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)