The failure logger skips fallback hops (has_logged_async_failure is already set), so
model_call_details.end_time still belongs to the previous hop and predates this hop's
api_call_start_time. The fallback cooldown guard measured a negative elapsed time and
cooled down deployments for caller-set timeouts.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
client_side_timeout records that the caller configured a timeout, not that
the timeout fired. A 408 the provider returns before that deadline is a
deployment failure and must still count toward cooldown.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
A 408 produced by a timeout the caller set (a timeout body field or an
x-litellm-timeout header, which the proxy marks as client_side_timeout)
says nothing about the deployment's health, yet the router's primary
failure callback counted it toward allowed_fails and cooled the
deployment down. The fallback path already skipped it.
The marker never reached that callback because get_litellm_params drops
kwargs outside OPTIONAL_KWARGS_KEYS, so it is listed there now, and
deployment_callback_on_failure returns before the failure counter when
is_caller_timeout_408 holds. A 408 from a timeout the deployment or the
provider set still counts and still cools the deployment down.
A stale RoutingPrismaWrapper.writer_unavailable flag made WriterPinnedClient hand back the routed wrapper, where query_raw is classified as a read, so the access-group UPDATE statements behind /key/regenerate, /key/generate with access_group_ids and model rename/delete went to the read replica and failed with SQLSTATE 25006. Route those raw statements through the underlying writer regardless of the flag; a raw SQL write has no replica fallback. WriterPinnedClient keeps yielding to the replica for degraded reads.
The model sync's backing-row count stays on the writer too: it runs right after the row delete/update on the writer and a lagging replica could still report the removed row, which would leave the group naming a model nobody serves.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
_organization_max_budget_check returned early whenever org_max_budget
was <= 0, so an organization with an explicit max_budget of 0 was
treated as unlimited instead of zero allowance. Key, team, and user
budget checks already skip only on None; align organization budgets
with that convention.
validate_team_org_change had the same defect in a different shape: it
used a truthy check on the org's max_budget when validating a team
move, so an explicit 0 there silently skipped the guard too.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
POST /credentials let a duplicate name hit the unique index and handed back
Prisma's "Unique constraint failed" as a 500, so callers string-matched that
message to tell a caller mistake from a server fault. The unique violation now
maps to a 409 whose message names the PATCH route, two concurrent creates of
one name agree on it, and the detection lives in a repository helper the five
hand-rolled copies can move onto later
PATCH /credentials/{name} took a CredentialItem body, so the model_id the
Terraform adopt path sent was dropped. It now accepts UpdateCredentialItem and
shares the deployment lookup with create. Both handlers take the router as a
FastAPI dependency instead of reading the proxy global, which is what the
tests override
