Disable key circuit breaker for pool providers

This commit is contained in:
fawney19
2026-05-29 15:16:06 +08:00
parent 9fb600df1b
commit 1658925f52
12 changed files with 252 additions and 129 deletions
@@ -16,6 +16,6 @@ pub(crate) use self::reads::{
};
pub(crate) use self::status::build_admin_provider_pool_status_payload;
pub(crate) use self::writes::{
admin_provider_pool_key_circuit_breaker_reason, record_admin_provider_pool_error,
admin_provider_pool_key_terminal_error_reason, record_admin_provider_pool_error,
record_admin_provider_pool_stream_timeout, record_admin_provider_pool_success,
};
@@ -222,7 +222,7 @@ fn first_error_text(
})
}
pub(crate) fn admin_provider_pool_key_circuit_breaker_reason(
pub(crate) fn admin_provider_pool_key_terminal_error_reason(
status_code: u16,
error_body: Option<&str>,
) -> Option<String> {
@@ -475,7 +475,7 @@ pub(crate) async fn record_admin_provider_pool_error(
if status_code == 400 {
// Bad Request is usually attributable to the caller payload, not key health.
// Account-level 400s are handled by the orchestration circuit-breaker path.
// Account-level 400s are handled by orchestration pool-score feedback.
return;
}
@@ -588,7 +588,7 @@ pub(crate) async fn record_admin_provider_pool_stream_timeout(
#[cfg(test)]
mod tests {
use super::{
admin_provider_pool_key_circuit_breaker_reason, parse_google_quota_cooldown_seconds_at,
admin_provider_pool_key_terminal_error_reason, parse_google_quota_cooldown_seconds_at,
record_admin_provider_pool_error, record_admin_provider_pool_stream_timeout,
record_admin_provider_pool_success,
};
@@ -747,9 +747,9 @@ mod tests {
}
#[test]
fn circuit_reason_detects_workspace_deactivated_errors() {
fn terminal_error_reason_detects_workspace_deactivated_errors() {
assert_eq!(
admin_provider_pool_key_circuit_breaker_reason(
admin_provider_pool_key_terminal_error_reason(
402,
Some(r#"{"error":{"message":"workspace has been deactivated"}}"#),
)
@@ -757,7 +757,7 @@ mod tests {
Some("workspace_deactivated_402:workspace has been deactivated")
);
assert_eq!(
admin_provider_pool_key_circuit_breaker_reason(
admin_provider_pool_key_terminal_error_reason(
400,
Some(r#"{"error":{"message":"deactivated_workspace"}}"#),
)
@@ -767,9 +767,9 @@ mod tests {
}
#[test]
fn circuit_reason_detects_account_ban_errors() {
fn terminal_error_reason_detects_account_ban_errors() {
assert_eq!(
admin_provider_pool_key_circuit_breaker_reason(
admin_provider_pool_key_terminal_error_reason(
403,
Some(r#"{"error":{"message":"AccountSuspendedException: account suspended"}}"#),
)
@@ -777,7 +777,7 @@ mod tests {
Some("forbidden_403")
);
assert_eq!(
admin_provider_pool_key_circuit_breaker_reason(
admin_provider_pool_key_terminal_error_reason(
423,
Some(r#"{"error":{"message":"account access denied"}}"#),
)
@@ -1099,7 +1099,7 @@ mod tests {
}
#[tokio::test]
async fn severe_account_errors_use_circuit_breaker_instead_of_pool_cooldown() {
async fn severe_account_errors_do_not_use_pool_cooldown() {
let Some(redis) = start_managed_redis_or_skip().await else {
return;
};
@@ -1109,7 +1109,7 @@ mod tests {
let key_ids = vec!["key-account-disabled".to_string()];
assert_eq!(
admin_provider_pool_key_circuit_breaker_reason(
admin_provider_pool_key_terminal_error_reason(
401,
Some(r#"{"error":{"message":"account has been deactivated"}}"#),
)
@@ -13,7 +13,6 @@ use aether_data_contracts::repository::provider_catalog::{
StoredProviderCatalogEndpoint, StoredProviderCatalogKey,
};
use aether_data_contracts::repository::usage::StoredProviderApiKeyWindowUsageSummary;
use aether_scheduler_core::provider_key_circuit_payload_is_active_open_at;
use serde_json::json;
use std::collections::BTreeMap;
use std::time::{SystemTime, UNIX_EPOCH};
@@ -934,18 +933,6 @@ fn admin_pool_health_score(key: &StoredProviderCatalogKey) -> f64 {
}
}
fn admin_pool_circuit_breaker_open(key: &StoredProviderCatalogKey, now_unix_secs: u64) -> bool {
key.circuit_breaker_by_format
.as_ref()
.and_then(serde_json::Value::as_object)
.map(|formats| {
formats
.values()
.any(|item| provider_key_circuit_payload_is_active_open_at(item, now_unix_secs))
})
.unwrap_or(false)
}
fn admin_pool_scheduling_payload(
key: &StoredProviderCatalogKey,
cooldown_reason: Option<&str>,
@@ -1040,7 +1027,7 @@ pub(super) fn build_admin_pool_key_payload(
.as_ref()
.and_then(|_| runtime.cooldown_ttl_by_key.get(&key.id).copied());
let health_score = admin_pool_health_score(key);
let circuit_breaker_open = admin_pool_circuit_breaker_open(key, now_unix_secs);
let circuit_breaker_open = false;
let auth_semantics = provider_key_auth_semantics(key, provider_type);
let account_quota_exhausted = pool_config
.as_ref()
@@ -2,7 +2,7 @@ pub(crate) use super::super::admin::provider::pool::config::{
admin_provider_pool_cache_affinity_enabled, admin_provider_pool_config_from_config_value,
};
pub(crate) use super::super::admin::provider::pool::runtime::{
admin_provider_pool_key_circuit_breaker_reason, read_admin_provider_pool_key_cooldown_reason,
admin_provider_pool_key_terminal_error_reason, read_admin_provider_pool_key_cooldown_reason,
read_admin_provider_pool_runtime_state, record_admin_provider_pool_error,
record_admin_provider_pool_stream_timeout, record_admin_provider_pool_success,
release_admin_provider_pool_key_lease,
+143 -58
View File
@@ -19,8 +19,7 @@ use tracing::warn;
use super::{
local_failover_error_message, project_local_adaptive_rate_limit,
project_local_adaptive_success, project_local_failure_health, project_local_key_circuit_closed,
project_local_key_circuit_failure, project_local_key_circuit_open,
project_local_success_health, LocalFailoverClassification,
project_local_key_circuit_failure, project_local_success_health, LocalFailoverClassification,
};
use crate::ai_serving::extract_pool_sticky_session_token;
use crate::client_session_affinity::{
@@ -29,7 +28,7 @@ use crate::client_session_affinity::{
use crate::clock::current_unix_secs;
use crate::handlers::shared::provider_pool::admin_provider_pool_config_from_config_value;
use crate::handlers::shared::provider_pool::{
admin_provider_pool_key_circuit_breaker_reason, record_admin_provider_pool_error,
admin_provider_pool_key_terminal_error_reason, record_admin_provider_pool_error,
record_admin_provider_pool_stream_timeout, record_admin_provider_pool_success,
release_admin_provider_pool_key_lease, AdminProviderPoolConfig,
};
@@ -574,6 +573,7 @@ async fn record_health_failure_effect(
else {
return;
};
let is_pool_provider = local_execution_plan_uses_pool(state, context.plan).await;
let observed_at_unix_secs = current_unix_secs();
let Some(health_by_format) = project_local_failure_health(
current_key.health_by_format.as_ref(),
@@ -589,16 +589,24 @@ async fn record_health_failure_effect(
.and_then(|value| value.get("consecutive_failures"))
.and_then(Value::as_u64)
.unwrap_or(0);
let circuit_breaker_by_format = project_local_key_circuit_failure(
current_key.circuit_breaker_by_format.as_ref(),
api_format,
observed_at_unix_secs,
consecutive_failures,
current_key.max_probe_interval_minutes,
);
let circuit_breaker_update = circuit_breaker_by_format
.as_ref()
.or(current_key.circuit_breaker_by_format.as_ref());
let circuit_breaker_update_owned = if is_pool_provider {
None
} else {
project_local_key_circuit_failure(
current_key.circuit_breaker_by_format.as_ref(),
api_format,
observed_at_unix_secs,
consecutive_failures,
current_key.max_probe_interval_minutes,
)
};
let circuit_breaker_update = if is_pool_provider {
None
} else {
circuit_breaker_update_owned
.as_ref()
.or(current_key.circuit_breaker_by_format.as_ref())
};
if let Err(err) = state
.update_provider_catalog_key_health_state(
@@ -636,23 +644,35 @@ async fn record_health_success_effect(
else {
return;
};
let is_pool_provider = local_execution_plan_uses_pool(state, context.plan).await;
let Some(health_by_format) =
project_local_success_health(current_key.health_by_format.as_ref(), api_format)
else {
return;
};
let circuit_breaker_by_format = current_key
.circuit_breaker_by_format
.as_ref()
.and_then(|current| project_local_key_circuit_closed(Some(current), api_format));
let circuit_breaker_update_owned = if is_pool_provider {
None
} else {
current_key
.circuit_breaker_by_format
.as_ref()
.and_then(|current| project_local_key_circuit_closed(Some(current), api_format))
};
if current_key.health_by_format.as_ref() == Some(&health_by_format)
&& circuit_breaker_by_format.as_ref() == current_key.circuit_breaker_by_format.as_ref()
&& ((is_pool_provider && current_key.circuit_breaker_by_format.is_none())
|| (!is_pool_provider
&& circuit_breaker_update_owned.as_ref()
== current_key.circuit_breaker_by_format.as_ref()))
{
return;
}
let circuit_breaker_update = circuit_breaker_by_format
.as_ref()
.or(current_key.circuit_breaker_by_format.as_ref());
let circuit_breaker_update = if is_pool_provider {
None
} else {
circuit_breaker_update_owned
.as_ref()
.or(current_key.circuit_breaker_by_format.as_ref())
};
if let Err(err) = state
.update_provider_catalog_key_health_state(
@@ -711,9 +731,9 @@ async fn record_pool_error_effect(
context: LocalExecutionEffectContext<'_>,
effect: LocalPoolErrorEffect<'_>,
) {
let circuit_reason =
admin_provider_pool_key_circuit_breaker_reason(effect.status_code, effect.error_body);
if circuit_reason.is_none()
let terminal_error_reason =
admin_provider_pool_key_terminal_error_reason(effect.status_code, effect.error_body);
if terminal_error_reason.is_none()
&& !local_candidate_failure_should_record_pool_error(
effect.classification,
effect.status_code,
@@ -726,10 +746,7 @@ async fn record_pool_error_effect(
return;
};
if let Some(reason) = circuit_reason {
open_pool_key_circuit_breaker(state, context, &reason).await;
}
clear_pool_key_circuit_breaker(state, context).await;
record_admin_provider_pool_error(
state.runtime_state.as_ref(),
&context.plan.provider_id,
@@ -757,16 +774,10 @@ async fn record_pool_error_effect(
.await;
}
async fn open_pool_key_circuit_breaker(
async fn clear_pool_key_circuit_breaker(
state: &AppState,
context: LocalExecutionEffectContext<'_>,
reason: &str,
) {
let api_format = context.plan.provider_api_format.trim();
if api_format.is_empty() {
return;
}
let Some(current_key) = state
.read_provider_catalog_keys_by_ids(std::slice::from_ref(&context.plan.key_id))
.await
@@ -775,27 +786,21 @@ async fn open_pool_key_circuit_breaker(
else {
return;
};
let Some(circuit_breaker_by_format) = project_local_key_circuit_open(
current_key.circuit_breaker_by_format.as_ref(),
api_format,
reason,
current_unix_secs(),
current_key.max_probe_interval_minutes,
) else {
if current_key.circuit_breaker_by_format.is_none() {
return;
};
}
if let Err(err) = state
.update_provider_catalog_key_health_state(
&context.plan.key_id,
current_key.is_active,
current_key.health_by_format.as_ref(),
Some(&circuit_breaker_by_format),
None,
)
.await
{
warn!(
"gateway orchestration effects: failed to open pool key circuit for provider {} endpoint {} key {}: {:?}",
"gateway orchestration effects: failed to clear pool key circuit for provider {} endpoint {} key {}: {:?}",
context.plan.provider_id, context.plan.endpoint_id, context.plan.key_id, err
);
}
@@ -984,6 +989,10 @@ fn pool_score_hard_state_for_status(
status_code: u16,
error_body: Option<&str>,
) -> Option<PoolMemberHardState> {
if let Some(reason) = admin_provider_pool_key_terminal_error_reason(status_code, error_body) {
return Some(pool_score_hard_state_for_terminal_error_reason(&reason));
}
match status_code {
401 | 403 => Some(PoolMemberHardState::AuthInvalid),
402 => Some(PoolMemberHardState::QuotaExhausted),
@@ -1006,6 +1015,16 @@ fn pool_score_hard_state_for_status(
}
}
fn pool_score_hard_state_for_terminal_error_reason(reason: &str) -> PoolMemberHardState {
if reason.starts_with("payment_required_") {
PoolMemberHardState::QuotaExhausted
} else if reason.starts_with("forbidden_") {
PoolMemberHardState::AuthInvalid
} else {
PoolMemberHardState::Banned
}
}
fn pool_score_delta_for_status(status_code: u16) -> i32 {
match status_code {
401 | 403 => -2_000,
@@ -1028,6 +1047,7 @@ mod tests {
use aether_data_contracts::repository::candidates::{
RequestCandidateStatus, StoredRequestCandidate,
};
use aether_data_contracts::repository::pool_scores::PoolMemberHardState;
use aether_data_contracts::repository::provider_catalog::{
StoredProviderCatalogEndpoint, StoredProviderCatalogKey, StoredProviderCatalogProvider,
};
@@ -1036,9 +1056,10 @@ mod tests {
use super::{
apply_local_execution_effect, local_candidate_failure_should_record_pool_error,
LocalAdaptiveRateLimitEffect, LocalAdaptiveSuccessEffect, LocalAttemptFailureEffect,
LocalExecutionEffect, LocalExecutionEffectContext, LocalHealthFailureEffect,
LocalHealthSuccessEffect, LocalOAuthInvalidationEffect, LocalPoolErrorEffect,
pool_score_hard_state_for_status, LocalAdaptiveRateLimitEffect, LocalAdaptiveSuccessEffect,
LocalAttemptFailureEffect, LocalExecutionEffect, LocalExecutionEffectContext,
LocalHealthFailureEffect, LocalHealthSuccessEffect, LocalOAuthInvalidationEffect,
LocalPoolErrorEffect,
};
use crate::data::{GatewayDataConfig, GatewayDataState};
use crate::orchestration::LocalFailoverClassification;
@@ -1943,13 +1964,46 @@ mod tests {
));
}
#[test]
fn terminal_pool_account_errors_project_pool_hard_state() {
assert_eq!(
pool_score_hard_state_for_status(
400,
Some(r#"{"error":{"message":"deactivated_workspace"}}"#),
),
Some(PoolMemberHardState::Banned)
);
assert_eq!(
pool_score_hard_state_for_status(
402,
Some(r#"{"error":{"message":"payment required"}}"#),
),
Some(PoolMemberHardState::QuotaExhausted)
);
}
#[tokio::test]
async fn pool_account_error_opens_key_circuit() {
async fn pool_account_error_does_not_open_key_circuit() {
let Some(redis) = start_managed_redis_or_skip().await else {
return;
};
let state = codex_state_with_redis(redis.redis_url(), "orchestration_pool_circuit");
let plan = sample_codex_plan();
let legacy_circuit = json!({
"openai:responses": {
"open": true,
"reason": "legacy"
}
});
state
.update_provider_catalog_key_health_state(
&plan.key_id,
true,
None,
Some(&legacy_circuit),
)
.await
.expect("legacy circuit should seed");
apply_local_execution_effect(
&state,
@@ -1973,15 +2027,7 @@ mod tests {
.into_iter()
.next()
.expect("stored key should exist");
let circuit = stored_key
.circuit_breaker_by_format
.as_ref()
.and_then(|value| value.get("openai:responses"))
.expect("format circuit should be stored");
assert_eq!(circuit["open"], json!(true));
assert_eq!(circuit["reason"], json!("account_deactivated_401"));
assert!(circuit["next_probe_at"].is_string());
assert!(circuit["next_probe_at_unix_secs"].as_u64().is_some());
assert_eq!(stored_key.circuit_breaker_by_format, None);
}
#[tokio::test]
@@ -2144,6 +2190,45 @@ mod tests {
);
}
#[tokio::test]
async fn pool_health_failure_does_not_open_key_circuit_after_eight_consecutive_failures() {
let state = pool_health_state();
let plan = sample_plan();
for _ in 0..8 {
apply_local_execution_effect(
&state,
LocalExecutionEffectContext {
plan: &plan,
report_context: None,
},
LocalExecutionEffect::HealthFailure(LocalHealthFailureEffect {
status_code: 503,
classification: LocalFailoverClassification::RetryUpstreamFailure,
}),
)
.await;
}
let stored_key = state
.read_provider_catalog_keys_by_ids(std::slice::from_ref(&plan.key_id))
.await
.expect("provider catalog keys should load")
.into_iter()
.next()
.expect("stored key should exist");
assert_eq!(stored_key.circuit_breaker_by_format, None);
assert_eq!(
stored_key
.health_by_format
.as_ref()
.and_then(|value| value.get("openai:chat"))
.and_then(|value| value.get("consecutive_failures"))
.and_then(Value::as_u64),
Some(8)
);
}
#[tokio::test]
async fn health_success_projection_resets_key_health_for_format() {
let state = health_state();
+1 -2
View File
@@ -35,8 +35,7 @@ pub(crate) use self::effects::{
};
pub(crate) use self::health::{
project_local_failure_health, project_local_key_circuit_closed,
project_local_key_circuit_failure, project_local_key_circuit_open,
project_local_success_health,
project_local_key_circuit_failure, project_local_success_health,
};
pub(crate) use self::policy::{
append_local_failover_policy_to_value, local_failover_policy_from_report_context,
@@ -128,6 +128,7 @@ pub(super) fn is_candidate_selectable(
.get(candidate.key_id.as_str())
.copied()
.unwrap_or(false),
enforce_key_circuit_breaker: !pool_group,
rpm_reset_at: (!pool_group)
.then(|| {
snapshot
@@ -182,6 +183,7 @@ pub(super) fn current_candidate_runtime_skip_reason(
.get(candidate.key_id.as_str())
.copied()
.unwrap_or(false),
enforce_key_circuit_breaker: !pool_group,
rpm_reset_at,
})
}
@@ -1553,6 +1553,71 @@ async fn selects_next_candidate_when_first_provider_key_circuit_is_open() {
assert_eq!(selected.key_id, "key-b");
}
#[tokio::test]
async fn pool_provider_ignores_key_circuit_open_runtime_skip() {
let mut first = sample_row();
first.provider_id = "provider-a".to_string();
first.provider_name = "openai-a".to_string();
first.endpoint_id = "endpoint-a".to_string();
first.key_id = "key-a".to_string();
first.key_name = "alpha".to_string();
first.key_global_priority_by_format = Some(serde_json::json!({"openai:chat": 1}));
let mut second = sample_row();
second.provider_id = "provider-b".to_string();
second.provider_name = "openai-b".to_string();
second.endpoint_id = "endpoint-b".to_string();
second.key_id = "key-b".to_string();
second.key_name = "beta".to_string();
second.key_global_priority_by_format = Some(serde_json::json!({"openai:chat": 2}));
let candidates = Arc::new(InMemoryMinimalCandidateSelectionReadRepository::seed(vec![
first, second,
]));
let mut pool_provider = sample_provider("provider-a", None);
pool_provider.config = Some(serde_json::json!({"pool_advanced": {"enabled": true}}));
let provider_catalog = Arc::new(InMemoryProviderCatalogReadRepository::seed(
vec![pool_provider, sample_provider("provider-b", None)],
Vec::new(),
vec![
sample_key("key-a", "provider-a", Some(10)).with_health_fields(
Some(serde_json::json!({"openai:chat": {"health_score": 0.2}})),
Some(serde_json::json!({"openai:chat": {"open": true}})),
),
sample_key("key-b", "provider-b", Some(10)),
],
));
let quotas = Arc::new(InMemoryProviderQuotaRepository::seed(vec![]));
let request_candidates = Arc::new(InMemoryRequestCandidateRepository::seed(vec![]));
let state = AppState::new()
.expect("state should build")
.with_data_state_for_tests(
GatewayDataState::with_candidate_selection_provider_catalog_quota_and_request_candidates_for_tests(
candidates,
provider_catalog,
quotas,
request_candidates,
),
);
let (selected, skipped) = collect_selectable_candidates_with_skip_reasons(
state.data.as_ref(),
&state,
"openai:chat",
"gpt-4.1",
false,
None,
100,
)
.await
.expect("selection should succeed");
assert!(selected
.iter()
.any(|candidate| candidate.provider_id == "provider-a"));
assert!(skipped.is_empty());
}
#[tokio::test]
async fn exposes_runtime_skipped_candidates_with_skip_reasons() {
let mut first = sample_row();
@@ -1595,7 +1595,7 @@ async fn gateway_ignores_health_signals_in_pool_scheduling_status() {
assert_eq!(keys.len(), 2);
assert_eq!(keys[0]["key_name"], json!("circuit-open"));
assert_eq!(keys[0]["circuit_breaker_open"], json!(true));
assert_eq!(keys[0]["circuit_breaker_open"], json!(false));
assert_eq!(keys[0]["scheduling_status"], json!("available"));
assert_eq!(keys[0]["scheduling_reason"], json!("available"));
assert_eq!(keys[0]["scheduling_label"], json!("可用"));