configuration down to HTTPRequestTimeoutManager. When a developer sets
networkTimeout, it replaces the built-in base/flat tiers while the reduced
fail-fast tiers stay fixed.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add NetworkTimeout.swift to RevenueCat.xcodeproj
The file was only picked up by the SPM/Tuist directory globs, so the
hand-maintained Xcode project build failed to resolve the type.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Gate re-tiered no-fallback timeouts behind usesRemoteConfigAPISources
The aggressive no-fallback fail-fast timeouts and their per-host timeout
recording now apply to main-API requests only when the usesRemoteConfigAPISources
dangerous setting is enabled. With it disabled (the default) the legacy network
timeout is preserved. The gating is a per-call timeout(...) parameter so
blob-source downloads can opt into the re-tiered timeouts independently.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Keep the legacy flat timeout when re-tiered timeouts are off, and share one manager
The fallback-host and proxied tiers returned the new 30s flat timeout regardless of
`usesRemoteConfigAPISources`, halving the legacy 60s (or custom) network timeout for
proxy users and every fallback-host request. They now honor `reTieredTimeoutsEnabled`
like the no-fallback tier does, so only the fallback-supported tiers behave the same
either way.
`Backend` also built one timeout manager per HTTPClient, so the main and remote-config
lanes kept independent per-host memories despite talking to the same hosts. They now
share one, and `HTTPClient` requires both the manager and the network timeout so this
wiring can't be dropped silently.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Clear a host's fail-fast memory as soon as it answers
The memory exists to remember how quickly a host responds, so a non-error
response is evidence the host recovered even when decoding or verifying it
fails afterwards. Deciding off the raw status code instead of the parsed result
keeps those failures from leaving a responsive host stuck on the reduced
timeout tier.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Name the requestTimeoutManager in the fail-fast comment
Co-authored-by: Cursor <cursoragent@cursor.com>
* Tighten comments on the Backend-shared failover and timeout manager
Name the instances the comments describe so it's clear what is shared.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Record API source timeouts before health checks
* Test shared timeout state across backend lanes
* Normalize URL timeout classification across platforms
* Remove unused RemoteConfigAPI test hook
* Skip HTTP stubs lane tests on watchOS
* Treat the remote config fallback path as a fallback host request
`FallbackPath` bakes the fallback host into its `serverHostURL`, but `fallbackUrlIndex` is
nil on those requests, so `isFallbackURLRequest` was false and they were classified as
main-source: with API sources enabled they drew the aggressive 15s/5s no-fallback tiers
instead of the flat fallback timeout, their timeouts armed the per-host fail-fast memory
for the fallback host, and their successes cleared the main source's entry.
Add `isFallbackHostPath` so a path can declare that it targets a fallback host on its own,
and use it alongside the index when picking the tier and updating the memory.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Rick van der Linden <rick.vanderlinden@revenuecat.com>
* Implemented initial HTTPRequestTimeoutManager + added first couple of tests
* Cleaned up the code and added a bunch more unit tests based on the Android test suite
* Updated doc format
* Add new files to workspace
* Added HTTPClient tests testing interaction with timeout manager
* Swiftlint fixes
* Updated comment
* Add missing file to workspace
* Fix after rebase
* Added extra HTTPClient <-> HTTPRequestTimeoutManager tests asserting on the recorded events
* Implemented PR feedback
* Fixed compile issue
* Fix compile issue
* Implemented PR feedback
* Removed deleted file reference from workspace
* Fixed test after changing reset timeout
* Make sure the default HTTPRequistTimeoutManager is using the default value from the HTTPClient
* Add a sophisticated test testing the full timeout flow against an unroutable IP for a real world scenario
* Updated timeout case name
* Fixed case where .timeoutOnMainBackendForFallbackSupportedEndpoint was also recorded for non timeouts and added tests to prevent regression
* Fixed only stubbing the HTTP response based on path, since that could be the same across hosts, resulting in flaky tests
