mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-26 01:12:21 +00:00
f7ae9efad2
* test(e2e): hold every worker under an idle RSS budget before any traffic The harness reads /debug/memory/summary on every replica once at collection time, right after the readiness gate and before this pytest process sends any traffic, and the memory suite's first test fails when any worker idles past E2E_MEMORY_IDLE_RSS_BUDGET_MB (768 MB by default) or gives no reading at all. A v1.100.x worker with a database idled at 836-886 MB where v1.101.0rc1 idled at 544 MB on the same database: prisma-client-py's default recursive type depth generated 91k TypedDict classes that v1.101.0's recursive_type_depth = -1 cut to 19k. The budget starts at the rc1 reading plus headroom. * test(e2e): read idle RSS only when the idle budget test is selected Gate the collection-time /debug/memory/summary read on a selected test using the idle_rss fixture and skip it under --collect-only, so sessions that never run the idle budget test pay no round trip. Drop the markerless unit test file the e2e guide bans and assert live that every configured replica was measured * test(e2e): take the idle RSS read after collection settles Read every replica's RSS from a tryfirst pytest_collection_finish hook so -k and -m deselection has already run, and only when a selected test still asks for the idle_rss fixture and the run is not --collect-only * test(e2e): record the heaviest idle RSS reading as junit properties * test(e2e): attach the idle RSS properties from the harness's setup hook --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
42 lines
15 KiB
YAML
42 lines
15 KiB
YAML
# Reliability & Performance (behavior features). Grounded in litellm/router.py + router_strategy/ + router_utils/.
|
|
- {id: reliability.fallback.5xx.routes_to_fallback, module: reliability, tier: P0, behavior: fallback, variant: "5xx", assertions: [routes_to_fallback], exercised_on: [chat_completions], source: "litellm/router.py:2024", rationale: "Reroute on provider 5xx to alternate deployment"}
|
|
- {id: reliability.fallback.context_window.routes_to_fallback, module: reliability, tier: P0, behavior: fallback, variant: context_window, assertions: [routes_to_fallback], exercised_on: [chat_completions], source: "litellm/router.py:6108", rationale: "Fallback when model exceeds context limit"}
|
|
- {id: reliability.fallback.content_policy.routes_to_fallback, module: reliability, tier: P0, behavior: fallback, variant: content_policy, assertions: [routes_to_fallback], exercised_on: [chat_completions], source: "litellm/router.py:6023", rationale: "Reroute on content-policy violation"}
|
|
- {id: reliability.fallback.timeout.routes_to_fallback, module: reliability, tier: P0, behavior: fallback, variant: "timeout", assertions: [routes_to_fallback], exercised_on: [chat_completions], source: "litellm/router.py:2766", rationale: "Fallback on request timeout"}
|
|
- {id: reliability.retry.5xx.succeeds_within_retries, module: reliability, tier: P0, behavior: retry, variant: "5xx", assertions: [succeeds_within_retries], exercised_on: [chat_completions], source: "litellm/router.py:6414", rationale: "Transient 5xx often succeeds on retry"}
|
|
- {id: reliability.retry.timeout.succeeds_within_retries, module: reliability, tier: P0, behavior: retry, variant: timeout, assertions: [succeeds_within_retries], exercised_on: [chat_completions], source: "get_retry_from_policy.py:44", rationale: "Timeout retried per policy"}
|
|
- {id: reliability.retry.429.succeeds_within_retries, module: reliability, tier: P0, behavior: retry, variant: "429", assertions: [succeeds_within_retries], exercised_on: [chat_completions], source: "get_retry_from_policy.py:46", rationale: "429 retried per RateLimitErrorRetries policy"}
|
|
- {id: reliability.retry.auth.succeeds_within_retries, module: reliability, tier: P1, behavior: retry, variant: auth, assertions: [succeeds_within_retries], exercised_on: [chat_completions], source: "get_retry_from_policy.py:42", rationale: "Transient auth glitch retry"}
|
|
- {id: reliability.retry.context_window.succeeds_within_retries, module: reliability, tier: P1, behavior: retry, variant: context_window, assertions: [succeeds_within_retries], exercised_on: [chat_completions], source: "get_retry_from_policy.py:51", fail_before_fix: proven, rationale: "A context-window 400 under BadRequestErrorRetries retries onto a sibling deployment in the same model group, instead of coming straight back as the 400 the deployment that just refused it returned"}
|
|
- {id: reliability.cooldown.5xx.trips_then_recovers, module: reliability, tier: P0, behavior: cooldown, variant: "5xx", assertions: [trips_then_recovers], exercised_on: [chat_completions], source: "cooldown_handlers.py:40", rationale: "Deployment cools after repeated 5xx, recovers after cooldown_time"}
|
|
- {id: reliability.cooldown.sibling_replica.serves_backup_within_read_interval, module: reliability, tier: P1, behavior: cooldown, variant: sibling_replica, assertions: [serves_backup_within_read_interval], exercised_on: [chat_completions], source: "cooldown_cache.py:44", fail_before_fix: proven, rationale: "A bench taken on one gateway reaches a sibling that already holds the key's read timer within the 1s Redis read interval plus margin, so its next call lands on the backup"}
|
|
- {id: reliability.cooldown.429.trips_then_recovers, module: reliability, tier: P0, behavior: cooldown, variant: "429", assertions: [trips_then_recovers], exercised_on: [chat_completions], source: "cooldown_handlers.py:69", rationale: "Cools on 429, avoids hammering exhausted provider"}
|
|
- {id: reliability.cooldown.auth.trips_then_recovers, module: reliability, tier: P1, behavior: cooldown, variant: auth, assertions: [trips_then_recovers], exercised_on: [chat_completions], source: "cooldown_handlers.py:74", rationale: "Cools on 401 auth error"}
|
|
- {id: reliability.cooldown.timeout.trips_then_recovers, module: reliability, tier: P1, behavior: cooldown, variant: timeout, assertions: [trips_then_recovers], exercised_on: [chat_completions], source: "cooldown_handlers.py:77", rationale: "Cools on 408 timeout"}
|
|
- {id: reliability.cooldown.client_disconnect.stays_healthy, module: reliability, tier: P0, behavior: cooldown, variant: client_disconnect, assertions: [stays_healthy], exercised_on: [chat_completions], source: "llms/azure/azure.py:484", fail_before_fix: proven, rationale: "A client hanging up mid-request under cancel_on_disconnect never benches the Azure deployment it was talking to: the cancellation used to surface as a fake 500 that tripped the cooldown and sent every caller behind it to billed fallbacks (GitHub issues #35329 and #42222)"}
|
|
- {id: reliability.routing.simple_shuffle.picks_healthy_deployment, module: reliability, tier: P1, behavior: routing, variant: simple_shuffle, assertions: [picks_healthy_deployment], exercised_on: [chat_completions], source: "router_strategy/simple_shuffle.py", rationale: "Baseline weighted/uniform pick"}
|
|
- {id: reliability.routing.latency_based.picks_lowest_latency, module: reliability, tier: P1, behavior: routing, variant: latency_based, assertions: [picks_lowest_latency], exercised_on: [chat_completions], source: "router_strategy/lowest_latency.py", rationale: "Routes to lowest-latency deployment"}
|
|
- {id: reliability.routing.cost_based.picks_lowest_cost, module: reliability, tier: P1, behavior: routing, variant: cost_based, assertions: [picks_lowest_cost], exercised_on: [chat_completions], source: "router_strategy/lowest_cost.py", rationale: "Spend-aware routing"}
|
|
- {id: reliability.routing.usage_based.picks_under_tpm, module: reliability, tier: P0, behavior: routing, variant: usage_based, assertions: [picks_under_tpm], exercised_on: [chat_completions], source: "router_strategy/lowest_tpm_rpm_v2.py", rationale: "Routes to lowest-TPM deployment; prevents over-allocation"}
|
|
- {id: reliability.routing.least_busy.picks_lowest_traffic, module: reliability, tier: P1, behavior: routing, variant: least_busy, assertions: [picks_lowest_traffic], exercised_on: [chat_completions], source: "router_strategy/least_busy.py", rationale: "Fewest in-flight requests"}
|
|
- {id: reliability.routing.complexity_llm_classifier.routes_by_llm_tier, module: reliability, tier: P1, behavior: routing, variant: complexity_llm_classifier, assertions: [routes_by_llm_tier], exercised_on: [chat_completions], source: "router_strategy/complexity_router/complexity_router.py", fail_before_fix: proven, rationale: "v2 auto-router LLM complexity classifier runs over the proxy and routes by semantic tier instead of silently crashing on absent litellm_metadata and falling back to heuristic scoring"}
|
|
- {id: reliability.routing.tagged_marker.request_tag_selects_marker, module: reliability, tier: P0, behavior: routing, variant: tagged_marker, assertions: [request_tag_selects_marker], exercised_on: [chat_completions], source: "litellm/router.py:11445", rationale: "Tagged request selects the tagged strategy marker under a shared model_name instead of the plain deployment registered first (GitHub issue #36619)"}
|
|
- {id: reliability.routing.tagged_marker.untagged_request_served_by_plain_deployment, module: reliability, tier: P0, behavior: routing, variant: tagged_marker, assertions: [untagged_request_served_by_plain_deployment], exercised_on: [chat_completions, messages, responses], source: "litellm/router.py:11445", rationale: "Untagged requests to a shared model_name are served by the plain deployment on every call, never captured or errored by the tagged marker (GitHub issue #36620)"}
|
|
- {id: reliability.routing.tagged_marker.header_tag_selects_marker, module: reliability, tier: P0, behavior: routing, variant: tagged_marker, assertions: [header_tag_selects_marker], exercised_on: [messages], source: "litellm/router.py:11445", rationale: "A request tagged only via the x-litellm-tags header selects the tagged marker on Anthropic-native /v1/messages (GitHub issue #36621)"}
|
|
- {id: reliability.routing.tagged_marker.untagged_tier_deployments_still_served, module: reliability, tier: P1, behavior: routing, variant: tagged_marker, assertions: [untagged_tier_deployments_still_served], exercised_on: [chat_completions, messages], source: "litellm/router_strategy/tag_based_routing.py:433", rationale: "Routing tags the marker consumed no longer constrain deployment selection inside the routed tier group, so untagged tier deployments serve the rewrite (GitHub issue #36621)"}
|
|
- {id: reliability.routing.tagged_marker.tag_semantics_stay_strict, module: reliability, tier: P1, behavior: routing, variant: tagged_marker, assertions: [tag_semantics_stay_strict], exercised_on: [chat_completions], source: "litellm/router_strategy/tag_based_routing.py:299", rationale: "Tag consumption must not loosen strict semantics: a tagged call aimed straight at an untagged deployment still gets the 401 tags-configuration denial"}
|
|
- {id: reliability.routing.tagged_marker.responses_input_routes_through_marker, module: reliability, tier: P0, behavior: routing, variant: tagged_marker, assertions: [responses_input_routes_through_marker], exercised_on: [responses], source: "litellm/router.py:11489", rationale: "Tagged /v1/responses (header or litellm_metadata.tags, string or list input) routes through the marker to its tier, extending the GitHub issues #36620/#36621 tag split to the Responses surface"}
|
|
- {id: reliability.routing.tagged_marker.alias_connection_params_stay_with_tier, module: reliability, tier: P0, behavior: routing, variant: tagged_marker, assertions: [alias_connection_params_stay_with_tier], exercised_on: [chat_completions], source: "litellm/router.py:11567", rationale: "An api_key or api_base on the marker alias is never forwarded onto the routed request; the tier deployment calls its provider with its own credential (GitHub PR #36626)"}
|
|
- {id: reliability.routing.semantic_auto_router.responses_input_routed, module: reliability, tier: P0, behavior: routing, variant: semantic_auto_router, assertions: [responses_input_routed], exercised_on: [responses], source: "litellm/router_strategy/auto_router/auto_router.py:131", fail_before_fix: proven, rationale: "/v1/responses input is resolved into messages for the semantic auto-router pre-routing hook instead of failing 400 Unmapped LLM provider auto_router (GitHub PR #37333)"}
|
|
- {id: reliability.routing.strategy_alias.custom_pricing_ignored, module: reliability, tier: P1, behavior: routing, variant: strategy_alias, assertions: [custom_pricing_ignored], exercised_on: [chat_completions], source: "litellm/router.py:11489", rationale: "Custom pricing on a strategy-router alias never prices the routed request; spend logs at the routed tier deployment's own rate (GitHub PR #36691)"}
|
|
- {id: reliability.routing.complexity_heuristic.scores_current_ask_only, module: reliability, tier: P1, behavior: routing, variant: complexity_heuristic, assertions: [scores_current_ask_only], exercised_on: [chat_completions], source: "router_strategy/complexity_router/complexity_router.py:942", rationale: "The heuristic complexity classifier scores the caller's current ask only, so a keyword-heavy agent system prompt cannot inflate the tier (GitHub PR #36721)"}
|
|
- {id: reliability.cache.exact.returns_cached, module: reliability, tier: P1, behavior: cache, variant: exact, assertions: [returns_cached], exercised_on: [chat_completions, messages, embeddings], source: "litellm/caching/caching.py", rationale: "Response cache returns cached on exact match"}
|
|
- {id: reliability.cache.prompt_caching_model_select.returns_cached, module: reliability, tier: P1, behavior: cache, variant: prompt_caching_model_select, assertions: [returns_cached], exercised_on: [chat_completions], source: "router_utils/prompt_caching_cache.py", rationale: "Selects model supporting prompt caching for cacheable prefix; runs only on a stack with the prompt_caching pre-call check enabled (E2E_PROMPT_CACHING_STACK)"}
|
|
- {id: reliability.circuit_breaker.redis.trips_then_recovers, module: reliability, tier: P0, behavior: circuit_breaker, variant: redis, assertions: [trips_then_recovers], exercised_on: [chat_completions, messages, embeddings], source: "litellm/caching/redis_cache.py:99", rationale: "Redis breaker CLOSED->OPEN->HALF_OPEN; guards all cache/rate-limit ops"}
|
|
- {id: reliability.circuit_breaker.redis_timeout.stays_responsive, module: reliability, tier: P1, behavior: circuit_breaker, variant: redis_timeout, assertions: [stays_responsive], exercised_on: [chat_completions, messages], source: "litellm/proxy/hooks/proxy_track_cost_callback.py:386", fail_before_fix: proven, rationale: "Under locust load split round robin over /chat/completions and /v1/messages with every request retrying through failing mock deployments, holding Redis in CLIENT PAUSE ALL for the phase trips the breaker and every request still succeeds, with latency, RSS, and CPU reported as p50/p90/p99 against the pre-pause baseline; on v1.100.0 the failed-tracking alert body doubled per request until the worker OOMed (LIT-6780)"}
|
|
- {id: reliability.timeout.request_timeout.exceeds_deadline, module: reliability, tier: P1, behavior: timeout, variant: request_timeout, assertions: [exceeds_deadline], exercised_on: [chat_completions, messages], source: "litellm/router.py:545-551", rationale: "Per-request timeout raises Timeout"}
|
|
- {id: reliability.timeout.stream_timeout.exceeds_deadline, module: reliability, tier: P1, behavior: timeout, variant: stream_timeout, assertions: [exceeds_deadline], exercised_on: [chat_completions], source: "litellm/router.py:551", rationale: "Streaming chunk-delivery timeout"}
|
|
- {id: reliability.perf.throughput.under_slo, module: reliability, tier: P1, behavior: perf, variant: throughput, assertions: [under_slo], exercised_on: [chat_completions, messages], source: grammar, rationale: "Throughput SLO under load"}
|
|
- {id: reliability.perf.memory.under_slo, module: reliability, tier: P1, behavior: perf, variant: memory, assertions: [under_slo], exercised_on: [chat_completions], source: grammar, rationale: "Proxy RSS and the stored request snapshot stay within fixed budgets across a few hundred failing requests with retries and fallbacks, the v1.100.0 retry-breadcrumb leak shape (MAT-335)"}
|
|
- {id: reliability.perf.idle_memory.under_slo, module: reliability, tier: P1, behavior: perf, variant: idle_memory, assertions: [under_slo], exercised_on: [], source: grammar, rationale: "Every worker's RSS as read right after the readiness gate and before any test traffic stays under a fixed idle budget; a DB-backed v1.100.x worker idled at 886 MB against a 2 GiB pod limit where v1.101.0rc1 idled at 544 MB"}
|
|
- {id: reliability.perf.session_anomaly.under_slo, module: reliability, tier: P1, behavior: perf, variant: session_anomaly, assertions: [under_slo], exercised_on: [messages], source: grammar, rationale: "Weekly Claude Code-shaped multi-turn session load against real providers; ceilings on error rate, warm-turn cache read/write, p95 turn time, and gateway-recorded spend (LIT-4562)"}
|