World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
Reuse the vector build connection for index status checks (#734)
mayakid0c08b86310e0
diff --git a/crates/utopia-store/src/vector_index.rs b/crates/utopia-store/src/vector_index.rsindex e114eb2a..8a3a7d15 100644--- a/crates/utopia-store/src/vector_index.rs+++ b/crates/utopia-store/src/vector_index.rs@@ -103,13 +103,21 @@ pub const RESORT: &str = "distance + 0, id";/// 索引现在的状态:`None` 没有;`Some(valid)` 有,`false` 是上次建到一半留下的pub async fn status(pool: &PgPool, target: Target, dims: usize) -> AppResult<Option<bool>> {+ status_on(pool, target, dims).await+}++async fn status_on<'a>(+ executor: impl Executor<'a, Database = Postgres>,+ target: Target,+ dims: usize,+) -> AppResult<Option<bool>> {let row: Option<(bool,)> = sqlx::query_as("SELECT i.indisvalid FROM pg_class cJOIN pg_index i ON i.indexrelid = c.oidWHERE c.relname = $1 AND c.relkind = 'i'",).bind(index_name(target, dims))- .fetch_optional(pool)+ .fetch_optional(executor).await?;Ok(row.map(|(v,)| v))}@@ -186,7 +194,9 @@ pub async fn build(pool: &PgPool, target: Target, dims: usize) -> AppResult<Buil// 默认 64 MB 的 maintenance_work_mem 到一万四千行 1024 维就装不下图,之后// 每一行都要落盘再读,5 万行建了 4 分 20 秒;512 MB 只在建的这一会儿占着conn.execute("SET maintenance_work_mem = '512MB'").await?;- let before = status(pool, target, dims).await?;+ // Reuse the held connection: acquiring another from a busy pool can+ // make builds wait for connections they are themselves holding.+ let before = status_on(&mut *conn, target, dims).await?;if before == Some(false) {conn.execute(format!("DROP INDEX CONCURRENTLY IF EXISTS {name}").as_str()).await?;diff --git a/crates/utopia-store/tests/a_vector_build_uses_one_connection.rs b/crates/utopia-store/tests/a_vector_build_uses_one_connection.rsnew file mode 100644index 00000000..682d593e--- /dev/null+++ b/crates/utopia-store/tests/a_vector_build_uses_one_connection.rs@@ -0,0 +1,38 @@+//! A build already owns a connection; checking its index must not need another.+use std::time::Duration;+use utopia_store::vector_index::{self, Target};++#[tokio::test]+async fn a_build_finishes_with_only_one_connection_available() -> anyhow::Result<()> {+ let Some(url) = utopia_store::test_db::url() else {+ return Ok(());+ };+ // Two is the application's supported minimum. The timeout only bounds a+ // broken nested acquire; no assertion depends on index-building latency.+ let pool = sqlx::postgres::PgPoolOptions::new()+ .max_connections(2)+ .acquire_timeout(Duration::from_secs(2))+ .connect(&url)+ .await?;+ utopia_store::db::migrate(&pool).await?;+ for (target, dims) in [(Target::Chunks, 17), (Target::EntityProfiles, 18)] {+ vector_index::drop(&pool, target, dims).await?;+ // Stand in for another request holding one connection. This makes the+ // pool pressure deterministic, without racing two build tasks.+ let busy = pool.acquire().await?;+ let run = async {+ let built = vector_index::build(&pool, target, dims).await?;+ assert!(built.created);+ assert_eq!(vector_index::status(&pool, target, dims).await?, Some(true));+ let again = vector_index::build(&pool, target, dims).await?;+ assert!(!again.created, "an existing valid index is reused");+ Ok::<_, anyhow::Error>(())+ }+ .await;+ drop(busy);+ vector_index::drop(&pool, target, dims).await?;+ run?;+ }+ pool.close().await;+ Ok(())+}