World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
fix(store): serialize vector index drops with builds (#908)
WaylandYanga9da69c94df9
diff --git a/crates/utopia-store/src/vector_index.rs b/crates/utopia-store/src/vector_index.rsindex 23f86e1f..c22180f9 100644--- a/crates/utopia-store/src/vector_index.rs+++ b/crates/utopia-store/src/vector_index.rs@@ -27,7 +27,7 @@//! 占表大头的库才走 HNSW(实测 6 万行:20 行和 1 万行的库走精确,5 万的走索引)。//! 应用侧不设阈值——阈值是对规划器的猜测,猜错了两边都慢。-use sqlx::{Executor, PgPool, Postgres, Transaction};+use sqlx::{Executor, PgConnection, PgPool, Postgres, Transaction};use std::collections::HashSet;use std::sync::{Mutex, OnceLock};use utopia_core::{AppError, AppResult};@@ -45,6 +45,25 @@ const BUILD_UNLOCK: &str = "SELECT pg_advisory_unlock(hashtextextended('vector_i/// 没抢到锁时隔多久再探。建一次索引几十秒到几分钟,四分之一秒的粒度够了const BUILD_LOCK_POLL: std::time::Duration = std::time::Duration::from_millis(250);+async fn lock_build(conn: &mut PgConnection) -> AppResult<()> {+ loop {+ let got: bool = sqlx::query_scalar(BUILD_TRY_LOCK)+ .fetch_one(&mut *conn)+ .await?;+ if got {+ return Ok(());+ }+ tokio::time::sleep(BUILD_LOCK_POLL).await;+ }+}++async fn unlock_build(mut conn: sqlx::pool::PoolConnection<Postgres>) {+ let unlocked: Result<bool, _> = sqlx::query_scalar(BUILD_UNLOCK).fetch_one(&mut *conn).await;+ if !matches!(unlocked, Ok(true)) {+ let _ = conn.close().await;+ }+}+/// pgvector 的 HNSW 对 `vector` 类型的上限。超过的维度(text-embedding-3-large/// 是 3072)不建索引,查询照常走精确路径。`halfvec` 能到 4000,但那是另一种/// 精度,等有人用到再说@@ -209,15 +228,7 @@ pub async fn build(pool: &PgPool, target: Target, dims: usize) -> AppResult<Buil// 构建任务被同时认领就是这个局面。锁是会话级的:CONCURRENTLY 不能进事务,事务级// 咨询锁没处放;跟着这条连接走,跨 worker、跨实例都排队。建索引一次几十秒到几分钟,// 排队比死锁后重试便宜- loop {- let got: bool = sqlx::query_scalar(BUILD_TRY_LOCK)- .fetch_one(&mut *conn)- .await?;- if got {- break;- }- tokio::time::sleep(BUILD_LOCK_POLL).await;- }+ lock_build(&mut conn).await?;let outcome = async {conn.execute("SET max_parallel_maintenance_workers = 0").await?;@@ -250,11 +261,11 @@ pub async fn build(pool: &PgPool, target: Target, dims: usize) -> AppResult<Buil// 还回池子:带着锁回池,之后所有构建都会卡在它后面let _ = conn.execute("RESET max_parallel_maintenance_workers").await;let _ = conn.execute("RESET maintenance_work_mem").await;- if conn.execute(BUILD_UNLOCK).await.is_err() {- let _ = conn.close().await;+ if outcome.is_ok() {+ remember(&name);}+ unlock_build(conn).await;let created = outcome?;- remember(&name);Ok(Built {name,created,@@ -265,10 +276,18 @@ pub async fn build(pool: &PgPool, target: Target, dims: usize) -> AppResult<Buil/// 删掉(测试与手工维护用;写路径不会走到这里)pub async fn drop(pool: &PgPool, target: Target, dims: usize) -> AppResult<()> {let name = index_name(target, dims);- forget(&name);let mut conn = pool.acquire().await?;- conn.execute(format!("DROP INDEX CONCURRENTLY IF EXISTS {name}").as_str())- .await?;+ // Dropping an index must take the same lock as building one: separate+ // concurrent index operations on the table can deadlock each other.+ lock_build(&mut conn).await?;+ let outcome = conn+ .execute(format!("DROP INDEX CONCURRENTLY IF EXISTS {name}").as_str())+ .await;+ if outcome.is_ok() {+ forget(&name);+ }+ unlock_build(conn).await;+ outcome?;Ok(())}diff --git a/crates/utopia-store/tests/a_vector_index_is_built_by_a_job.rs b/crates/utopia-store/tests/a_vector_index_is_built_by_a_job.rsindex fc2838a4..33f9235d 100644--- a/crates/utopia-store/tests/a_vector_index_is_built_by_a_job.rs+++ b/crates/utopia-store/tests/a_vector_index_is_built_by_a_job.rs@@ -23,6 +23,41 @@ async fn queued(pool: &PgPool, dims: usize) -> anyhow::Result<i64> {.await?)}+#[tokio::test]+async fn dropping_an_index_waits_for_a_build() -> anyhow::Result<()> {+ let Some(url) = utopia_store::test_db::url() else {+ return Ok(());+ };+ let pool = PgPool::connect(&url).await?;+ let mut holder = pool.acquire().await?;+ let lock = "SELECT pg_try_advisory_lock(hashtextextended('vector_index:build', 0))";+ loop {+ if sqlx::query_scalar::<_, bool>(lock)+ .fetch_one(&mut *holder)+ .await?+ {+ break;+ }+ tokio::time::sleep(std::time::Duration::from_millis(10)).await;+ }++ // An absent index makes DROP CONCURRENTLY quick unless it takes the build lock.+ let blocked = tokio::time::timeout(+ std::time::Duration::from_millis(100),+ vector_index::drop(&pool, Target::Chunks, 17),+ )+ .await+ .is_err();+ let unlocked: bool =+ sqlx::query_scalar("SELECT pg_advisory_unlock(hashtextextended('vector_index:build', 0))")+ .fetch_one(&mut *holder)+ .await?;+ assert!(unlocked);+ assert!(blocked, "DROP must wait for an in-progress index build");+ vector_index::drop(&pool, Target::Chunks, 17).await?;+ Ok(())+}+#[tokio::test]async fn the_first_write_of_a_dimension_queues_one_build() -> anyhow::Result<()> {let Some(url) = utopia_store::test_db::url() else {