Merge branch 'fawney19:main' into main

This commit is contained in:
ZheFox
2026-07-28 13:56:34 +08:00
committed by GitHub
117 changed files with 10467 additions and 1253 deletions
@@ -1,7 +1,7 @@
use aether_data::DataLayerError;
use aether_data_contracts::repository::candidate_selection::{
StoredMinimalCandidateSelectionRow, StoredPoolKeyCandidateRowsQuery,
StoredRequestedModelCandidateRowsQuery,
StoredApiFormatCandidateRowsQuery, StoredMinimalCandidateSelectionRow,
StoredPoolKeyCandidateRowsQuery, StoredRequestedModelCandidateRowsQuery,
};
use aether_scheduler_core::{
auth_constraints_allow_api_format, collect_global_model_names_for_required_capability,
@@ -39,6 +39,19 @@ pub(crate) trait MinimalCandidateSelectionRowSource {
api_format: &str,
) -> Result<Vec<StoredMinimalCandidateSelectionRow>, DataLayerError>;
async fn read_minimal_candidate_selection_rows_for_api_format_page(
&self,
query: &StoredApiFormatCandidateRowsQuery,
) -> Result<Vec<StoredMinimalCandidateSelectionRow>, DataLayerError> {
Ok(self
.read_minimal_candidate_selection_rows_for_api_format(&query.api_format)
.await?
.into_iter()
.skip(query.offset as usize)
.take(query.limit as usize)
.collect())
}
async fn read_pool_key_candidate_rows_for_group(
&self,
query: &StoredPoolKeyCandidateRowsQuery,
@@ -231,6 +244,30 @@ pub(crate) async fn read_requested_model_rows_fast_path_page(
})
}
pub(crate) async fn read_api_format_rows_fallback_page(
state: &(impl MinimalCandidateSelectionRowSource + Sync),
api_format: &str,
offset: u32,
limit: u32,
) -> Result<RequestedModelCandidateRowsPage, DataLayerError> {
let limit = limit.max(1);
let rows = state
.read_minimal_candidate_selection_rows_for_api_format_page(
&StoredApiFormatCandidateRowsQuery {
api_format: api_format.to_string(),
offset,
limit,
},
)
.await?;
let scanned_rows = rows.len() as u32;
Ok(RequestedModelCandidateRowsPage {
rows,
scanned_rows,
end_of_requested_name: scanned_rows < limit,
})
}
pub(crate) async fn enumerate_minimal_candidate_selection_with_required_capabilities(
state: &(impl MinimalCandidateSelectionRowSource + Sync),
api_format: &str,
@@ -7,9 +7,10 @@ use std::time::Duration;
use aether_cache::ExpiringMap;
use aether_data::DataLayerError;
use aether_data_contracts::repository::candidate_selection::{
MinimalCandidateSelectionReadRepository, StoredMinimalCandidateSelectionRow,
StoredPoolKeyCandidateOrder, StoredPoolKeyCandidateRowsByKeyIdsQuery,
StoredPoolKeyCandidateRowsQuery, StoredRequestedModelCandidateRowsQuery,
MinimalCandidateSelectionReadRepository, StoredApiFormatCandidateRowsQuery,
StoredMinimalCandidateSelectionRow, StoredPoolKeyCandidateOrder,
StoredPoolKeyCandidateRowsByKeyIdsQuery, StoredPoolKeyCandidateRowsQuery,
StoredRequestedModelCandidateRowsQuery,
};
use async_trait::async_trait;
use tokio::sync::{Notify, OwnedSemaphorePermit, Semaphore};
@@ -389,6 +390,19 @@ impl MinimalCandidateSelectionReadRepository for CachedMinimalCandidateSelection
.await
}
async fn list_for_exact_api_format_page(
&self,
query: &StoredApiFormatCandidateRowsQuery,
) -> Result<Vec<StoredMinimalCandidateSelectionRow>, DataLayerError> {
let key = CandidateSelectionCacheKey::ApiFormatPage {
api_format: normalize_api_format_key(&query.api_format),
offset: query.offset,
limit: query.limit,
};
self.get_or_load(key, || self.inner.list_for_exact_api_format_page(query))
.await
}
async fn list_for_exact_api_format_and_global_model(
&self,
api_format: &str,
@@ -480,6 +494,11 @@ enum CandidateSelectionCacheKey {
ApiFormat {
api_format: String,
},
ApiFormatPage {
api_format: String,
offset: u32,
limit: u32,
},
ApiFormatAndGlobalModel {
api_format: String,
global_model_name: String,
@@ -1044,6 +1063,36 @@ mod tests {
assert!(cache.inflight.lock().unwrap().is_empty());
}
#[tokio::test]
async fn candidate_selection_cache_keys_api_format_pages_by_offset() {
let inner = Arc::new(StubCandidateSelectionRepository::new(Duration::ZERO));
let cache = CachedMinimalCandidateSelectionReadRepository::new(inner.clone());
let first = StoredApiFormatCandidateRowsQuery {
api_format: "openai:chat".to_string(),
offset: 0,
limit: 256,
};
let second = StoredApiFormatCandidateRowsQuery {
offset: 256,
..first.clone()
};
cache
.list_for_exact_api_format_page(&first)
.await
.expect("first page should load");
cache
.list_for_exact_api_format_page(&first)
.await
.expect("first page should be cached");
cache
.list_for_exact_api_format_page(&second)
.await
.expect("second page should load independently");
assert_eq!(inner.calls(), 2);
}
#[tokio::test]
async fn candidate_selection_cache_releases_inflight_when_leader_is_cancelled() {
let inner = Arc::new(FirstLoadPendingThenFastRepository::new());
@@ -176,6 +176,14 @@ impl MinimalCandidateSelectionRowSource for GatewayDataState {
.await
}
async fn read_minimal_candidate_selection_rows_for_api_format_page(
&self,
query: &aether_data_contracts::repository::candidate_selection::StoredApiFormatCandidateRowsQuery,
) -> Result<Vec<StoredMinimalCandidateSelectionRow>, DataLayerError> {
self.list_minimal_candidate_selection_rows_for_api_format_page(query)
.await
}
async fn read_pool_key_candidate_rows_for_group(
&self,
query: &aether_data_contracts::repository::candidate_selection::StoredPoolKeyCandidateRowsQuery,
+3 -3
View File
@@ -97,9 +97,9 @@ use aether_data_contracts::repository::billing::{
UserPlanEntitlementRecord,
};
use aether_data_contracts::repository::candidate_selection::{
MinimalCandidateSelectionReadRepository, StoredMinimalCandidateSelectionRow,
StoredPoolKeyCandidateRowsByKeyIdsQuery, StoredPoolKeyCandidateRowsQuery,
StoredRequestedModelCandidateRowsQuery,
MinimalCandidateSelectionReadRepository, StoredApiFormatCandidateRowsQuery,
StoredMinimalCandidateSelectionRow, StoredPoolKeyCandidateRowsByKeyIdsQuery,
StoredPoolKeyCandidateRowsQuery, StoredRequestedModelCandidateRowsQuery,
};
use aether_data_contracts::repository::candidates::{
PublicHealthStatusCount, PublicHealthTimelineBucket, RequestCandidateReadRepository,
+23 -5
View File
@@ -2,11 +2,12 @@ use super::{
AdminGlobalModelListQuery, AdminProviderModelListQuery, CreateAdminGlobalModelRecord,
DataLayerError, GatewayDataState, PublicCatalogModelListQuery, PublicCatalogModelSearchQuery,
PublicGlobalModelQuery, StoredAdminGlobalModel, StoredAdminGlobalModelPage,
StoredAdminProviderModel, StoredMinimalCandidateSelectionRow,
StoredPoolKeyCandidateRowsByKeyIdsQuery, StoredPoolKeyCandidateRowsQuery,
StoredProviderActiveGlobalModel, StoredProviderModelStats, StoredPublicCatalogModel,
StoredPublicGlobalModel, StoredPublicGlobalModelPage, StoredRequestedModelCandidateRowsQuery,
UpdateAdminGlobalModelRecord, UpsertAdminProviderModelRecord,
StoredAdminProviderModel, StoredApiFormatCandidateRowsQuery,
StoredMinimalCandidateSelectionRow, StoredPoolKeyCandidateRowsByKeyIdsQuery,
StoredPoolKeyCandidateRowsQuery, StoredProviderActiveGlobalModel, StoredProviderModelStats,
StoredPublicCatalogModel, StoredPublicGlobalModel, StoredPublicGlobalModelPage,
StoredRequestedModelCandidateRowsQuery, UpdateAdminGlobalModelRecord,
UpsertAdminProviderModelRecord,
};
impl GatewayDataState {
@@ -98,6 +99,23 @@ impl GatewayDataState {
.await
}
pub(crate) async fn list_minimal_candidate_selection_rows_for_api_format_page(
&self,
query: &StoredApiFormatCandidateRowsQuery,
) -> Result<Vec<StoredMinimalCandidateSelectionRow>, DataLayerError> {
crate::request_diagnostics::observe_db_operation(
"candidate_selection",
self.database_pool_summary(),
async {
match &self.minimal_candidate_selection_reader {
Some(repository) => repository.list_for_exact_api_format_page(query).await,
None => Ok(Vec::new()),
}
},
)
.await
}
pub(crate) async fn list_pool_key_candidate_rows_for_group(
&self,
query: &StoredPoolKeyCandidateRowsQuery,