Three defects kept LiteLLM's OTel metrics from reaching an OTLP backend.
OTEL_EXPORTER_OTLP_HEADERS is W3C Baggage encoded per the OTLP spec, so its
values are percent-encoded. litellm split the string on "," and "=" and passed
the raw value straight to the exporter, so a vendor that documents
"Authorization=Basic%20<token>" got a literal "%20" on the wire and the backend
rejected the credential. Grafana Cloud documents exactly that shape, which made
its OTLP gateway unreachable. Header parsing now delegates to the OTel SDK's own
W3C Baggage parser in liberal mode, so percent-encoded values decode and values
that were never encoded keep working. It moves from model/utils.py to
plumbing/providers.py because model/ is deliberately free of opentelemetry
imports; providers.parse_headers was already the entry point every caller used.
The OTLP metric exporters then overrode histogram temporality to delta.
Prometheus and Mimir, which back Grafana Cloud's OTLP gateway, reject delta
histograms outright: the gateway answers 400 "invalid temporality and type
combination" and drops the entire batch, so every GenAI metric was silently lost
while traces kept flowing. Backends that prefer delta still accept cumulative, so
the SDK default is the compatible choice in both directions, and the enterprise
billing exporter already relies on it.
Three GenAI instruments also carried names no convention or backend defines, so
nothing downstream could chart them. Time to first token and time per output
token take their semconv names, gen_ai.server.time_to_first_token and
gen_ai.server.time_per_output_token; the gen_ai.client.response.* spellings
litellm used are not conventions at all. Cost has no semconv instrument, so it
takes gen_ai.usage.cost, the name backends already query for spend. All three are
listed verbatim in Grafana Cloud's AI Observability integration reference, so its
prebuilt panels find them. Both engines now read the names from the shared Metric
constants rather than repeating string literals, so v1 and v2 cannot drift.
The renames are breaking for anyone charting the former names; the docs and the
release changelog carry the migration note.
* feat(ui): shareable log links via log_id query param on the logs page
Clicking a log row now writes ?log_id=<request_id> to the URL, closing the
drawer removes it, and loading the logs page with ?log_id= opens the drawer
for that log. When the log is not in the loaded page, it is fetched by
request_id (the backend already drops the date window for id lookups), so
links keep working for logs of any age. Drawer open state derives from the
URL, mirroring the models page ?model= pattern.
* fix(ui): close the log drawer on browser back after opening via session id
Session opens now write ?session_id= to the URL instead of holding local
state, so back removes both params and the drawer closes (Greptile P1).
Session views become shareable links as a side effect. In-drawer log
switching now replaces the history entry instead of pushing, so back
always closes the drawer in one step rather than replaying every viewed
log.
* fix(proxy): scope /spend/logs/session/ui to the requesting user's visible logs
Non-admin callers now only receive session rows they could already see on
/spend/logs/ui: their own logs plus logs of teams where they hold the
spend-logs permission. Previously any authenticated user could read any
session's log metadata by id, which shareable ?session_id= links made
trivial to trigger. Admin views are unchanged. Also, clicking a log row
now clears a lingering ?session_id= from the URL so the drawer shows the
clicked log instead of a stale session (Greptile P1).
* feat(ui): chart failed requests as their own series on the cache dashboard
Spend logs for failed requests are stored with an empty call_type, so the
Cache Hits vs API Requests chart lumped them into an Unknown bar that read
as normal LLM API traffic. The activity query now also returns a per-group
failed_rows count (status = 'failure') and the dashboard charts it as a
third stacked series, so failures are visibly separate from successful
requests and cache hits. The chart data transform moves into a pure
summarizeCacheActivity helper with unit tests; header stats keep their
existing semantics (cache hit ratio still counts failures in the
denominator).
* refactor(ui): move cache dashboard aggregation server-side with a typed response
The /global/activity/cache_hits endpoint previously returned raw per
(key, call_type, model) spend-log aggregates typed as LiteLLM_SpendLogs
(wrong), and the dashboard reduced them in the browser: grouping by
call_type, relabeling empty call_type as Unknown, and computing the stat
card totals. All of that now happens server-side. The SQL groups per
call_type and splits cache hits vs successful vs failed requests, a new
cache_activity module validates rows into Pydantic models and computes
totals plus the key-alias/model filter options, and the endpoint declares
a real response_model so schema.d.ts types it correctly. The dashboard
consumes it through a typed $api react-query hook (filters ride the
query key and are applied in SQL instead of the browser), the hand-rolled
summarizeCacheActivity transform and the adminGlobalCacheActivity fetch
helper are deleted, and the refresh button now actually refetches.
The endpoint is UI-internal (hidden from the public swagger), so the
response reshape is not a public API break.
* fix(scim): stop provisioning nested group ids as internal users
POST/PUT/PATCH /scim/v2/Groups treated every member.value as a user id, so
with the default scim_upsert_user=true an unknown id was auto-created as an
internal user. Entra sends nested groups as members carrying "type": "Group",
which meant every nested group produced a phantom internal user whose id and
email were the group GUID, and those users counted toward licensed seats.
Group members are now classified before they are used: members typed "Group"
are skipped without a database hit, an id that names an existing team is
skipped too (Okta sends untyped ids through filtered paths, so the type alone
is not enough), and only ids that resolve to a user, or that resolve to
nothing at all, keep today's behavior. The user lookup runs before the team
lookup so a user whose id collides with a team id keeps syncing.
The type was previously dropped at parse time on POST/PUT because SCIMMember
had no such field, and on PATCH because the raw member dicts were reduced to
bare ids; both paths now share one resolver and one parser that preserves it.
Member removals no longer upsert: a remove of an id we do not know is an
idempotent no-op rather than a reason to create a user and immediately drop
it, and strict mode (scim_upsert_user=false) no longer rejects it. Removal of
an id that is on the roster but has no user row still cleans up membership.
Responses now state members are of type "User" instead of emitting a null,
and the advertised Group schema documents the members.type sub-attribute.
* fix(scim): harden group member classification after adversarial review
Removals now bypass classification and drop exactly the ids they name,
restoring cleanup of roster entries the old bug left behind. The
team-id fallback only applies to untyped members, so an explicit User
type always provisions even when the id collides with a team. Member
types are normalized before matching; a type other than User or Group
only skips when the id is not an existing user. Non-string type values
are tolerated as absent on every verb instead of failing validation.
Admitted member ids are deduped order-preserving, which also closes a
pre-existing duplicate-row hazard on group creation.
* fix(scim): only treat scim-managed teams as nested groups
A PR reviewer flagged that an untyped SCIM member whose id collides
with an admin-created team was silently skipped, suppressing that
user's provisioning. SCIM group writes (POST, PUT, and every PATCH)
now stamp the team with scim_managed metadata, and the typeless
team-id skip only applies to teams carrying that marker or the
scim_data blob older PUTs already wrote. Admin-created teams stay
unmarked, so a colliding untyped member provisions the user in
permissive mode and returns the standard unknown-user 400 in strict
mode. Teams SCIM touched before this change adopt the marker on their
next group write.
A live Vertex batch run showed the documented "embed_content_config" sibling of "request" is rejected by the API ("unsupported type"), failing the whole job rather than the row; the same fields inside the EmbedContentRequest succeed and honor output_dimensionality. Real output rows also report usage under response.usageMetadata.promptTokenCount, not the documented response.tokenCount, so every row came back with zero tokens.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Vertex batch files sent every jsonl line through the generateContent transform, so embeddings rows went out as {"request": {"contents": [...]}} and Vertex rejected each one with "no such field: 'contents'"; the OpenAI "input" was dropped along the way too. Route lines by their own url: embeddings lines now emit the EmbedContentRequest shape (singular content, embed_content_config sibling, custom_id round-tripping through the top-level key), and matching output rows come back as OpenAI embeddings responses.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Add regression tests for the db-fetch paths whose converted construction
lines were uncovered: the auth_checks getters (default end user budget, end
user, team membership, access group, team by alias, org by alias, object
permission, managed vector stores, project), get_all_team_memberships and
list_available_teams in team_endpoints, and the proxy admin user info
helper. Each test feeds a mocked prisma row through the real function and
asserts the validated model's fields, so a bad model_validate conversion on
any of these paths now fails a test instead of only dropping coverage.
Convert pydantic table-model construction from Cls(**row.model_dump())
kwargs-unpacking to Cls.model_validate(...) across the management endpoint
hotspot files (team, key, internal user, scim, model management, spend
tracking, auth checks, proxy_server). Unpacking an untyped dict reports one
Any-typed argument per matched model field, so each converted site clears
10-35 diagnostics while running the exact same pydantic validation.
Conversions were limited to models verified to use pydantic's default
__init__; UserAPIKeyAuth and LiteLLM_VerificationTokenView keep their custom
kwargs-rewriting __init__ and are untouched. Two locally-verified helper
params move from Any to object.
Whole-tree basedpyright, measured against the branch point in the same
environment: reportAny 24,431 -> 22,741 (-1,690), reportArgumentType
2,189 -> 2,136 (-53), reportUnknownArgumentType 34,370 -> 34,067 (-303),
reportExplicitAny 7,285 -> 7,283 (-2); total 154,882 -> 152,834 (-2,048)
with no rule increasing anywhere and no per-file increases. No casts, no
suppressions, no behavior changes. Budgets ratcheted: basedpyright -2,048
across 4 rules, ruff ANN401 -2.
* fix(mcp): resolve call_tool by registry without requiring tool map
Multi-worker reloads put MCP servers in the registry from the DB but do
not re-run tools/list on every process. Gating call_tool on
tool_name_to_mcp_server_name_mapping made cold workers 500 with Tool not
found after another worker had already listed the tool. Treat a registry
match on server id/name/alias as enough; upstream rejects unknown tools
* test(e2e): poll MCP register, tools/list, and tools/call across multi-worker lag
Stage multi-worker gateways only load MCP servers and tool maps on the
process that handled the request. Poll until the server is listed, the
tool appears on tools/list, and tools/call is not a cold-worker 500 so
key-access and Datadog MCP e2e stop racing the LB
* Revert "fix(mcp): resolve call_tool by registry without requiring tool map"
This reverts commit 8b56e51e39.
* test(e2e): tighten MCP multi-worker lag classifier
Only retry tools/call on gateway shapes Tool <name> not found and
server_not_found, not any 500 that mentions tool/server not found, so
upstream failures are not retried until the poll deadline
* test(e2e): drop unit file for MCP lag classifier
The live await_call_tool polls already cover multi-worker lag; a separate
string-match unit module is not worth keeping
(cherry picked from commit c274cf321c)
* fix(mcp): resolve call_tool by registry without requiring tool map
Multi-worker reloads put MCP servers in the registry from the DB but do
not re-run tools/list on every process. Gating call_tool on
tool_name_to_mcp_server_name_mapping made cold workers 500 with Tool not
found after another worker had already listed the tool. Treat a registry
match on server id/name/alias as enough; upstream rejects unknown tools
* test(e2e): poll MCP register, tools/list, and tools/call across multi-worker lag
Stage multi-worker gateways only load MCP servers and tool maps on the
process that handled the request. Poll until the server is listed, the
tool appears on tools/list, and tools/call is not a cold-worker 500 so
key-access and Datadog MCP e2e stop racing the LB
* Revert "fix(mcp): resolve call_tool by registry without requiring tool map"
This reverts commit 8b56e51e39.
* test(e2e): tighten MCP multi-worker lag classifier
Only retry tools/call on gateway shapes Tool <name> not found and
server_not_found, not any 500 that mentions tool/server not found, so
upstream failures are not retried until the poll deadline
* test(e2e): drop unit file for MCP lag classifier
The live await_call_tool polls already cover multi-worker lag; a separate
string-match unit module is not worth keeping
Gemini returns each thoughtSignature on exactly one part. LiteLLM
stores a function-call signature both message-level (thought_signatures)
and on the tool call itself, then re-attached it to BOTH the text part
and the function-call part when serializing history. gemini-3 and newer
models bill every replayed copy as the previous turn's full reasoning
token count, so long agentic sessions doubled their context growth and
hit the 1,048,576-token limit
Only attach a message-level signature to the text part when the same
signature is not already carried by a tool-call part:
- compare signature values instead of boolean presence so a distinct
text-part signature is never dropped
- ignore the gemini-3 dummy-signature fallback during detection so
replaying gemini-2.5 history to a newer model keeps the real text
signature
- count signatures carried by server-side tool invocations so they are
not re-attached to the text part
gemini-2.5 responses (signature on the text part, function call
unsigned) are unaffected: the text signature is preserved as before
An auto-router deployment's litellm_params.model (auto_router/...) is the
discriminator the router loads it by, but the model management endpoints
accepted any client-supplied value verbatim; a doubled or stripped prefix
made router init fail on the next load and ignore_invalid_deployments
silently dropped the deployment. Validate writes that supply
litellm_params.model at all three endpoints against the merged params and
reject incoherent values with an actionable 400. Classification is
extracted to router_utils/auto_router_model_naming.py so the Router
predicates and the validation share one source
Every model-write endpoint returned 200 off the DB write alone; a model the
reload dropped (ignore_invalid_deployments, or a wholesale reload failure)
stayed invisible on every channel at once, which is how the registry-leak
defect went undiagnosed for three weeks. ProxyConfig.add_deployment and
clear_cache now return whether the reload pass completed, and each write
endpoint verifies the rows it wrote are live in this pod's router afterwards,
distinguishing a deliberately environment-inactive model via the same
predicate the Router's own gate uses. The access-group writers return the
mutated id set instead of discarding it
A poll may start with remaining budget and still return after started+timeout
if the transport overruns its clamp. Recheck the first-listing deadline after
the response so a late listing does not open the continuous DB-sync phase
(cherry picked from commit 7ff2bcbf14)
When less than one full poll interval remained in the first-listing budget,
the pre-sleep check returned NotServable without another /v1/models call.
Sleep only min(interval, time left) so a model that becomes listable in the
last seconds of the timeout still gets a clamped final poll
(cherry picked from commit 8439195922)
create_model returned after the first /v1/models hit that listed the model,
so chat could still land on a cold gateway worker (numWorkers>1 / peer pod)
and 400 Invalid model name. Require continuous listing for the product
default add_deployment interval (30s) after first sight so every worker has
synced from the DB; first listing still bounded at 40s
(cherry picked from commit 7d1ee2ff86)
_await_model_servable used poll_timeout (120s), the spend/log read-back
budget. A stuck model reload therefore stalled every suite that creates a
deployment for two minutes before failing
Give create_model a fixed harness middle ground: model_servable_timeout=40s,
polled every 2s, with each /v1/models call capped at 5s and clamped to the
remaining deadline so one slow GET cannot overrun the wait. Happy path still
returns on the first listing. Not derived from proxy general_settings or env
Transport.get accepts an optional per-call timeout for that clamp. Unit tests
cover the deadline arithmetic and clamp without a live proxy
(cherry picked from commit c082a0e648)
* fix(proxy): warm rotate Prisma client for IAM refresh
* fix(proxy): drain Prisma operations during IAM rotation
* fix(proxy): bound the drain wait when retiring a replaced prisma engine
A replaced engine waited indefinitely for its drain tracker to empty.
Hung queries self-release via prisma's 30s default HTTP timeout, but a
transaction whose owner is hard-cancelled before commit/rollback leaks
its drain count forever, keeping the retired engine and its DB
connection pool alive indefinitely; at one rotation per 12 minutes such
engines accumulate. Cap the wait at 90 seconds, which exceeds every
legitimate operation bound (30s HTTP timeout, 60s max interactive
transaction timeout in this codebase), then kill the engine anyway.
Work killed at the deadline degrades to the pre-drain behavior and is
retried by the existing reconnect/backoff layers.
---------
Co-authored-by: ryan-crabbe-berri <ryan@berri.ai>
Scrub aliases on delete only when the deleted deployment's model_name no
longer resolves in the router. A legacy load-balanced team model can have
several deployment rows sharing one internal name; deleting one replica
must not remove aliases that still route to the survivors, in any team
A team's model_aliases can map a public name like gpt-4 to the internal
routing key (model_name_{team_id}_{uuid}) of a team deployment that has
since been deleted, e.g. after replacing per-team duplicates with one
gateway-level model. The pre-call rewrite then sent every request to a
name the router cannot serve, failing with "no healthy deployments for
model_name_..." even though the requested name still resolves at the
gateway level. The rewrite is now skipped when the alias target has no
live deployment in the router
delete_model also skipped the team alias scan for internal-shaped names
on the assumption they can never be alias values, which is exactly the
shape legacy team model aliases have, so deleting a legacy team model
left the stale alias behind. The scan now always runs, and a public
name that still resolves to a live router deployment (e.g. a shared
gateway-level model group) stays in team.models so the delete does not
revoke the team's access to it
