World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
Two migrations claimed version 25, and nothing was watching (#50)
WaylandYangf8e7be11139f
diff --git a/.env.example b/.env.exampleindex 46219c2d..26953953 100644--- a/.env.example+++ b/.env.example@@ -9,6 +9,10 @@ UTOPIA_WEB_DIST=web/distUTOPIA_DATA_DIR=data# 开放注册;false 时仅首个用户可注册,其余由管理员开放UTOPIA_OPEN_REGISTRATION=true+# 数据库连接池上限。缺省 32,与 worker 并发的缺省对齐。+# 池子小于并发时症状是请求变慢、而不是任何一处说"池子不够"——调高 worker+# 并发时一并考虑这里,也别超过 Postgres 自己的 max_connections(默认 100)。+# UTOPIA_DB_MAX_CONNECTIONS=32# --- 以下几项由 docker-compose 读取,不是应用配置 ---# 数据库口令。默认值只适合本机开发,公开部署务必改。diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.ymlindex ce079e07..f33a69b8 100644--- a/.github/workflows/ci.yml+++ b/.github/workflows/ci.yml@@ -26,6 +26,43 @@ jobs:- name: Buildrun: cargo build --workspace+ # 迁移得真的在库上跑一遍。此前 CI 只有 fmt/clippy/test/build,于是两个 PR+ # 各带一个 0025、各自通过,合进 main 之后服务直接起不来——sqlx 按版本号索引,+ # 而**合并后的状态从来没有被任何一次 CI 跑过**。+ migrations:+ runs-on: ubuntu-latest+ services:+ postgres:+ image: pgvector/pgvector:pg16+ env:+ POSTGRES_USER: utopia+ POSTGRES_PASSWORD: utopia+ POSTGRES_DB: utopia+ ports: ["5432:5432"]+ options: >-+ --health-cmd pg_isready --health-interval 5s+ --health-timeout 5s --health-retries 10+ steps:+ - uses: actions/checkout@v4+ # 撞号在文件层面不是冲突,git 合得干干净净——只能显式查+ - name: 版本号不得重复+ run: |+ dupes=$(ls migrations/ | cut -d_ -f1 | sort | uniq -d)+ if [ -n "$dupes" ]; then+ echo "::error::迁移版本号重复:$dupes"+ ls migrations/+ exit 1+ fi+ - uses: dtolnay/rust-toolchain@stable+ - uses: Swatinem/rust-cache@v2+ - name: 装 sqlx-cli+ run: cargo install sqlx-cli --no-default-features --features rustls,postgres --locked+ - name: 全新库上跑一遍+ run: sqlx migrate run --database-url postgres://utopia:utopia@localhost:5432/utopia+ # 第二遍必须也过:迁移改号后要能在已经跑过旧号的库上安全重放+ - name: 再跑一遍+ run: sqlx migrate run --database-url postgres://utopia:utopia@localhost:5432/utopia+web:runs-on: ubuntu-latestdefaults:diff --git a/crates/utopia-core/src/config.rs b/crates/utopia-core/src/config.rsindex 831819ee..79c1e473 100644--- a/crates/utopia-core/src/config.rs+++ b/crates/utopia-core/src/config.rs@@ -15,6 +15,9 @@ pub struct AppConfig {pub web_dist: String,/// 数据目录:原始文件(files/)与 Tantivy 索引(index/)。pub data_dir: String,+ /// 数据库连接池上限。缺省 32,与 worker 并发的缺省对齐——池子小于并发时+ /// 症状是请求变慢而不是任何一处说"池子不够",所以它必须可调。+ pub db_max_connections: Option<u32>,/// 是否开放注册。false 时仅首个用户(引导部署)可注册,其余需管理员开放。pub open_registration: bool,}@@ -27,6 +30,7 @@ impl Default for AppConfig {jwt_secret: "dev-secret-change-me".into(),web_dist: "web/dist".into(),data_dir: "data".into(),+ db_max_connections: None,open_registration: true,}}diff --git a/crates/utopia-server/src/main.rs b/crates/utopia-server/src/main.rsindex 136e8063..8ea7e5af 100644--- a/crates/utopia-server/src/main.rs+++ b/crates/utopia-server/src/main.rs@@ -40,7 +40,7 @@ async fn main() -> anyhow::Result<()> {.init();let cfg = AppConfig::load()?;- let pool = utopia_store::db::connect(&cfg.database_url).await?;+ let pool = utopia_store::db::connect(&cfg.database_url, cfg.db_max_connections).await?;utopia_store::db::migrate(&pool).await?;tracing::info!("数据库迁移完成");diff --git a/crates/utopia-store/src/db.rs b/crates/utopia-store/src/db.rsindex f849a891..a43708c2 100644--- a/crates/utopia-store/src/db.rs+++ b/crates/utopia-store/src/db.rs@@ -1,11 +1,25 @@use sqlx::postgres::PgPoolOptions;use sqlx::PgPool;+use std::time::Duration;-pub async fn connect(database_url: &str) -> anyhow::Result<PgPool> {+/// 连接池上限的缺省。+///+/// 与 worker 并发的缺省(32)对齐:任务大部分时间在等 HTTP 应答、并不占着连接,+/// 但每分块的 epoch 检查、未匹配统计这些短查询会成串涌来,池子太小就变成排队。+/// 曾经写死 10,而 worker 默认 32——三倍超发,撞上来会先是请求变慢再是超时,+/// 而不是任何一处报错说"池子不够"。+const DEFAULT_MAX_CONNECTIONS: u32 = 32;++pub async fn connect(database_url: &str, max_connections: Option<u32>) -> anyhow::Result<PgPool> {+ let max = max_connections.unwrap_or(DEFAULT_MAX_CONNECTIONS).max(2);let pool = PgPoolOptions::new()- .max_connections(10)+ .max_connections(max)+ // 取不到连接时早点响亮地失败,而不是把请求悬在默认的 30 秒上——+ // 池子配小了要看得出来是池子的问题+ .acquire_timeout(Duration::from_secs(10)).connect(database_url).await?;+ tracing::info!(max_connections = max, "数据库连接池已建立");Ok(pool)}diff --git a/docker-compose.yml b/docker-compose.ymlindex 8c534847..39e57862 100644--- a/docker-compose.yml+++ b/docker-compose.yml@@ -36,6 +36,10 @@ services:UTOPIA_JWT_SECRET: ${UTOPIA_JWT_SECRET:-please-change-me-in-production}UTOPIA_WEB_DIST: /app/web-distUTOPIA_DATA_DIR: /app/data+ # 连接池上限。缺省 32,与 worker 并发的缺省对齐;池子小于并发时症状是+ # 请求变慢而不是报错说池子不够,所以调并发时别忘了它。+ # 上调前先确认 Postgres 的 max_connections(默认 100)容得下。+ UTOPIA_DB_MAX_CONNECTIONS: ${UTOPIA_DB_MAX_CONNECTIONS:-32}volumes:# 原始文件(files/)与全文索引(index/)。不挂出来的话,容器一重建# 它们就没了,而库里的 documents 记录还在——引用全断且不报错。diff --git a/migrations/0025_proposed_type.sql b/migrations/0028_proposed_type.sqlsimilarity index 57%rename from migrations/0025_proposed_type.sqlrename to migrations/0028_proposed_type.sqlindex 5be4c75d..f81db152 100644--- a/migrations/0025_proposed_type.sql+++ b/migrations/0028_proposed_type.sql@@ -3,9 +3,16 @@-- 事实那边已经有了(surface_predicate),实体这边还没有:类型不在本体里就降级成-- concept,模型说的 "model" 只剩一个「model ×43」的总数,实体行上一字未留。-- 后果是想加 model 类时找不出那 43 个实体——它们混在 322 个 concept 里,只能整库重抽。-ALTER TABLE entities ADD COLUMN proposed_type TEXT;+ALTER TABLE entities ADD COLUMN IF NOT EXISTS proposed_type TEXT;-- 顺手改名。"表层"是造出来的词,得先解释才懂;"提议的"一眼就对,而且更准确——-- 那不是猜测:模型读了原文得出 model,通常是对的,只是本体表达不了。-- 真正带兜底性质的是我们存进去的 concept。-ALTER TABLE fact_evidence RENAME COLUMN surface_predicate TO proposed_predicate;+-- 幂等:改号之前这两条已在部分库里跑过(0025/0026 与另一条改动撞号,见 0030 的+-- 说明),renumber 后必须能在那些库上安全重放。+DO $$ BEGIN+ IF EXISTS (SELECT 1 FROM information_schema.columns+ WHERE table_name = 'fact_evidence' AND column_name = 'surface_predicate') THEN+ ALTER TABLE fact_evidence RENAME COLUMN surface_predicate TO proposed_predicate;+ END IF;+END $$;diff --git a/migrations/0026_model_concurrency.sql b/migrations/0029_model_concurrency.sqlsimilarity index 79%rename from migrations/0026_model_concurrency.sqlrename to migrations/0029_model_concurrency.sqlindex 9e952eed..71a408c5 100644--- a/migrations/0026_model_concurrency.sql+++ b/migrations/0029_model_concurrency.sql@@ -5,7 +5,7 @@---- 限流放在 LLM 调用处而不是任务调度处:不调模型的任务(文件夹同步)不该受它-- 约束,调不同模型的任务(抽取用 chat、摄入用 embedding)之间也不该互相挤。-CREATE TABLE model_concurrency (+CREATE TABLE IF NOT EXISTS model_concurrency (base_url TEXT NOT NULL,model TEXT NOT NULL,max_concurrent INT NOT NULL CHECK (max_concurrent BETWEEN 1 AND 256),@@ -14,11 +14,13 @@ CREATE TABLE model_concurrency ();-- 没配过的模型走这个缺省+-- 幂等:改号前这条已在部分库跑过(0026 撞号),必须能安全重放ALTER TABLE deployment_settings- ADD COLUMN default_model_concurrency INT NOT NULL DEFAULT 10;+ ADD COLUMN IF NOT EXISTS default_model_concurrency INT NOT NULL DEFAULT 10;-- worker 并发降级成外层兜底:真正的节流交给按模型的信号量,这里只防任务-- 无限堆积。要明显大于各模型限额之和,否则被限流的任务会占满槽位把别的饿死。+-- 幂等:改号前这条已在部分库跑过(0026 撞号),必须能安全重放ALTER TABLE deployment_settingsALTER COLUMN worker_concurrency SET DEFAULT 32;UPDATE deployment_settings SET worker_concurrency = 32 WHERE worker_concurrency = 4;