Register s3_access_key_id, s3_secret_access_key and s3_encryption_key_id as
LiteLLM-owned batch params so they are no longer forwarded to Bedrock as
additionalModelRequestFields (which 400s ordinary chat on a batch-configured
deployment), keep them on CredentialLiteLLMParams so the batch/file paths
still receive them, and redact the S3 credential key names in debug logs.
Resolves LIT-8290
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Register s3_access_key_id, s3_secret_access_key and s3_encryption_key_id as
LiteLLM-owned batch params so they are no longer forwarded to Bedrock as
additionalModelRequestFields (which 400s ordinary chat on a batch-configured
deployment), keep them on CredentialLiteLLMParams so the batch/file paths
still receive them, and redact the S3 credential key names in debug logs.
Resolves LIT-8290
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Three holes in the new suite, all of which let a test pass without proving
what its name claims:
- A migration recorded twice, once per replica, each with
applied_steps_count = 1, slipped past both the step-count check and
migration_names(), which collapses the history into a set. Reject
duplicate migration_name rows outright.
- auth_traffic only asserted the failures it had seen by the time
keep_serving hit its target. A request failing after that, or on the
other replica while the test waited on one stream, was recorded and
never read. Assert the recorded failures once the thread has joined.
- The rolling test warmed the baseline replica's virtual-key cache before
the upgrade, and that cache holds for 60 seconds by default
(UserAPIKeyCacheTTLEnum.in_memory_cache_ttl). The candidate migrates
well inside that window, so the post-upgrade requests could be served
from cache without ever repeating the whole-row token lookup that the
stale prepared statement breaks. Drive the baseline replica with a key
minted after the schema moved, which it has never seen and must resolve
from the database.
Re-ran against v1.101.0 -> v1.102.0: 6 passed.
The migration e2e harness only ever used one image: it seeded the database
with the candidate build and then applied synthetic migrations on top. That
proves the migration machinery (locking, crash recovery, legacy baselining,
pooling) but never executes the real schema of release N against the real
migrations of release N+1, which is the path operators actually run.
Adds a baseline image alongside the candidate, so a test can seed with a
published release and upgrade with the build under test.
Suites:
- test_upgrade.py: the candidate applies the pending release migrations,
keys minted by the baseline release survive, and concurrent replicas
upgrade a baseline database exactly once.
- test_rolling_upgrade.py: a baseline replica keeps serving virtual-key
auth while the candidate migrates underneath it, and both releases serve
and resolve each other's keys during the overlap. This is the reported
failure: a new column on LiteLLM_VerificationToken invalidates prepared
plans on pods still running the old release, which the proxy reads
whole-row, and auth starts failing until those pods leave service.
- test_shaped_database.py: the upgrade completes and preserves rows on a
populated spend log, rather than on the empty database every other
migration test starts from.
Every upgrade assertion is gated on the candidate having actually applied
migrations the baseline had not, so a stale pin fails loudly instead of
passing on an empty delta.
CI adds two jobs to the migration_startup workflow. The baseline defaults
to a committed release pin and is overridable per pipeline, matching how
migration_candidate_image already works; only the upgrade jobs pull it.
Verified against a real v1.101.0 -> v1.102.0 upgrade: 6 passed, with the
baseline seeding 165 migrations and the candidate applying the 6 that
landed between the two releases.
The deleted wrapper put cache: {"no-cache": true} on every request body, so the
gateway's response cache never answered a re-sent prompt. The SDKs send nothing
of the sort, and the mid-conversation prompt-cache priming loop re-sends an
identical body until the provider reports a warm cache, which a cached reply
never does. NO_PROXY_CACHE in sdk_clients.py restores the field as extra_body
on every messages, responses, completions, and embeddings call.
The wrapper also returned a 4xx as a value where the SDKs raise. The Bedrock
safety_identifier test judges the captured Converse body, and Claude on Bedrock
rejects the forwarded field with a 400, so the /v1/responses leg now suppresses
openai.BadRequestError the way the chat leg carries the same 400 as a Result.
Resolves the six conflicts main accumulated against the SDK migration.
endpoints_client.py stays deleted (main only added the safety_identifier
kwarg to it), and its two new importers move onto the SDK clients: the
Bedrock safety_identifier forwarding test in test_responses_e2e.py now
sends its /v1/responses leg through the OpenAI SDK and its chat leg
through ProxyClient.chat, and test_messages_azure_foundry_e2e.py drops
the x-litellm-rust header check together with EXPECT_RUST, which main
removed from e2e_config in a84f68b6e3. proxy_client.py keeps both the
rerank and router-settings model imports. uv.lock is main's lock plus
the anthropic e2e-dev entry, and the pricing JSONs are taken verbatim
from main since this PR never meant to change them.
The Buildkite ephemeral stack runs the gateway in another pod, so it cannot reach the pytest host's provider edge. The GitHub changed-e2e lane runs gateways on the runner and sets E2E_PROVIDER_EDGE_HOST_REACHABLE
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The changed-test gate booted its stage-mirror stack without files_settings
or finetune_settings, so every raw upload with a custom_llm_provider hit a
500, and it exported the whole provider env into the gateways, so the
AWS_ROLE_NAME the assume-role test needs made the GovCloud deployment run
an AssumeRole with its static keys. The gate also deleted its pytest output,
so a red run left nothing to read. The mirror config now carries the
openai, azure, and vertex_ai file settings, gateways start without
AWS_ROLE_NAME, and the workflow uploads the pass logs and junit files with
every secret value, every field of a JSON-valued secret, and their
XML-escaped forms replaced before the raw files are removed.
AGENTS.md allows a vendor-owned literal only when its source and date are cited
next to it. A live realtime test cannot avoid naming a model, so record how the
id was checked, and record that a retired id fails as a hang rather than an
error so the next reader does not start by suspecting litellm.
Use contextlib.suppress for the cancelled task in stop_view_setup_task, make the legacy prisma setup test inject a plain mock for the synchronous start_view_setup_task and assert it is called, and drop the docstrings the branch added to tests
AWS retired amazon.nova-sonic-v1:0. GetFoundationModel now answers
ResourceNotFoundException "This model version has reached the end of its life",
and opening a bidirectional stream against it fails with ValidationException
"The provided model identifier is invalid". amazon.nova-2-sonic-v1:0 is the
active replacement.
The test passed on builds 219 (2026-09-16) and 246 (2026-09-17) and has failed
every run since, three attempts per build, with no litellm change to the
realtime path in between. The symptom was a clean websocket close: Bedrock ends
the stream rather than erroring, the forwarder treats a None receive as a normal
stream end and closes the client socket, so the client sees ConnectionClosedOK
and the test fails waiting for response.done.
litellm already carries both models in the cost map, with
"deprecation_date": "2026-09-14" on the old one, and the promptStart
transformation already sends the audioOutputConfiguration that nova-2-sonic
requires; only the test constant was left behind.
Verified against live Bedrock with the promptStart shape the transformation
builds: amazon.nova-sonic-v1:0 raises "The provided model identifier is
invalid", amazon.nova-2-sonic-v1:0 opens a session and returns a usageEvent.
The mocked handler and provider-cache tests keep the old id: it is only a label
there, no call reaches AWS.
The proxy used to boot with no master key (every request accepted without
authentication) and with sk-1234, the key every example used. It now stops at
startup, before it connects to the database, and prints how to fix it: where the
bad key came from, a copy-pastable command that generates a secure key, and,
when the public key is also encrypting a database, a link to the rotation guide
general_settings.dangerously_allow_unsafe_proxy: true or
LITELLM_DANGEROUSLY_ALLOW_UNSAFE_PROXY=true starts the proxy anyway, for local
development. CI and test boots that rely on sk-1234 or on no key set it
BREAKING CHANGE: deployments with no master key, an empty one, or sk-1234 no
longer start until they set a real key or opt in to the override
Two tests in the config/misc management suite were failing every run against
the Buildkite e2e stack, and one of them took the rest of the build with it.
test_add_allowed_ip_does_not_store_unrelated_config_value posted 127.0.0.1 to
/add/allowed_ip. That route sets the live general_settings["allowed_ips"] that
auth_utils._check_valid_ip reads before it persists anything, and the check is
exact string membership with no CIDR support, so from the moment the POST
returns only 127.0.0.1 can reach the proxy. The runner 403s on its very next
call, and the deferred /delete/allowed_ip sits behind the same auth dependency,
so the cleanup is locked out too and every later test in the build 403s. Build
254's first attempt lost 459 of its 465 failures to that one cascade.
There is no safe way to exercise the route against a shared proxy: nothing
reports the caller's address as the proxy sees it, so a test cannot allowlist
itself first. Move the claim to the route's own TestClient suite, where the
auth dependency is overridden and general_settings is per-test, and record the
route in the module docstring beside /cache/settings and the Vault override so
it is not re-added. save_config's end of the contract was already covered by
test_ProxyConfig_save_config_merges_changed_keys_without_copying_file_settings;
the new test covers the route's end, that what it hands save_config differs
from the loaded config in allowed_ips and nothing else.
The unrelated-key probe also only ever worked on one lane: max_parallel_requests
was added to tests/e2e/gateway/stage_mirror_ci_config.yml and never to the
Buildkite stack's config, where resolve() reports it as "unset" rather than
"config". That key is now unused, so drop it again.
test_config_update_persists_router_setting_to_get wrote router_settings.
num_retries, which both lanes declare in their config file, so the config-
ownership work correctly refuses it with a 400. Switch to retry_after, which is
declared by neither lane, is accepted by /config/update, and is reported back by
GET /router/settings. Verified against a live proxy: max_fallbacks also takes
the write but never reads back, so the read-back poll is what picks the key.
On a fresh database where the migrations run in a separate job while the
proxy boots with DISABLE_SCHEMA_UPDATE=true, the startup view check ran
as a fire-and-forget task, used up its three 10 second retries before
LiteLLM_SpendLogs existed, and died with an unretrieved exception. The
spend views were never created, so the /global/spend routes returned 500
until the pod was restarted
PrismaClient now holds a view setup task. It polls to_regclass for the
spend-log table every 5 seconds, creates the views and loads the spend
log row count once the table is there, keeps polling if an attempt raises
while the schema is still settling, and logs an ERROR with the last
failure if nothing worked after 15 minutes. Proxy shutdown cancels the
task
The spend route e2e tests for the five view-backed routes are no longer
skipped and wait for the views through the harness convergence helper
vLLM serves no /v1/files or /v1/batches, so a hosted_vllm deployment can never
host a batch. Batch inputs for such a deployment now land in a LiteLLM-owned
storage backend, the batch is executed line by line through the deployment's
own chat, completion, embedding, or responses route, and the batch plus its
output and error files are served back from the database under the creating key
