feat(routing): move sticky-key retries into routing policy with lazy attempts

Replace the provider/endpoint max_retries fields as the source of same-key
retries with a routing policy setting, sticky_key_attempts (default 2). Only
the first-ranked candidate is retried on the same key; every failover
candidate gets a single attempt so failover keeps advancing instead of
retrying each fallback key.

Materialize exactly one attempt per candidate and derive same-key retries in
the attempt loop after a candidate-scoped failure, so the retry budget no
longer inflates up-front materialization and needs no upper bound. The budget
travels in the report context; retries reuse the plan with a fresh candidate
id and incremented retry index. Pool groups only retry their first key within
the retry-index stride.

Expose the setting in the routing profile editor and the set_scheduling rule
action, and drop the max_retries input from the provider form.
This commit is contained in:
elky
2026-09-02 20:48:40 +08:00
parent 415b2da81b
commit 7323d41fbe
40 changed files with 851 additions and 570 deletions
@@ -47,13 +47,12 @@ use crate::cache::{
use crate::clock::current_unix_ms;
use crate::dispatch::refs::dispatch_ref_for_local_candidate;
use crate::handlers::shared::provider_pool::admin_provider_pool_config_from_config_value;
use crate::orchestration::{local_attempt_slot_count, ExecutionAttemptIdentity};
use crate::orchestration::{ExecutionAttemptIdentity, POOL_KEY_RETRY_INDEX_STRIDE};
use crate::scheduler::candidate::is_auth_api_key_concurrency_limit_skip_reason;
use crate::scheduler::config::SchedulerSchedulingMode;
use crate::stage_metrics::observe_gateway_stage_ms;
use crate::{AppState, GatewayError};
const POOL_KEY_RETRY_INDEX_STRIDE: u32 = 100;
const AUTH_API_KEY_CONCURRENCY_WAIT_BUDGET: Duration = Duration::from_millis(100);
const AUTH_API_KEY_CONCURRENCY_RETRY_DELAY: Duration = Duration::from_millis(10);
@@ -481,10 +480,6 @@ where
type ExtraData = Value;
type Error = Infallible;
fn attempt_slot_count(&self, candidate: &Self::Candidate) -> u32 {
local_attempt_slot_count(&candidate.transport)
}
fn build_extra_data(&self, candidate: &Self::Candidate) -> Option<Self::ExtraData> {
available_candidate_extra_data_with_dispatch_ref(candidate, &self.build_extra_data)
}
@@ -1610,7 +1605,8 @@ async fn persist_available_local_execution_candidate_at_index<F>(
where
F: Fn(&EligibleLocalExecutionCandidate) -> Option<Value> + Send + Sync,
{
let attempt_slots = local_attempt_slot_count(&candidate.transport).max(1);
// Exactly one attempt is materialized per candidate; same-key retries are
// derived lazily by the attempt loop after a failure.
let extra_data = ai_candidate_extra_data_with_ranking(
available_candidate_base_extra_data_with_dispatch_ref(&candidate, build_extra_data),
candidate.ranking.as_ref(),
@@ -1625,53 +1621,34 @@ where
Some(candidate_index),
extra_data,
);
let should_persist = should_persist_available_local_candidate(&candidate);
let mut attempts = Vec::with_capacity(attempt_slots as usize);
let mut owned_candidate = Some(candidate);
let retry_index = effective_retry_index(0, candidate.orchestration.pool_key_index);
let generated_candidate_id = Uuid::new_v4().to_string();
let candidate_id = if should_persist_available_local_candidate(&candidate) {
state
.persist_available_local_candidate(
trace_id,
context.user_id,
context.api_key_id,
&candidate.candidate,
candidate_index,
retry_index,
generated_candidate_id.as_str(),
context.required_capabilities,
extra_data,
current_unix_ms(),
context.error_context,
)
.await
} else {
generated_candidate_id
};
for retry_index in 0..attempt_slots {
let candidate_ref = owned_candidate
.as_ref()
.expect("candidate should remain available until final retry");
let generated_candidate_id = Uuid::new_v4().to_string();
let candidate_id = if should_persist {
state
.persist_available_local_candidate(
trace_id,
context.user_id,
context.api_key_id,
&candidate_ref.candidate,
candidate_index,
effective_retry_index(retry_index, candidate_ref.orchestration.pool_key_index),
generated_candidate_id.as_str(),
context.required_capabilities,
extra_data.clone(),
current_unix_ms(),
context.error_context,
)
.await
} else {
generated_candidate_id
};
let candidate = if retry_index + 1 == attempt_slots {
owned_candidate
.take()
.expect("final retry should consume owned candidate")
} else {
candidate_ref.clone()
};
let retry_index =
effective_retry_index(retry_index, candidate.orchestration.pool_key_index);
attempts.push(LocalExecutionCandidateAttempt {
eligible: candidate,
candidate_index,
retry_index,
candidate_id,
});
}
attempts
vec![LocalExecutionCandidateAttempt {
eligible: candidate,
candidate_index,
retry_index,
candidate_id,
}]
}
fn available_candidate_extra_data_with_dispatch_ref<F>(
@@ -1924,32 +1901,15 @@ fn build_unpersisted_local_execution_candidate_attempts(
candidate: EligibleLocalExecutionCandidate,
candidate_index: u32,
) -> VecDeque<LocalExecutionCandidateAttempt> {
let attempt_slots = local_attempt_slot_count(&candidate.transport).max(1);
let mut attempts = VecDeque::with_capacity(attempt_slots as usize);
let mut owned_candidate = Some(candidate);
for retry_index in 0..attempt_slots {
let candidate = if retry_index + 1 == attempt_slots {
owned_candidate
.take()
.expect("final retry should consume owned candidate")
} else {
owned_candidate
.as_ref()
.expect("candidate should remain available until final retry")
.clone()
};
let retry_index =
effective_retry_index(retry_index, candidate.orchestration.pool_key_index);
attempts.push_back(LocalExecutionCandidateAttempt {
eligible: candidate,
candidate_index,
retry_index,
candidate_id: Uuid::new_v4().to_string(),
});
}
attempts
// One attempt per candidate; same-key retries are derived lazily by the
// attempt loop after a failure.
let retry_index = effective_retry_index(0, candidate.orchestration.pool_key_index);
VecDeque::from([LocalExecutionCandidateAttempt {
eligible: candidate,
candidate_index,
retry_index,
candidate_id: Uuid::new_v4().to_string(),
}])
}
async fn persist_pool_group_exhaustion_skipped_candidate(
@@ -2277,6 +2237,8 @@ mod tests {
pool_key_index,
pool_key_lease: None,
scheduler_affinity_epoch: None,
// These tests cover persistence shape, not same-key retries.
sticky_key_attempts: Some(1),
},
ranking: None,
}
@@ -371,6 +371,7 @@ mod tests {
priority_mode: aether_routing_core::RoutingSetPriorityMode::Provider,
scheduling_mode: aether_routing_core::RoutingSchedulingMode::CacheAffinity,
keep_priority_on_conversion: false,
sticky_key_attempts: aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS,
ranking_overlay: aether_routing_core::RankingOverlay::default(),
mutation_plan: Default::default(),
pool_policy_overrides: BTreeMap::new(),
@@ -406,6 +407,7 @@ mod tests {
priority_mode: aether_routing_core::RoutingSetPriorityMode::Provider,
scheduling_mode: aether_routing_core::RoutingSchedulingMode::FixedOrder,
keep_priority_on_conversion: false,
sticky_key_attempts: aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS,
ranking_overlay: Default::default(),
mutation_plan: Default::default(),
pool_policy_overrides: Default::default(),
@@ -443,6 +445,7 @@ mod tests {
priority_mode: aether_routing_core::RoutingSetPriorityMode::GlobalKey,
scheduling_mode: aether_routing_core::RoutingSchedulingMode::CacheAffinity,
keep_priority_on_conversion: false,
sticky_key_attempts: aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS,
ranking_overlay: aether_routing_core::RankingOverlay {
pool_priority_overrides: BTreeMap::from([("provider-1".to_string(), 4)]),
key_priority_overrides: BTreeMap::from([("representative-key".to_string(), 1)]),
@@ -23,7 +23,9 @@ use crate::ai_serving::{
use crate::orchestration::LocalExecutionCandidateMetadata;
use crate::stage_metrics::observe_gateway_stage_ms;
use super::candidate_ranking::rank_eligible_local_execution_candidates;
use super::candidate_ranking::{
rank_eligible_local_execution_candidates, scheduler_ordering_config_for_routing_policy,
};
#[derive(Debug, Clone, PartialEq)]
pub(crate) struct EligibleLocalExecutionCandidate {
@@ -378,8 +380,18 @@ async fn resolve_and_rank_local_execution_candidates_with_pool_expansion(
"candidate_resolution_core",
started_at.elapsed().as_millis() as u64,
);
let sticky_key_attempts = if outcome.eligible_candidates.is_empty() {
None
} else {
Some(
scheduler_ordering_config_for_routing_policy(state, routing_policy)
.await
.sticky_key_attempts,
)
};
for candidate in &mut outcome.eligible_candidates {
candidate.orchestration.scheduler_affinity_epoch = Some(scheduler_affinity_epoch);
candidate.orchestration.sticky_key_attempts = sticky_key_attempts;
}
(outcome.eligible_candidates, outcome.skipped_candidates)
}
@@ -1889,6 +1889,7 @@ mod tests {
priority_mode: aether_routing_core::RoutingSetPriorityMode::Provider,
scheduling_mode: aether_routing_core::RoutingSchedulingMode::FixedOrder,
keep_priority_on_conversion: false,
sticky_key_attempts: aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS,
ranking_overlay: Default::default(),
mutation_plan: Default::default(),
pool_policy_overrides: Default::default(),
@@ -1952,6 +1953,7 @@ mod tests {
priority_mode: aether_routing_core::RoutingSetPriorityMode::Provider,
scheduling_mode: aether_routing_core::RoutingSchedulingMode::FixedOrder,
keep_priority_on_conversion: false,
sticky_key_attempts: aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS,
ranking_overlay: Default::default(),
mutation_plan: Default::default(),
pool_policy_overrides: Default::default(),
@@ -2689,6 +2691,7 @@ mod tests {
priority_mode: aether_routing_core::RoutingSetPriorityMode::Provider,
scheduling_mode: aether_routing_core::RoutingSchedulingMode::FixedOrder,
keep_priority_on_conversion: true,
sticky_key_attempts: aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS,
ranking_overlay: Default::default(),
mutation_plan: Default::default(),
pool_policy_overrides: Default::default(),
@@ -203,6 +203,7 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
client_session_affinity: input.client_session_affinity.as_ref(),
routing_policy: input.routing_policy.as_ref(),
scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
sticky_key_attempts: eligible.orchestration.sticky_key_attempts,
client_requested_stream: body_json
.get("stream")
.and_then(serde_json::Value::as_bool)
@@ -4,7 +4,7 @@ use aether_ai_serving::{
build_ai_execution_report_context,
insert_provider_stream_event_api_format as insert_ai_provider_stream_event_api_format,
provider_stream_event_api_format_for_provider_type as ai_provider_stream_event_api_format_for_provider_type,
AiExecutionReportContextParts, AiRequestOrigin,
AiExecutionReportContextParts, AiRequestOrigin, STICKY_KEY_ATTEMPTS_REPORT_FIELD,
};
use aether_routing_core::ResolvedRoutingPolicy;
use aether_runtime_state::RuntimeLockLease;
@@ -59,6 +59,9 @@ pub(crate) struct LocalExecutionReportContextParts<'a> {
pub(crate) client_session_affinity: Option<&'a ClientSessionAffinity>,
pub(crate) routing_policy: Option<&'a ResolvedRoutingPolicy>,
pub(crate) scheduler_affinity_epoch: Option<u64>,
/// Routing policy sticky-key attempt budget; read back by the attempt
/// loop to derive same-key retries lazily.
pub(crate) sticky_key_attempts: Option<u32>,
pub(crate) client_requested_stream: bool,
pub(crate) upstream_is_stream: bool,
pub(crate) has_envelope: bool,
@@ -124,6 +127,12 @@ pub(crate) fn build_local_execution_report_context(
Value::Number(epoch.into()),
);
}
if let Some(sticky_key_attempts) = parts.sticky_key_attempts {
extra_fields.insert(
STICKY_KEY_ATTEMPTS_REPORT_FIELD.to_string(),
Value::Number(sticky_key_attempts.into()),
);
}
insert_request_path_fields(
&mut extra_fields,
parts.request_path,
@@ -330,6 +339,7 @@ mod tests {
client_session_affinity: Some(&client_session_affinity),
routing_policy: None,
scheduler_affinity_epoch: None,
sticky_key_attempts: None,
client_requested_stream: false,
upstream_is_stream: false,
has_envelope: false,
@@ -413,6 +423,7 @@ mod tests {
client_session_affinity: None,
routing_policy: None,
scheduler_affinity_epoch: None,
sticky_key_attempts: None,
client_requested_stream: false,
upstream_is_stream: true,
has_envelope: false,
@@ -480,6 +491,7 @@ mod tests {
client_session_affinity: None,
routing_policy: None,
scheduler_affinity_epoch: None,
sticky_key_attempts: None,
client_requested_stream: false,
upstream_is_stream: false,
has_envelope: false,
@@ -109,6 +109,7 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
client_session_affinity: input.client_session_affinity.as_ref(),
routing_policy: input.routing_policy.as_ref(),
scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
sticky_key_attempts: eligible.orchestration.sticky_key_attempts,
client_requested_stream: spec_metadata.require_streaming,
upstream_is_stream: spec_metadata.require_streaming,
has_envelope: false,
@@ -122,6 +122,7 @@ pub(super) async fn maybe_build_local_openai_image_decision_payload_for_candidat
client_session_affinity: input.client_session_affinity.as_ref(),
routing_policy: input.routing_policy.as_ref(),
scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
sticky_key_attempts: eligible.orchestration.sticky_key_attempts,
client_requested_stream: spec_metadata.require_streaming,
upstream_is_stream,
has_envelope: false,
@@ -90,6 +90,7 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
client_session_affinity: input.client_session_affinity.as_ref(),
routing_policy: input.routing_policy.as_ref(),
scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
sticky_key_attempts: eligible.orchestration.sticky_key_attempts,
client_requested_stream: false,
upstream_is_stream: false,
has_envelope: false,
@@ -142,6 +142,7 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
client_session_affinity: input.client_session_affinity.as_ref(),
routing_policy: input.routing_policy.as_ref(),
scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
sticky_key_attempts: eligible.orchestration.sticky_key_attempts,
client_requested_stream: body_json
.get("stream")
.and_then(serde_json::Value::as_bool)
@@ -195,6 +195,7 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
client_session_affinity: input.client_session_affinity.as_ref(),
routing_policy: input.routing_policy.as_ref(),
scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
sticky_key_attempts: eligible.orchestration.sticky_key_attempts,
client_requested_stream: body_json
.get("stream")
.and_then(serde_json::Value::as_bool)
@@ -184,6 +184,7 @@ pub(crate) async fn maybe_build_local_openai_responses_decision_payload_for_cand
client_session_affinity: input.client_session_affinity.as_ref(),
routing_policy: input.routing_policy.as_ref(),
scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
sticky_key_attempts: eligible.orchestration.sticky_key_attempts,
client_requested_stream: body_json
.get("stream")
.and_then(serde_json::Value::as_bool)
@@ -1951,11 +1951,13 @@ fn apply_pool_orchestration(
orchestration: PoolCandidateOrchestration,
) -> EligibleLocalExecutionCandidate {
let scheduler_affinity_epoch = candidate.orchestration.scheduler_affinity_epoch;
let sticky_key_attempts = candidate.orchestration.sticky_key_attempts;
candidate.orchestration = LocalExecutionCandidateMetadata {
candidate_group_id: orchestration.candidate_group_id,
pool_key_index: orchestration.pool_key_index,
pool_key_lease: None,
scheduler_affinity_epoch,
sticky_key_attempts,
};
candidate
}
@@ -2159,6 +2161,7 @@ mod tests {
pool_key_index: Some(0),
pool_key_lease: None,
scheduler_affinity_epoch: None,
sticky_key_attempts: None,
}
);
assert_eq!(reordered[1].orchestration.pool_key_index, Some(1));
@@ -2176,6 +2179,7 @@ mod tests {
pool_key_index: None,
pool_key_lease: None,
scheduler_affinity_epoch: None,
sticky_key_attempts: None,
}
);
}
@@ -5072,6 +5076,7 @@ mod tests {
priority_mode: RoutingSetPriorityMode::Provider,
scheduling_mode: RoutingSchedulingMode::CacheAffinity,
keep_priority_on_conversion: false,
sticky_key_attempts: aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS,
ranking_overlay: RankingOverlay {
allowed_keys: key_ids.into_iter().map(str::to_string).collect(),
..RankingOverlay::default()
+1
View File
@@ -208,6 +208,7 @@ mod tests {
pool_key_index: None,
pool_key_lease: None,
scheduler_affinity_epoch: None,
sticky_key_attempts: None,
},
ranking: None,
}
@@ -252,6 +252,10 @@ where
Ok(())
}
async fn next_same_key_retry(&self, attempt: &T) -> Result<Option<T>, Self::Error> {
Ok(crate::orchestration::next_same_key_retry_attempt(attempt))
}
async fn record_attempt_failed(&self, attempt: &T) -> Result<(), Self::Error> {
record_provider_transfer_attempt_failed(
self.state,
@@ -788,18 +792,31 @@ where
{
let mut last_attempted = None;
let mut fallback_response = None;
// A same-key retry derived after a candidate-scoped failure runs before
// the source is asked for the next candidate.
let mut pending_same_key_retry: Option<Attempt> = None;
loop {
let next_started_at = std::time::Instant::now();
let next_attempt =
next_execution_attempt_with_timeout(source, trace_id, plan_kind, planning_timeout)
let attempt = match pending_same_key_retry.take() {
Some(attempt) => attempt,
None => {
let next_started_at = std::time::Instant::now();
let next_attempt = next_execution_attempt_with_timeout(
source,
trace_id,
plan_kind,
planning_timeout,
)
.await?;
observe_gateway_stage_ms(
"stream_candidate_next",
next_started_at.elapsed().as_millis() as u64,
);
let Some(attempt) = next_attempt else {
break;
observe_gateway_stage_ms(
"stream_candidate_next",
next_started_at.elapsed().as_millis() as u64,
);
let Some(attempt) = next_attempt else {
break;
};
attempt
}
};
if port.should_skip_attempt(&attempt).await? {
let provider_id = attempt.execution_plan().provider_id.clone();
@@ -839,6 +856,9 @@ where
if attempt_fallback_response.is_some() {
fallback_response = attempt_fallback_response;
}
if scope == AiAttemptRetryScope::Candidate {
pending_same_key_retry = port.next_same_key_retry(&attempt).await?;
}
apply_attempt_retry_scope(source, &attempt, scope).await?;
}
}
@@ -952,6 +972,10 @@ where
Ok(())
}
async fn next_same_key_retry(&self, attempt: &T) -> Result<Option<T>, Self::Error> {
Ok(crate::orchestration::next_same_key_retry_attempt(attempt))
}
async fn record_attempt_failed(&self, attempt: &T) -> Result<(), Self::Error> {
record_provider_transfer_attempt_failed(
self.state,
@@ -1663,6 +1663,22 @@ mod tests {
candidate_index: u32,
endpoint_id: &str,
candidate_id: &str,
) -> AiSyncAttempt {
test_openai_image_heartbeat_attempt_with_sticky_key_attempts(
candidate_index,
endpoint_id,
candidate_id,
1,
)
}
/// `sticky_key_attempts` is pinned so these tests exercise candidate
/// failover; the default same-key retry is covered separately.
fn test_openai_image_heartbeat_attempt_with_sticky_key_attempts(
candidate_index: u32,
endpoint_id: &str,
candidate_id: &str,
sticky_key_attempts: u32,
) -> AiSyncAttempt {
AiSyncAttempt {
plan: test_openai_image_heartbeat_plan(endpoint_id, candidate_id),
@@ -1670,6 +1686,7 @@ mod tests {
report_context: Some(json!({
"candidate_index": candidate_index,
"retry_index": 0,
"sticky_key_attempts": sticky_key_attempts,
})),
}
}
@@ -1828,6 +1845,9 @@ mod tests {
report_context: Some(json!({
"candidate_index": candidate_index,
"retry_index": 0,
// Pin to a single attempt so this helper exercises candidate
// failover rather than the default same-key retry.
"sticky_key_attempts": 1,
"client_api_format": client_api_format,
"provider_api_format": client_api_format,
})),
@@ -1980,6 +2000,90 @@ mod tests {
assert_eq!(body, json!({"data": [{"b64_json": "second-candidate"}]}));
}
#[tokio::test]
async fn openai_image_sync_heartbeat_retries_sticky_key_lazily_before_failover() {
let seen_plans = Arc::new(std::sync::Mutex::new(Vec::<(String, Option<String>)>::new()));
let seen_plans_for_override = Arc::clone(&seen_plans);
let state = AppState::new()
.expect("state should build")
.with_execution_runtime_sync_override_for_tests(move |plan| {
seen_plans_for_override
.lock()
.expect("mutex should lock")
.push((plan.endpoint_id.clone(), plan.candidate_id.clone()));
if plan.endpoint_id == "endpoint-retry" {
Ok(test_openai_image_execution_result(
plan,
StatusCode::TOO_MANY_REQUESTS.as_u16(),
json!({"error": {"message": "retry this candidate"}}),
))
} else {
Ok(test_openai_image_execution_result(
plan,
StatusCode::OK.as_u16(),
json!({"data": [{"b64_json": "second-candidate"}]}),
))
}
});
// Three total attempts on the sticky key; only one attempt is
// materialized up front, the other two are derived after each failure.
let attempts = vec![
test_openai_image_heartbeat_attempt_with_sticky_key_attempts(
0,
"endpoint-retry",
"candidate-retry",
3,
),
test_openai_image_heartbeat_attempt_with_sticky_key_attempts(
1,
"endpoint-success",
"candidate-success",
3,
),
];
let outcome = execute_openai_image_sync_heartbeat_attempts(
state,
"/v1/images/generations".to_string(),
"trace-image-heartbeat-sticky-retry".to_string(),
test_openai_image_heartbeat_decision(),
TEST_OPENAI_IMAGE_SYNC_PLAN_KIND.to_string(),
attempts,
ProviderTransferTracker::default(),
Instant::now(),
)
.await
.expect("heartbeat attempts should execute");
let LocalExecutionRequestOutcome::Responded(response) = outcome else {
panic!("second candidate should return a response");
};
let bytes = openai_image_sync_heartbeat_response_body_bytes(response).await;
let body: Value = serde_json::from_slice(&bytes).expect("body should decode");
let seen_plans = seen_plans.lock().expect("mutex should lock").clone();
assert_eq!(
seen_plans
.iter()
.map(|(endpoint_id, _)| endpoint_id.as_str())
.collect::<Vec<_>>(),
[
"endpoint-retry",
"endpoint-retry",
"endpoint-retry",
"endpoint-success"
]
);
let sticky_candidate_ids = seen_plans[..3]
.iter()
.map(|(_, candidate_id)| candidate_id.clone())
.collect::<std::collections::BTreeSet<_>>();
assert_eq!(
sticky_candidate_ids.len(),
3,
"each derived same-key retry must carry a fresh candidate id"
);
assert_eq!(body, json!({"data": [{"b64_json": "second-candidate"}]}));
}
#[tokio::test]
async fn openai_image_sync_heartbeat_honors_provider_transfer_limit() {
let call_count = Arc::new(AtomicUsize::new(0));
@@ -2013,6 +2117,7 @@ mod tests {
attempt.report_context = Some(json!({
"candidate_index": index,
"retry_index": 0,
"sticky_key_attempts": 1,
"local_failover_policy": {
"max_transfer_count": 1,
"max_transfer_timeout_seconds": 0
+180 -302
View File
@@ -1,8 +1,9 @@
use aether_ai_serving::{AiExecutionAttempt, STICKY_KEY_ATTEMPTS_REPORT_FIELD};
use aether_routing_core::DEFAULT_STICKY_KEY_ATTEMPTS;
use aether_runtime_state::RuntimeLockLease;
use aether_scheduler_core::parse_request_candidate_report_context;
use serde_json::Value;
use crate::provider_transport::GatewayProviderTransportSnapshot;
use uuid::Uuid;
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) struct ExecutionAttemptIdentity {
@@ -32,6 +33,9 @@ pub(crate) struct LocalExecutionCandidateMetadata {
pub(crate) pool_key_index: Option<u32>,
pub(crate) pool_key_lease: Option<RuntimeLockLease>,
pub(crate) scheduler_affinity_epoch: Option<u64>,
/// Routing-policy `sticky_key_attempts` in effect for this request. `None`
/// means the policy default applies.
pub(crate) sticky_key_attempts: Option<u32>,
}
pub(crate) const SCHEDULER_AFFINITY_EPOCH_REPORT_FIELD: &str = "scheduler_affinity_epoch";
@@ -42,6 +46,11 @@ pub(crate) const POOL_KEY_LEASE_TOKEN_REPORT_FIELD: &str = "pool_key_lease_token
pub(crate) const POOL_KEY_LEASE_FENCING_REPORT_FIELD: &str = "pool_key_lease_fencing_token";
pub(crate) const POOL_KEY_LEASE_TTL_MS_REPORT_FIELD: &str = "pool_key_lease_ttl_ms";
/// Pool-expanded keys encode `pool_key_index * STRIDE + retry_index` into the
/// persisted `retry_index` so a pool group's keys stay ordered in one candidate
/// slot. Same-key retries on a pool key are therefore bounded by the stride.
pub(crate) const POOL_KEY_RETRY_INDEX_STRIDE: u32 = 100;
pub(crate) fn attempt_identity_from_report_context(
report_context: Option<&Value>,
) -> Option<ExecutionAttemptIdentity> {
@@ -72,6 +81,10 @@ pub(crate) fn local_execution_candidate_metadata_from_report_context(
scheduler_affinity_epoch: report_context
.and_then(|value| value.get(SCHEDULER_AFFINITY_EPOCH_REPORT_FIELD))
.and_then(Value::as_u64),
sticky_key_attempts: report_context
.and_then(|value| value.get(STICKY_KEY_ATTEMPTS_REPORT_FIELD))
.and_then(Value::as_u64)
.and_then(|value| u32::try_from(value).ok()),
}
}
@@ -140,58 +153,55 @@ fn pool_key_lease_from_report_context(report_context: Option<&Value>) -> Option<
})
}
pub(crate) fn build_local_attempt_identities(
candidate_index: u32,
transport: &GatewayProviderTransportSnapshot,
) -> Vec<ExecutionAttemptIdentity> {
let attempt_slots = local_attempt_slot_count(transport);
(0..attempt_slots)
.map(|retry_index| ExecutionAttemptIdentity::new(candidate_index, retry_index))
.collect()
/// Retry index of the next same-key attempt, or `None` when the sticky-key
/// budget for this candidate is used up.
///
/// Only the first-ranked candidate (index `0`, the cache-affinity sticky key)
/// is retried on the same key; every later candidate gets exactly one attempt
/// so that once failover has started it keeps advancing. `sticky_key_attempts`
/// is the *total* attempt count on that key: `2` means one retry, `0` and `1`
/// mean none. There is no upper bound: attempts are derived one at a time
/// after each failure, never materialized ahead of time.
///
/// Inside a pool group only the first key (`pool_key_index == 0`) is treated
/// as sticky, and its retries stay below `POOL_KEY_RETRY_INDEX_STRIDE` so the
/// encoded retry index never collides with the next pool key.
pub(crate) fn next_same_key_retry_index(
identity: ExecutionAttemptIdentity,
sticky_key_attempts: Option<u32>,
) -> Option<u32> {
if identity.candidate_index != 0 {
return None;
}
let pool_limit = match identity.pool_key_index {
None => u32::MAX,
Some(0) => POOL_KEY_RETRY_INDEX_STRIDE,
Some(_) => return None,
};
let budget = sticky_key_attempts.unwrap_or(DEFAULT_STICKY_KEY_ATTEMPTS);
let attempts_so_far = identity.retry_index.checked_add(1)?;
if attempts_so_far >= budget || attempts_so_far >= pool_limit {
return None;
}
Some(attempts_so_far)
}
pub(crate) fn local_attempt_slot_count(transport: &GatewayProviderTransportSnapshot) -> u32 {
local_attempt_slots_from_transport(transport).unwrap_or(1)
}
/// For endpoint/provider table fields, `2` is the legacy admin default and is
/// treated as "not explicitly configured" so existing local-execution behaviour
/// (one attempt slot per candidate) stays unchanged. Values `0`, `1`, and `>2`
/// are treated as explicit.
const LEGACY_DEFAULT_MAX_RETRIES: u32 = 2;
/// Upper bound on local attempt slots. This is intentionally stricter than
/// admin max_retries validation to prevent unbounded pre-materialization from
/// arbitrarily large JSON config values.
const MAX_LOCAL_ATTEMPT_SLOTS: u32 = 99;
fn local_attempt_slots_from_transport(transport: &GatewayProviderTransportSnapshot) -> Option<u32> {
let rules = transport
.provider
.config
.as_ref()
.and_then(|config| config.get("failover_rules"))
.and_then(Value::as_object);
rules
.and_then(|value| value.get("max_retries"))
.and_then(Value::as_u64)
.and_then(|value| u32::try_from(value).ok())
.or_else(|| {
transport
.endpoint
.max_retries
.and_then(|value| u32::try_from(value).ok())
.filter(|&value| value != LEGACY_DEFAULT_MAX_RETRIES)
})
.or_else(|| {
transport
.provider
.max_retries
.and_then(|value| u32::try_from(value).ok())
.filter(|&value| value != LEGACY_DEFAULT_MAX_RETRIES)
})
.map(|value| value.clamp(1, MAX_LOCAL_ATTEMPT_SLOTS))
/// Derive the next same-key attempt for `attempt` after a candidate-scoped
/// failure, reading the attempt identity and sticky budget from its report
/// context. Returns `None` when no further same-key retry is allowed.
pub(crate) fn next_same_key_retry_attempt<A: AiExecutionAttempt>(attempt: &A) -> Option<A> {
let owned_report_context = attempt
.report_context_ref()
.is_none()
.then(|| attempt.report_context())
.flatten();
let report_context = attempt
.report_context_ref()
.or(owned_report_context.as_ref());
let identity = attempt_identity_from_report_context(report_context)?;
let metadata = local_execution_candidate_metadata_from_report_context(report_context);
let retry_index = next_same_key_retry_index(identity, metadata.sticky_key_attempts)?;
attempt.with_same_key_retry(retry_index, Uuid::new_v4().to_string())
}
#[cfg(test)]
@@ -199,277 +209,143 @@ mod tests {
use serde_json::json;
use super::{
attempt_identity_from_report_context, build_local_attempt_identities,
local_execution_candidate_metadata_from_report_context, ExecutionAttemptIdentity,
LocalExecutionCandidateMetadata,
};
use crate::provider_transport::snapshot::{
GatewayProviderTransportEndpoint, GatewayProviderTransportKey,
GatewayProviderTransportProvider, GatewayProviderTransportSnapshot,
attempt_identity_from_report_context,
local_execution_candidate_metadata_from_report_context, next_same_key_retry_attempt,
next_same_key_retry_index, ExecutionAttemptIdentity, LocalExecutionCandidateMetadata,
POOL_KEY_RETRY_INDEX_STRIDE,
};
use aether_ai_serving::{AiExecutionAttempt, AiSyncAttempt};
use aether_runtime_state::RuntimeLockLease;
fn sample_transport(
provider_max_retries: Option<i32>,
endpoint_max_retries: Option<i32>,
provider_config: Option<serde_json::Value>,
) -> GatewayProviderTransportSnapshot {
GatewayProviderTransportSnapshot {
provider: GatewayProviderTransportProvider {
id: "provider-1".to_string(),
name: "OpenAI".to_string(),
provider_type: "llm".to_string(),
website: None,
is_active: true,
keep_priority_on_conversion: false,
enable_format_conversion: true,
concurrent_limit: None,
max_retries: provider_max_retries,
proxy: None,
request_timeout_secs: None,
stream_first_byte_timeout_secs: None,
config: provider_config,
},
endpoint: GatewayProviderTransportEndpoint {
id: "endpoint-1".to_string(),
provider_id: "provider-1".to_string(),
api_format: "openai:chat".to_string(),
api_family: Some("openai".to_string()),
endpoint_kind: Some("chat".to_string()),
is_active: true,
base_url: "https://example.com".to_string(),
header_rules: None,
body_rules: None,
max_retries: endpoint_max_retries,
custom_path: None,
config: None,
format_acceptance_config: None,
proxy: None,
},
key: GatewayProviderTransportKey {
id: "key-1".to_string(),
provider_id: "provider-1".to_string(),
name: "primary".to_string(),
auth_type: "bearer".to_string(),
is_active: true,
api_formats: None,
auth_type_by_format: None,
allow_auth_channel_mismatch_formats: None,
#[test]
fn first_candidate_defaults_to_one_same_key_retry() {
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(0, 0), None),
Some(1)
);
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(0, 1), None),
None
);
}
allowed_models: None,
capabilities: None,
rate_multipliers: None,
global_priority_by_format: None,
expires_at_unix_secs: None,
proxy: None,
fingerprint: None,
upstream_metadata: None,
decrypted_api_key: "secret".to_string(),
decrypted_auth_config: None,
},
#[test]
fn first_candidate_uses_policy_sticky_key_attempts_without_upper_bound() {
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(0, 2), Some(3)),
None
);
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(0, 1), Some(3)),
Some(2)
);
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(0, 4_999), Some(10_000)),
Some(5_000)
);
}
#[test]
fn zero_and_one_mean_single_attempt() {
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(0, 0), Some(0)),
None
);
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(0, 0), Some(1)),
None
);
}
#[test]
fn failover_candidates_never_retry_on_the_same_key() {
for candidate_index in 1..5 {
assert_eq!(
next_same_key_retry_index(ExecutionAttemptIdentity::new(candidate_index, 0), None),
None
);
assert_eq!(
next_same_key_retry_index(
ExecutionAttemptIdentity::new(candidate_index, 0),
Some(50)
),
None
);
}
}
#[test]
fn build_local_attempt_identities_defaults_to_single_attempt() {
let identities = build_local_attempt_identities(3, &sample_transport(None, None, None));
assert_eq!(identities, vec![ExecutionAttemptIdentity::new(3, 0)]);
}
#[test]
fn build_local_attempt_identities_prefer_failover_rules_over_endpoint_and_provider() {
let identities = build_local_attempt_identities(
1,
&sample_transport(
Some(5),
Some(4),
Some(json!({
"failover_rules": {
"max_retries": 2
}
})),
),
);
fn pool_groups_only_retry_their_first_key_within_the_stride() {
let first_pool_key = ExecutionAttemptIdentity::new(0, 0).with_pool_key_index(Some(0));
assert_eq!(next_same_key_retry_index(first_pool_key, Some(3)), Some(1));
let at_stride_limit = ExecutionAttemptIdentity::new(0, POOL_KEY_RETRY_INDEX_STRIDE - 1)
.with_pool_key_index(Some(0));
assert_eq!(
identities,
vec![
ExecutionAttemptIdentity::new(1, 0),
ExecutionAttemptIdentity::new(1, 1),
]
next_same_key_retry_index(at_stride_limit, Some(10_000)),
None
);
}
#[test]
fn build_local_attempt_identities_falls_back_to_endpoint_max_retries() {
let identities =
build_local_attempt_identities(2, &sample_transport(Some(5), Some(3), None));
let second_pool_key = ExecutionAttemptIdentity::new(0, POOL_KEY_RETRY_INDEX_STRIDE)
.with_pool_key_index(Some(1));
assert_eq!(
identities,
vec![
ExecutionAttemptIdentity::new(2, 0),
ExecutionAttemptIdentity::new(2, 1),
ExecutionAttemptIdentity::new(2, 2),
]
next_same_key_retry_index(second_pool_key, Some(10_000)),
None
);
}
#[test]
fn build_local_attempt_identities_falls_back_to_provider_max_retries() {
let identities = build_local_attempt_identities(0, &sample_transport(Some(4), None, None));
fn next_same_key_retry_attempt_rewrites_candidate_id_and_retry_index() {
let attempt = AiSyncAttempt {
plan: aether_contracts::ExecutionPlan {
request_id: "trace-1".to_string(),
candidate_id: Some("candidate-a".to_string()),
provider_name: None,
provider_id: "provider-1".to_string(),
endpoint_id: "endpoint-1".to_string(),
key_id: "key-1".to_string(),
method: "POST".to_string(),
url: "https://example.com".to_string(),
headers: Default::default(),
content_type: None,
content_encoding: None,
body: aether_contracts::RequestBody {
json_body: None,
body_bytes_b64: None,
body_ref: None,
},
stream: false,
client_api_format: "openai:chat".to_string(),
provider_api_format: "openai:chat".to_string(),
model_name: None,
proxy: None,
transport_profile: None,
timeouts: None,
},
report_kind: None,
report_context: Some(json!({
"candidate_id": "candidate-a",
"candidate_index": 0,
"retry_index": 0,
"sticky_key_attempts": 2,
})),
};
assert_eq!(
identities,
vec![
ExecutionAttemptIdentity::new(0, 0),
ExecutionAttemptIdentity::new(0, 1),
ExecutionAttemptIdentity::new(0, 2),
ExecutionAttemptIdentity::new(0, 3),
]
let retry = next_same_key_retry_attempt(&attempt).expect("one same-key retry remains");
let retry_candidate_id = retry.plan.candidate_id.clone().expect("fresh candidate id");
assert_ne!(retry_candidate_id, "candidate-a");
assert_eq!(retry.plan.key_id, "key-1");
let context = retry.report_context_ref().expect("context retained");
assert_eq!(context["candidate_id"], json!(retry_candidate_id));
assert_eq!(context["retry_index"], json!(1));
assert_eq!(context["candidate_index"], json!(0));
assert!(
next_same_key_retry_attempt(&retry).is_none(),
"budget of 2 attempts is exhausted after one retry"
);
}
#[test]
fn build_local_attempt_identities_endpoint_overrides_provider() {
let identities =
build_local_attempt_identities(7, &sample_transport(Some(10), Some(3), None));
assert_eq!(
identities,
vec![
ExecutionAttemptIdentity::new(7, 0),
ExecutionAttemptIdentity::new(7, 1),
ExecutionAttemptIdentity::new(7, 2),
]
);
}
#[test]
fn build_local_attempt_identities_default_two_treated_as_unset() {
let identities =
build_local_attempt_identities(5, &sample_transport(Some(2), Some(2), None));
assert_eq!(identities, vec![ExecutionAttemptIdentity::new(5, 0)]);
}
#[test]
fn build_local_attempt_identities_endpoint_two_falls_back_to_provider_ten() {
let identities =
build_local_attempt_identities(1, &sample_transport(Some(10), Some(2), None));
assert_eq!(
identities,
vec![
ExecutionAttemptIdentity::new(1, 0),
ExecutionAttemptIdentity::new(1, 1),
ExecutionAttemptIdentity::new(1, 2),
ExecutionAttemptIdentity::new(1, 3),
ExecutionAttemptIdentity::new(1, 4),
ExecutionAttemptIdentity::new(1, 5),
ExecutionAttemptIdentity::new(1, 6),
ExecutionAttemptIdentity::new(1, 7),
ExecutionAttemptIdentity::new(1, 8),
ExecutionAttemptIdentity::new(1, 9),
]
);
}
#[test]
fn build_local_attempt_identities_failover_rules_zero_produces_one_slot() {
let identities = build_local_attempt_identities(
1,
&sample_transport(
Some(5),
Some(4),
Some(json!({
"failover_rules": {
"max_retries": 0
}
})),
),
);
assert_eq!(identities, vec![ExecutionAttemptIdentity::new(1, 0)]);
}
#[test]
fn build_local_attempt_identities_endpoint_zero_produces_one_slot() {
let identities =
build_local_attempt_identities(3, &sample_transport(Some(5), Some(0), None));
assert_eq!(identities, vec![ExecutionAttemptIdentity::new(3, 0)]);
}
#[test]
fn build_local_attempt_identities_provider_zero_produces_one_slot() {
let identities = build_local_attempt_identities(3, &sample_transport(Some(0), None, None));
assert_eq!(identities, vec![ExecutionAttemptIdentity::new(3, 0)]);
}
#[test]
fn build_local_attempt_identities_provider_ten_creates_ten_slots() {
let identities = build_local_attempt_identities(2, &sample_transport(Some(10), None, None));
assert_eq!(identities.len(), 10);
assert_eq!(identities[0], ExecutionAttemptIdentity::new(2, 0));
assert_eq!(identities[9], ExecutionAttemptIdentity::new(2, 9));
}
#[test]
fn build_local_attempt_identities_failover_rules_over_limit_clamped_to_max() {
let identities = build_local_attempt_identities(
0,
&sample_transport(
Some(3),
Some(5),
Some(json!({
"failover_rules": {
"max_retries": 1000
}
})),
),
);
assert_eq!(identities.len(), 99);
}
#[test]
fn build_local_attempt_identities_failover_rules_u32_max_clamped_to_max() {
let identities = build_local_attempt_identities(
0,
&sample_transport(
None,
None,
Some(json!({
"failover_rules": {
"max_retries": u32::MAX
}
})),
),
);
assert_eq!(identities.len(), 99);
}
#[test]
fn build_local_attempt_identities_endpoint_over_limit_clamped_to_max() {
let identities =
build_local_attempt_identities(0, &sample_transport(None, Some(2000), None));
assert_eq!(identities.len(), 99);
}
#[test]
fn build_local_attempt_identities_provider_over_limit_clamped_to_max() {
let identities =
build_local_attempt_identities(0, &sample_transport(Some(5000), None, None));
assert_eq!(identities.len(), 99);
}
#[test]
fn parse_attempt_identity_from_report_context_reads_candidate_and_retry_indices() {
let identity = attempt_identity_from_report_context(Some(&json!({
@@ -499,6 +375,7 @@ mod tests {
"pool_key_lease_token": "gateway-1:token-1",
"pool_key_lease_fencing_token": 7,
"pool_key_lease_ttl_ms": 900000,
"sticky_key_attempts": 3,
})));
assert_eq!(
@@ -514,6 +391,7 @@ mod tests {
ttl_ms: 900000,
}),
scheduler_affinity_epoch: None,
sticky_key_attempts: Some(3),
}
);
}
+4 -5
View File
@@ -20,11 +20,10 @@ pub(crate) use self::adaptive::{
LocalAdaptiveRateLimitProjection, LocalAdaptiveSuccessProjection,
};
pub(crate) use self::attempt::{
attempt_identity_from_report_context, build_local_attempt_identities,
insert_pool_key_lease_report_context_fields, local_attempt_slot_count,
local_execution_candidate_metadata_from_report_context, ExecutionAttemptIdentity,
LocalExecutionCandidateMetadata, ROUTING_POOL_POLICY_OVERRIDE_REPORT_FIELD,
SCHEDULER_AFFINITY_EPOCH_REPORT_FIELD,
attempt_identity_from_report_context, insert_pool_key_lease_report_context_fields,
local_execution_candidate_metadata_from_report_context, next_same_key_retry_attempt,
ExecutionAttemptIdentity, LocalExecutionCandidateMetadata, POOL_KEY_RETRY_INDEX_STRIDE,
ROUTING_POOL_POLICY_OVERRIDE_REPORT_FIELD, SCHEDULER_AFFINITY_EPOCH_REPORT_FIELD,
};
pub(crate) use self::classifier::{
classify_anthropic_failure_disposition, classify_failure_disposition, classify_local_failover,
+21 -16
View File
@@ -1,7 +1,7 @@
use aether_routing_core::{
resolve_routing_policy, MutationPlan, RankingOverlay, ResolvedRoutingPolicy,
RoutingGroupConfig, RoutingPolicyInput, RoutingRulePhase, RoutingSchedulingMode,
RoutingSetPriorityMode,
RoutingDefaultPolicy, RoutingGroupConfig, RoutingPolicyInput, RoutingRulePhase,
RoutingSchedulingMode, RoutingSetPriorityMode, DEFAULT_STICKY_KEY_ATTEMPTS,
};
use http::StatusCode;
use serde_json::Value;
@@ -81,9 +81,7 @@ pub(crate) fn resolve_gateway_routing_policy(
pub(crate) fn resolve_gateway_static_default_routing_policy(
input: GatewayStaticRoutingPolicyInput<'_>,
) -> Result<Option<ResolvedRoutingPolicy>, GatewayError> {
let Some((priority_mode, scheduling_mode, keep_priority_on_conversion)) =
static_default_policy_fields(input.group_config_json)?
else {
let Some(default_policy) = static_default_policy_fields(input.group_config_json)? else {
return Ok(None);
};
@@ -93,9 +91,10 @@ pub(crate) fn resolve_gateway_static_default_routing_policy(
selection_source: input.selection_source.to_string(),
requested_model: input.requested_model.to_string(),
resolved_model: input.resolved_model.to_string(),
priority_mode,
scheduling_mode,
keep_priority_on_conversion,
priority_mode: default_policy.priority_mode,
scheduling_mode: default_policy.scheduling_mode,
keep_priority_on_conversion: default_policy.keep_priority_on_conversion,
sticky_key_attempts: default_policy.sticky_key_attempts,
ranking_overlay: RankingOverlay::default(),
mutation_plan: MutationPlan::default(),
pool_policy_overrides: BTreeMap::new(),
@@ -105,7 +104,7 @@ pub(crate) fn resolve_gateway_static_default_routing_policy(
fn static_default_policy_fields(
config_json: &Value,
) -> Result<Option<(RoutingSetPriorityMode, RoutingSchedulingMode, bool)>, GatewayError> {
) -> Result<Option<RoutingDefaultPolicy>, GatewayError> {
let Some(object) = config_json.as_object() else {
return Ok(None);
};
@@ -117,11 +116,7 @@ fn static_default_policy_fields(
}
let Some(default_policy) = object.get("default_policy") else {
return Ok(Some((
RoutingSetPriorityMode::default(),
RoutingSchedulingMode::default(),
false,
)));
return Ok(Some(RoutingDefaultPolicy::default()));
};
let Some(default_policy) = default_policy.as_object() else {
return Ok(None);
@@ -141,12 +136,22 @@ fn static_default_policy_fields(
})?,
None => false,
};
let sticky_key_attempts = match default_policy.get("sticky_key_attempts") {
Some(value) => value
.as_u64()
.and_then(|value| u32::try_from(value).ok())
.ok_or_else(|| {
invalid_routing_group_config("sticky_key_attempts must be a non-negative integer")
})?,
None => DEFAULT_STICKY_KEY_ATTEMPTS,
};
Ok(Some((
Ok(Some(RoutingDefaultPolicy {
priority_mode,
scheduling_mode,
keep_priority_on_conversion,
)))
sticky_key_attempts,
}))
}
fn routing_array_field_is_missing_or_empty(
+11 -1
View File
@@ -1,6 +1,7 @@
use aether_data_contracts::repository::routing_profiles::RoutingGroupLookupKey;
use aether_routing_core::{
ResolvedRoutingPolicy, RoutingDefaultPolicy, RoutingSchedulingMode, RoutingSetPriorityMode,
DEFAULT_STICKY_KEY_ATTEMPTS,
};
use aether_scheduler_core::SchedulerPriorityMode;
use tracing::warn;
@@ -37,6 +38,8 @@ pub(crate) struct SchedulerOrderingConfig {
pub(crate) priority_mode: SchedulerPriorityMode,
pub(crate) scheduling_mode: SchedulerSchedulingMode,
pub(crate) keep_priority_on_conversion: bool,
/// Total attempts on the first-ranked (sticky) candidate before failover.
pub(crate) sticky_key_attempts: u32,
}
impl Default for SchedulerOrderingConfig {
@@ -45,6 +48,7 @@ impl Default for SchedulerOrderingConfig {
priority_mode: SchedulerPriorityMode::Provider,
scheduling_mode: SchedulerSchedulingMode::CacheAffinity,
keep_priority_on_conversion: false,
sticky_key_attempts: DEFAULT_STICKY_KEY_ATTEMPTS,
}
}
}
@@ -57,6 +61,7 @@ impl SchedulerOrderingConfig {
priority_mode: scheduler_priority_mode_from_routing(policy.priority_mode),
scheduling_mode: scheduler_scheduling_mode_from_routing(policy.scheduling_mode),
keep_priority_on_conversion: policy.keep_priority_on_conversion,
sticky_key_attempts: policy.sticky_key_attempts,
}
}
@@ -65,6 +70,7 @@ impl SchedulerOrderingConfig {
priority_mode: scheduler_priority_mode_from_routing(policy.priority_mode),
scheduling_mode: scheduler_scheduling_mode_from_routing(policy.scheduling_mode),
keep_priority_on_conversion: policy.keep_priority_on_conversion,
sticky_key_attempts: policy.sticky_key_attempts,
}
}
@@ -80,6 +86,7 @@ impl SchedulerOrderingConfig {
SchedulerSchedulingMode::LoadBalance => RoutingSchedulingMode::LoadBalance,
},
keep_priority_on_conversion: self.keep_priority_on_conversion,
sticky_key_attempts: self.sticky_key_attempts,
}
}
@@ -221,6 +228,8 @@ pub(crate) async fn read_legacy_scheduler_ordering_config(
priority_mode,
scheduling_mode,
keep_priority_on_conversion,
// Legacy config never carried a sticky-key setting; use the routing default.
sticky_key_attempts: DEFAULT_STICKY_KEY_ATTEMPTS,
})
}
@@ -358,7 +367,8 @@ mod tests {
json!({
"priority_mode": "global_key",
"scheduling_mode": "load_balance",
"keep_priority_on_conversion": true
"keep_priority_on_conversion": true,
"sticky_key_attempts": DEFAULT_STICKY_KEY_ATTEMPTS
})
);
@@ -2220,7 +2220,9 @@ async fn gateway_retries_next_local_openai_chat_stream_candidate_after_retryable
.to_string(),
});
let frames = if attempt == 1 {
// The primary key gets two attempts under the default
// sticky_key_attempts; both must fail to reach the backup.
let frames = if attempt <= 2 {
concat!(
"{\"type\":\"headers\",\"payload\":{\"kind\":\"headers\",\"status_code\":429,\"headers\":{\"content-type\":\"application/json\"}}}\n",
"{\"type\":\"data\",\"payload\":{\"kind\":\"data\",\"text\":\"{\\\"error\\\":{\\\"message\\\":\\\"rate limited\\\",\\\"type\\\":\\\"rate_limit_error\\\"}}\"}}\n",
@@ -2362,14 +2364,25 @@ async fn gateway_retries_next_local_openai_chat_stream_candidate_after_retryable
.lock()
.expect("mutex should lock")
.len()
>= 2
>= 3
})
.await;
let seen_execution_runtime_requests = seen_execution_runtime
.lock()
.expect("mutex should lock")
.clone();
assert_eq!(seen_execution_runtime_requests.len(), 2);
// Default sticky_key_attempts is 2: the primary key is retried once on
// the same key, then failover moves to the backup with a single attempt.
assert_eq!(seen_execution_runtime_requests.len(), 3);
assert_eq!(
seen_execution_runtime_requests
.iter()
.filter(|request| {
request.url == "https://api.openai.primary.example/chat/completions"
})
.count(),
2
);
let primary_request = seen_execution_runtime_requests
.iter()
.find(|request| request.url == "https://api.openai.primary.example/chat/completions")
@@ -2404,7 +2417,15 @@ async fn gateway_retries_next_local_openai_chat_stream_candidate_after_retryable
.list_by_request_id("trace-openai-chat-local-stream-failover-123")
.await
.expect("request candidate trace should read");
assert_eq!(stored_candidates.len(), 2);
assert_eq!(stored_candidates.len(), 3);
assert_eq!(
stored_candidates
.iter()
.filter(|candidate| candidate.status == RequestCandidateStatus::Failed)
.count(),
2,
"both sticky-key attempts on the primary should be recorded as failed"
);
let failed_candidate = stored_candidates
.iter()
.find(|candidate| {
@@ -2465,7 +2486,7 @@ async fn gateway_retries_next_local_openai_chat_stream_candidate_after_retryable
assert_eq!(
execution_runtime_hits.load(std::sync::atomic::Ordering::SeqCst),
2
3
);
assert_eq!(*decision_hits.lock().expect("mutex should lock"), 0);
assert_eq!(*plan_hits.lock().expect("mutex should lock"), 0);
@@ -112,7 +112,7 @@ async fn gateway_skips_unsupported_local_openai_chat_sync_candidate_before_tryin
false,
false,
None,
Some(2),
Some(1),
None,
Some(20.0),
None,
@@ -134,7 +134,7 @@ async fn gateway_skips_unsupported_local_openai_chat_sync_candidate_before_tryin
"https://api.openai.skip.example".to_string(),
None,
None,
Some(2),
Some(1),
None,
None,
None,
@@ -520,7 +520,7 @@ async fn gateway_surfaces_local_execution_runtime_miss_reason_when_all_openai_ch
false,
false,
None,
Some(2),
Some(1),
None,
Some(20.0),
None,
@@ -542,7 +542,7 @@ async fn gateway_surfaces_local_execution_runtime_miss_reason_when_all_openai_ch
"https://chatgpt.com/backend-api/codex".to_string(),
None,
None,
Some(2),
Some(1),
None,
None,
None,
@@ -802,7 +802,7 @@ async fn gateway_retries_next_local_openai_chat_sync_candidate_after_auth_failur
false,
false,
None,
Some(2),
Some(1),
None,
Some(20.0),
None,
@@ -828,7 +828,7 @@ async fn gateway_retries_next_local_openai_chat_sync_candidate_after_auth_failur
base_url.to_string(),
None,
None,
Some(2),
Some(1),
None,
None,
None,
@@ -970,7 +970,9 @@ async fn gateway_retries_next_local_openai_chat_sync_candidate_after_auth_failur
.to_string(),
});
if attempt == 1 {
// The primary key gets two attempts under the default
// sticky_key_attempts; both must fail to reach the backup.
if attempt <= 2 {
return Json(json!({
"request_id": "trace-openai-chat-local-failover-123",
"status_code": 401,
@@ -1125,56 +1127,63 @@ async fn gateway_retries_next_local_openai_chat_sync_candidate_after_auth_failur
.lock()
.expect("mutex should lock")
.clone();
assert_eq!(seen_execution_runtime_requests.len(), 2);
// Default sticky_key_attempts is 2: the primary key is retried once on
// the same key, then failover moves to the backup with a single attempt.
assert_eq!(seen_execution_runtime_requests.len(), 3);
for primary_request in &seen_execution_runtime_requests[..2] {
assert_eq!(
primary_request.trace_id,
"trace-openai-chat-local-failover-123"
);
assert_eq!(
primary_request.url,
"https://api.openai.primary.example/chat/completions"
);
assert_eq!(
primary_request.authorization,
"Bearer sk-upstream-openai-primary"
);
}
assert_eq!(
seen_execution_runtime_requests[0].trace_id,
"trace-openai-chat-local-failover-123"
);
assert_eq!(
seen_execution_runtime_requests[0].url,
"https://api.openai.primary.example/chat/completions"
);
assert_eq!(
seen_execution_runtime_requests[0].authorization,
"Bearer sk-upstream-openai-primary"
);
assert_eq!(
seen_execution_runtime_requests[1].url,
seen_execution_runtime_requests[2].url,
"https://api.openai.backup.example/chat/completions"
);
assert_eq!(
seen_execution_runtime_requests[1].model,
seen_execution_runtime_requests[2].model,
"gpt-5-upstream-backup"
);
assert_eq!(
seen_execution_runtime_requests[1].authorization,
seen_execution_runtime_requests[2].authorization,
"Bearer sk-upstream-openai-backup"
);
let stored_candidates = request_candidate_repository
.list_by_request_id("trace-openai-chat-local-failover-123")
.await
.expect("request candidate trace should read");
assert_eq!(stored_candidates.len(), 2);
assert_eq!(stored_candidates[0].candidate_index, 0);
assert_eq!(stored_candidates[0].status, RequestCandidateStatus::Failed);
assert_eq!(stored_candidates[0].status_code, Some(401));
assert_eq!(
stored_candidates[0].error_message.as_deref(),
Some("invalid auth token")
);
let failed_upstream_response = stored_candidates[0]
.extra_data
.as_ref()
.and_then(|value| value.get("upstream_response"))
.expect("failed candidate should keep its upstream response");
assert_eq!(failed_upstream_response["status_code"], json!(401));
assert_eq!(
failed_upstream_response["body"]["error"]["message"],
json!("invalid auth token")
);
assert_eq!(stored_candidates[1].candidate_index, 1);
assert_eq!(stored_candidates[1].status, RequestCandidateStatus::Success);
assert_eq!(stored_candidates[1].status_code, Some(200));
assert_eq!(stored_candidates.len(), 3);
for (retry_index, failed_candidate) in stored_candidates[..2].iter().enumerate() {
assert_eq!(failed_candidate.candidate_index, 0);
assert_eq!(failed_candidate.retry_index, retry_index as u32);
assert_eq!(failed_candidate.status, RequestCandidateStatus::Failed);
assert_eq!(failed_candidate.status_code, Some(401));
assert_eq!(
failed_candidate.error_message.as_deref(),
Some("invalid auth token")
);
let failed_upstream_response = failed_candidate
.extra_data
.as_ref()
.and_then(|value| value.get("upstream_response"))
.expect("failed candidate should keep its upstream response");
assert_eq!(failed_upstream_response["status_code"], json!(401));
assert_eq!(
failed_upstream_response["body"]["error"]["message"],
json!("invalid auth token")
);
}
assert_eq!(stored_candidates[2].candidate_index, 1);
assert_eq!(stored_candidates[2].status, RequestCandidateStatus::Success);
assert_eq!(stored_candidates[2].status_code, Some(200));
tokio::time::sleep(std::time::Duration::from_millis(100)).await;
assert!(
@@ -1184,7 +1193,7 @@ async fn gateway_retries_next_local_openai_chat_sync_candidate_after_auth_failur
assert_eq!(
*execution_runtime_hits.lock().expect("mutex should lock"),
2
3
);
assert_eq!(*decision_hits.lock().expect("mutex should lock"), 0);
assert_eq!(*plan_hits.lock().expect("mutex should lock"), 0);
@@ -140,7 +140,7 @@ async fn gateway_executes_codex_search_with_responses_permission_and_search_cont
false,
false,
None,
Some(2),
Some(1),
None,
Some(900.0),
None,
@@ -162,7 +162,7 @@ async fn gateway_executes_codex_search_with_responses_permission_and_search_cont
"https://chatgpt.com/backend-api/codex".to_string(),
None,
None,
Some(2),
Some(1),
None,
None,
None,
@@ -570,9 +570,12 @@ async fn gateway_executes_codex_search_with_responses_permission_and_search_cont
.filter(|plan| plan["request_id"] == "trace-search-failover-1")
.map(|plan| plan["provider_id"].clone())
.collect::<Vec<_>>();
// Default sticky_key_attempts is 2: the first provider is retried once on
// the same key before failover advances to the second provider.
assert_eq!(
failover_plans,
vec![
json!("provider-codex-search-1"),
json!("provider-codex-search-1"),
json!("provider-codex-search-2")
]
@@ -581,17 +584,16 @@ async fn gateway_executes_codex_search_with_responses_permission_and_search_cont
.list_by_request_id("trace-search-failover-1")
.await
.expect("failover request candidates should read");
assert_eq!(failover_candidates.len(), 2);
assert_eq!(failover_candidates.len(), 3);
for failed_candidate in &failover_candidates[..2] {
assert_eq!(failed_candidate.status, RequestCandidateStatus::Failed);
assert_eq!(failed_candidate.status_code, Some(500));
}
assert_eq!(
failover_candidates[0].status,
RequestCandidateStatus::Failed
);
assert_eq!(failover_candidates[0].status_code, Some(500));
assert_eq!(
failover_candidates[1].status,
failover_candidates[2].status,
RequestCandidateStatus::Success
);
assert_eq!(failover_candidates[1].status_code, Some(200));
assert_eq!(failover_candidates[2].status_code, Some(200));
gateway_handle.abort();
execution_runtime_handle.abort();
@@ -1714,7 +1714,6 @@ fn ai_serving_candidate_materialization_owns_affinity_and_candidate_runtime_pers
"pub fn ai_should_persist_available_candidate_for_pool_key",
"pub fn ai_should_persist_skipped_candidate_for_pool_membership",
"pub fn ai_candidate_extra_data_with_ranking",
"attempt_slot_count",
"should_persist_available_candidate",
"persist_available_candidate",
"build_attempt",
+2 -2
View File
@@ -122,7 +122,7 @@ pub(super) fn sample_local_openai_provider() -> StoredProviderCatalogProvider {
false,
false,
None,
Some(2),
Some(1),
None,
Some(20.0),
None,
@@ -144,7 +144,7 @@ pub(super) fn sample_local_openai_endpoint() -> StoredProviderCatalogEndpoint {
"https://api.openai.example/v1".to_string(),
None,
None,
Some(2),
Some(1),
None,
None,
None,
+22 -17
View File
@@ -879,9 +879,13 @@ async fn gateway_records_failed_usage_when_all_local_openai_chat_candidates_exha
.list_by_request_id("trace-openai-chat-local-report-sync-failure-123")
.await
.expect("request candidate trace should read");
assert_eq!(stored_candidates.len(), 1);
assert_eq!(stored_candidates[0].status, RequestCandidateStatus::Failed);
assert_eq!(stored_candidates[0].status_code, Some(503));
// The only candidate is the sticky first key: the default policy retries
// it once on the same key before the request is exhausted.
assert_eq!(stored_candidates.len(), 2);
for candidate in &stored_candidates {
assert_eq!(candidate.status, RequestCandidateStatus::Failed);
assert_eq!(candidate.status_code, Some(503));
}
}
#[test]
@@ -957,7 +961,8 @@ async fn gateway_records_failed_usage_when_sync_runtime_transport_is_unavailable
let response = send_request(gateway, request).await;
assert_eq!(response.status(), StatusCode::SERVICE_UNAVAILABLE);
assert_eq!(*execution_hits.lock().expect("mutex should lock"), 1);
// The sticky first key is retried once on the same key before exhaustion.
assert_eq!(*execution_hits.lock().expect("mutex should lock"), 2);
let stored_usage = wait_for_usage_status(
usage_repository.as_ref(),
@@ -982,15 +987,15 @@ async fn gateway_records_failed_usage_when_sync_runtime_transport_is_unavailable
.list_by_request_id("trace-openai-chat-local-transport-unavailable-123")
.await
.expect("request candidate trace should read");
assert_eq!(stored_candidates.len(), 1);
assert_eq!(stored_candidates[0].status, RequestCandidateStatus::Failed);
assert!(stored_candidates[0]
.latency_ms
.is_some_and(|value| value >= 5));
assert_eq!(
stored_candidates[0].error_type.as_deref(),
Some("execution_runtime_unavailable")
);
assert_eq!(stored_candidates.len(), 2);
for candidate in &stored_candidates {
assert_eq!(candidate.status, RequestCandidateStatus::Failed);
assert!(candidate.latency_ms.is_some_and(|value| value >= 5));
assert_eq!(
candidate.error_type.as_deref(),
Some("execution_runtime_unavailable")
);
}
}
#[test]
@@ -1786,7 +1791,7 @@ async fn gateway_records_failed_usage_when_all_local_claude_cli_candidates_are_s
false,
false,
None,
Some(2),
Some(1),
None,
Some(20.0),
None,
@@ -1808,7 +1813,7 @@ async fn gateway_records_failed_usage_when_all_local_claude_cli_candidates_are_s
"https://right.codes/codex".to_string(),
None,
None,
Some(2),
Some(1),
Some("/v1/messages".to_string()),
None,
None,
@@ -2111,7 +2116,7 @@ fn gateway_keeps_failed_usage_request_capture_lightweight_for_large_local_claude
false,
false,
None,
Some(2),
Some(1),
None,
Some(20.0),
None,
@@ -2133,7 +2138,7 @@ fn gateway_keeps_failed_usage_request_capture_lightweight_for_large_local_claude
"https://right.codes/codex".to_string(),
None,
None,
Some(2),
Some(1),
Some("/v1/messages".to_string()),
None,
None,