mirror of
https://github.com/mudler/LocalAI.git
synced 2026-09-14 23:28:23 -04:00
Richard Palethorpe
7768fab1f9
Assisted-by: Codex:gpt-5 Signed-off-by: Richard Palethorpe <io@richiejp.com>
78 lines
3.8 KiB
Markdown
78 lines
3.8 KiB
Markdown
---
|
|
title: "Offline test resources"
|
|
---
|
|
|
|
LocalAI tests separate resource acquisition from test execution. Resources are
|
|
declared by resource set in `test-resources/manifests/`; files and packed container
|
|
images are content-addressed by SHA-256 under
|
|
`.cache/test-resources/blobs/sha256/`.
|
|
|
|
Prepare the resources before running a target:
|
|
|
|
```sh
|
|
make prepare-offline-test-cache TEST_RESOURCE_SET=default
|
|
```
|
|
|
|
Preparation verifies every cached blob and fails closed. It never substitutes
|
|
a live request for a missing or corrupt entry. Maintainers can populate a
|
|
cache from pinned declarations only by explicitly enabling online mode:
|
|
|
|
```sh
|
|
LOCALAI_TEST_RESOURCES_ONLINE=1 make update-offline-test-cache TEST_RESOURCE_SET=default
|
|
```
|
|
|
|
The update command records declared responses, files, and digest-pinned images,
|
|
then writes a deterministic, zstd level-1 bundle at
|
|
`.cache/test-resources/bundles/<resource-set>.tar.zst`. Its SHA-256 is written to the
|
|
lock file. The test workflow transfers the bundle as a workflow artifact and
|
|
verifies it after deleting the recording cache; the scheduled refresh workflow
|
|
also publishes verified bundles to GHCR as OCI artifacts.
|
|
|
|
HTTP declarations may include `request_headers`. `Range` participates in the
|
|
cache key, and authorization values participate only through a SHA-256 value;
|
|
credentials are never written verbatim to the cache index. Redirect responses
|
|
are recorded without following them, so every hop needed by a test must be
|
|
declared explicitly.
|
|
|
|
File and HTTP declarations may list HTTPS `mirrors`. Recording tries the
|
|
canonical URL twice, then each mirror twice, and reports the duration of every
|
|
attempt. Every candidate must produce the same declared SHA-256; mirrors are
|
|
alternate transports, not alternate content.
|
|
|
|
A digest mismatch is never accepted automatically. The updater prints the
|
|
observed failure for every source and directs maintainers to compare upstream
|
|
checksums, signatures, release notes, and redirects, then check the GitHub
|
|
Advisory Database and OSV before approving a new digest. Repeated mismatches can
|
|
mean a legitimate upstream release, a corrupt mirror, or a supply-chain event.
|
|
|
|
Ordinary test recipes execute through `scripts/run-test-offline.sh`. Its
|
|
supervised replay proxy terminates HTTP and HTTPS and returns an immediate
|
|
error containing the method and URL for undeclared requests. Linux CI also
|
|
runs the command in a cgroup with public IPv4 and IPv6 rejected; macOS relies
|
|
on replay, declared resources, guarded Go transports, and static lint because
|
|
kernel-level subprocess enforcement is Linux-only.
|
|
|
|
Testcontainer images must be registry-digest pinned and loaded during
|
|
preparation. Container helpers check that an image exists before startup and
|
|
attach services to internal-only Docker networks, preventing testcontainers
|
|
from silently pulling a missing tag.
|
|
|
|
The default Linux and macOS suites use separate resource sets because
|
|
Docker archives are platform-specific. Backend and hardware resources remain
|
|
separate targets so ordinary contributors do not acquire large model fixtures
|
|
that their test command does not use.
|
|
|
|
Coverage runs print a wall-clock summary for each test root and list every
|
|
Ginkgo spec or hook taking at least three seconds, including its source
|
|
location. Set `COVERAGE_SLOW_SPEC_THRESHOLD=<seconds>` to tune the reporting
|
|
threshold. This measures the whole spec or hook, so it exposes time spent in
|
|
sleeps, polling, channel waits, cleanup, and resource contention without
|
|
replacing Go's global clock or changing test semantics. The same timings are
|
|
written to `coverage/timings.tsv` for CI artifacts and comparisons. The report
|
|
shows the slowest 25 entries per root by default; set
|
|
`COVERAGE_SLOW_SPEC_LIMIT=<count>` to change the cap.
|
|
|
|
Real third-party compatibility checks belong in separately named
|
|
`external-probe-*` scheduled workflows and must not be part of deterministic
|
|
test or coverage gates.
|