World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
Keep superseded document processing from overwriting the current revision (#1217)
Daniel0bae9c419ef8
diff --git a/crates/utopia-server/src/pipeline.rs b/crates/utopia-server/src/pipeline.rsindex 48503451..c012352e 100644--- a/crates/utopia-server/src/pipeline.rs+++ b/crates/utopia-server/src/pipeline.rs@@ -6,6 +6,7 @@ use crate::state::AppState;use futures_util::{stream, StreamExt};use utopia_core::models::{LlmSettings, Proposer};use utopia_llm::LlmClient;+use utopia_store::documents::{self, ProcessingInput};use uuid::Uuid;/// 这份文档的来源要不要抽取。没有来源的文档(直接上传、记忆片段)照旧抽。@@ -47,14 +48,19 @@ const EMBED_BATCH: usize = 16;const EMBED_JOBS: usize = 4;pub async fn process_document(state: &AppState, document_id: Uuid) -> anyhow::Result<()> {- match run(state, document_id).await {+ let input = documents::processing_input(&state.pool, document_id).await?;+ let doc = &input.document;+ match run(state, &input).await {Ok(()) => Ok(()),// 在等读字的服务读完:文档照旧是 parsing,任务过一会儿再来问- Err(e) if utopia_core::is_deferred(&e).is_some() => Err(e),+ Err(e) if utopia_core::is_deferred(&e).is_some() => {+ if documents::processing_is_current(&state.pool, &input).await? {+ Err(e)+ } else {+ Ok(())+ }+ }Err(e) => {- let doc = utopia_store::documents::get(&state.pool, document_id)- .await- .ok();// 字要靠模型读、而那种模型没配(0040):**降级**。文件留着,文档停在 failed// 并记下缺哪一种,报一条库级告警;配上模型会自己重新处理。重试没用——// 模型不会在两分钟里自己配上,所以挂 Terminal。@@ -66,33 +72,40 @@ pub async fn process_document(state: &AppState, document_id: Uuid) -> anyhow::Ree.downcast_ref::<utopia_ingest::NoSpeakers>().map(|_| utopia_ingest::Reader::Transcribe)});+ if !documents::set_failed_if_current(+ &state.pool,+ &input,+ &e.to_string(),+ waiting.map(|reader| reader.as_str()),+ )+ .await?+ {+ // 过期模型结果可丢弃,数据库故障仍要让队列看见。+ if e.is::<sqlx::Error>()+ || matches!(+ e.downcast_ref::<utopia_core::AppError>(),+ Some(utopia_core::AppError::Db(_))+ )+ {+ return Err(e);+ }+ tracing::info!(%document_id, "discarding a superseded document failure");+ return Ok(());+ }if let Some(reader) = waiting {- let _ = utopia_store::documents::set_needs_reader(- &state.pool,+ crate::alerting::observe_document_needs_reader(+ state,+ doc.kb_id,document_id,- reader.as_str(),+ &doc.filename,+ reader,&e.to_string(),).await;- if let Some(doc) = &doc {- crate::alerting::observe_document_needs_reader(- state,- doc.kb_id,- document_id,- &doc.filename,- reader,- &e.to_string(),- )- .await;- state.emit_document(doc.kb_id, document_id);- }- return Err(e.context(utopia_core::Terminal));- }- let _ =- utopia_store::documents::set_failed(&state.pool, document_id, &e.to_string()).await;- if let Some(doc) = &doc {state.emit_document(doc.kb_id, document_id);+ return Err(e.context(utopia_core::Terminal));}+ state.emit_document(doc.kb_id, document_id);// 读不了的格式(视频、二进制、空文件)换多少次也一样if e.downcast_ref::<utopia_ingest::Unreadable>().is_some() {return Err(e.context(utopia_core::Terminal));@@ -102,8 +115,9 @@ pub async fn process_document(state: &AppState, document_id: Uuid) -> anyhow::Re}}-async fn run(state: &AppState, document_id: Uuid) -> anyhow::Result<()> {- let doc = utopia_store::documents::get(&state.pool, document_id).await?;+async fn run(state: &AppState, input: &ProcessingInput) -> anyhow::Result<()> {+ let doc = &input.document;+ let document_id = doc.id;// 排队之后被删了(#268):墓碑不重建分块、不回索引;清过的连原文都没了if doc.deleted_at.is_some() {tracing::info!(document = %document_id, "skipping a deleted document");@@ -111,7 +125,9 @@ async fn run(state: &AppState, document_id: Uuid) -> anyhow::Result<()> {}// 1. 解析(CPU 密集,放 blocking 线程)- utopia_store::documents::set_status(&state.pool, document_id, "parsing").await?;+ if !documents::set_status_if_current(&state.pool, input, "parsing").await? {+ return Ok(());+ }state.emit_document(doc.kb_id, document_id);let bytes = state.blob.get(&doc.sha256).await?;let filename = doc.filename.clone();@@ -157,13 +173,13 @@ async fn run(state: &AppState, document_id: Uuid) -> anyhow::Result<()> {let reading = match (needs.reader, settings.as_ref()) {(utopia_ingest::Reader::Ocr, Some(s)) => {match crate::readers::Ocr::from_settings(s) {- Some(ocr) => ocr.read(state, &doc, bytes).await?,+ Some(ocr) => ocr.read(state, doc, bytes).await?,None => return Err(e),}}(utopia_ingest::Reader::Transcribe, Some(s)) => {match crate::readers::Transcriber::from_settings(s) {- Some(t) => t.read(&doc, bytes).await?,+ Some(t) => t.read(doc, bytes).await?,None => return Err(e),}}@@ -180,14 +196,8 @@ async fn run(state: &AppState, document_id: Uuid) -> anyhow::Result<()> {}};let text_len = text.chars().count() as i32;- let Some(chunk_pairs) = utopia_store::documents::replace_chunks_if_current(- &state.pool,- doc.kb_id,- document_id,- &pieces,- &doc.sha256,- )- .await?+ let Some(chunk_pairs) =+ documents::replace_chunks_if_current(&state.pool, input, &pieces).await?else {// 读取期间源文档可能已更新或删除,丢弃过期结果,// 不再改写新任务的索引、状态和抽取队列。@@ -208,7 +218,9 @@ async fn run(state: &AppState, document_id: Uuid) -> anyhow::Result<()> {}// 3. 全文索引(Tantivy)- utopia_store::documents::set_status(&state.pool, document_id, "indexing").await?;+ if !documents::set_status_if_current(&state.pool, input, "indexing").await? {+ return Ok(());+ }state.emit_document(doc.kb_id, document_id);let search = state.search.clone();let kb = doc.kb_id.to_string();@@ -217,36 +229,39 @@ async fn run(state: &AppState, document_id: Uuid) -> anyhow::Result<()> {// 4. embedding(工作区配置了 embedding 模型才做;没配也算 ready,先享受 BM25 搜索)if let Some((settings, client)) = embedder(settings.as_ref()) {- utopia_store::documents::set_status(&state.pool, document_id, "embedding").await?;+ if !documents::set_status_if_current(&state.pool, input, "embedding").await? {+ return Ok(());+ }state.emit_document(doc.kb_id, document_id);- embed_pending(state, settings, &client, document_id).await?;+ embed_pending(state, settings, &client, input).await?;}- utopia_store::documents::set_ready(&state.pool, document_id, text_len, chunk_count).await?;-// 来源说了不抽取的,到这里为止:可搜、可问,不进图。//// schema 文档就是这一类(0035 决定 7)——它是给问数检索表结构的语料,进抽取的// 结果是抽取器把每个列名当成一个实体(宽表语料上四十个概念实体里二十八个是// 列名,#553)。状态记成 `skipped` 而不是留在 `none`:`none` 在 Library 里读作// 「还没排到」,而它永远不会排到- if !source_extracts(state, doc.source_id).await? {- utopia_store::documents::set_graph_status(&state.pool, document_id, "skipped").await?;- state.emit_document(doc.kb_id, document_id);- tracing::info!(%document_id, chunks = chunk_count, "文档处理完成,来源不抽取");- return Ok(());- }-// 两段式:索引就绪后,若配置了对话模型则排队图谱抽取(不阻塞可搜可问)。// 推送来的陈述不问模型(0054),没配也排- if pushed_statements || settings.as_ref().is_some_and(|s| s.chat_ready()) {- utopia_store::documents::set_graph_status(&state.pool, document_id, "queued").await?;- utopia_store::jobs::enqueue(- &state.pool,- "extract_document",- serde_json::json!({ "document_id": document_id }),- )- .await?;+ let graph_status = if !source_extracts(state, doc.source_id).await? {+ Some("skipped")+ } else if pushed_statements || settings.as_ref().is_some_and(|s| s.chat_ready()) {+ Some("queued")+ } else {+ None+ };+ if !documents::finish_processing_if_current(+ &state.pool,+ input,+ text_len,+ chunk_count,+ graph_status,+ )+ .await?+ {+ tracing::info!(%document_id, "discarding a superseded document completion");+ return Ok(());}state.emit_document(doc.kb_id, document_id);@@ -363,10 +378,9 @@ async fn embed_pending(state: &AppState,settings: &LlmSettings,client: &LlmClient,- document_id: Uuid,+ input: &ProcessingInput,) -> anyhow::Result<usize> {- let pending =- utopia_store::documents::chunks_pending_embedding(&state.pool, document_id).await?;+ let pending = documents::chunks_pending_embedding(&state.pool, input.document.id).await?;// 批次先切成拥有的数据:借来的切片捕进并发驱动的 future 过不了 Send 边界//(ontology_index 里同一句话)let batches: Vec<Vec<(Uuid, String)>> =@@ -385,7 +399,9 @@ async fn embed_pending(}let items: Vec<(Uuid, Vec<f32>)> =batch.iter().map(|(id, _)| *id).zip(embeddings).collect();- utopia_store::documents::set_embeddings(&state.pool, &items).await?;+ if !documents::set_embeddings_if_current(&state.pool, input, &items).await? {+ return Ok(0);+ }Ok::<usize, anyhow::Error>(items.len())}).buffer_unordered(EMBED_JOBS);diff --git a/crates/utopia-server/src/pipeline_tests.rs b/crates/utopia-server/src/pipeline_tests.rsindex 4db5fa41..c88e92d7 100644--- a/crates/utopia-server/src/pipeline_tests.rs+++ b/crates/utopia-server/src/pipeline_tests.rs@@ -22,6 +22,7 @@//! 没有 `UTOPIA_DATABASE_URL` 时跳过而不是失败。自建自拆,绝不碰已有的库。use std::fmt::Write as _;+use std::sync::atomic::{AtomicBool, Ordering};use std::sync::{Arc, Mutex};use std::time::{Duration, Instant};use utopia_core::models::Proposer;@@ -231,7 +232,8 @@ impl Fx {.await?.expect("settings were written by the fixture");let client = crate::llm_util::embed_client(&settings).expect("embed model is configured");- super::embed_pending(&self.state, &settings, &client, doc).await+ let input = utopia_store::documents::processing_input(&self.pool, doc).await?;+ super::embed_pending(&self.state, &settings, &client, &input).await}/// (正文, 向量) 按 seq;向量为空就是没嵌@@ -353,16 +355,278 @@ async fn a_failed_batch_does_not_strand_the_document() -> anyhow::Result<()> {return Ok(());};let doc = f.document_to_process(20).await?;- assert!(super::process_document(&f.state, doc).await.is_err());+ let original = utopia_store::documents::get(&f.pool, doc).await?;+ let error = super::process_document(&f.state, doc)+ .await+ .expect_err("the current revision really failed");+ assert!(utopia_core::is_deferred(&error).is_none());+ assert!(!utopia_core::is_terminal(&error));+ let error_message = error.to_string();+ assert!(error_message.contains("model is down"));let row = utopia_store::documents::get(&f.pool, doc).await?;assert_eq!(row.status, "failed", "not left sitting in embedding");- assert!(- row.error.as_deref().is_some_and(|e| !e.is_empty()),- "the reason is on the document"+ assert_eq!(row.error.as_deref(), Some(error_message.as_str()));+ assert_eq!(+ (row.text_len, row.chunk_count),+ (original.text_len, original.chunk_count));f.cleanup().await}+/// #1215:第一批嵌入必须等测试放行;新版请求立即完成,不靠延时猜并发顺序。+async fn pause_embedding(+ f: &Fx,+ doc: Uuid,+ fail_old: bool,+) -> anyhow::Result<(+ Arc<tokio::sync::Notify>,+ tokio::task::JoinHandle<anyhow::Result<()>>,+ tokio::task::JoinHandle<()>,+)> {+ let entered = Arc::new(tokio::sync::Notify::new());+ let resume = Arc::new(tokio::sync::Notify::new());+ let first_request = Arc::new(AtomicBool::new(true));+ let app = axum::Router::new().route(+ "/embeddings",+ axum::routing::post({+ let entered = entered.clone();+ let resume = resume.clone();+ move |axum::Json(body): axum::Json<serde_json::Value>| {+ let entered = entered.clone();+ let resume = resume.clone();+ let first_request = first_request.clone();+ async move {+ let is_old = first_request.swap(false, Ordering::SeqCst);+ if is_old {+ entered.notify_one();+ resume.notified().await;+ if fail_old {+ return (+ axum::http::StatusCode::INTERNAL_SERVER_ERROR,+ axum::Json(serde_json::json!({"error": "old embedding failed"})),+ );+ }+ }+ let data: Vec<_> = body["input"]+ .as_array()+ .expect("embedding input is an array")+ .iter()+ .map(|input| {+ let mut vector = vector_of(input.as_str().expect("input is text"));+ // 同 SHA 的新版本可能复用 chunk ID,旧向量仍必须丢弃。+ if is_old {+ vector[0] += 10_000.0;+ }+ serde_json::json!({"embedding": vector})+ })+ .collect();+ (+ axum::http::StatusCode::OK,+ axum::Json(serde_json::json!({"data": data})),+ )+ }+ }+ }),+ );+ let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await?;+ let base = format!("http://{}", listener.local_addr()?);+ let server = tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });+ utopia_store::settings::upsert(+ &f.pool,+ f.ws,+ Some(&base),+ None,+ Some("fake-chat"),+ Some(&base),+ None,+ Some("fake-embed"),+ Some(4),+ )+ .await?;+ let state = f.state.clone();+ let processing = tokio::spawn(async move { super::process_document(&state, doc).await });+ tokio::time::timeout(Duration::from_secs(10), entered.notified()).await?;+ Ok((resume, processing, server))+}++async fn replace_document_text(f: &Fx, doc: Uuid, text: &str) -> anyhow::Result<()> {+ use sha2::{Digest, Sha256};+ let sha: String = Sha256::digest(text.as_bytes())+ .iter()+ .map(|byte| format!("{byte:02x}"))+ .collect();+ f.state.blob.put(&sha, text.as_bytes()).await?;+ utopia_store::documents::replace_content_and_enqueue_processing(+ &f.pool,+ doc,+ "long.md",+ "text/markdown",+ text.len() as i64,+ &sha,+ None,+ )+ .await?;+ Ok(())+}++async fn processing_snapshot(f: &Fx, doc: Uuid) -> anyhow::Result<serde_json::Value> {+ Ok(sqlx::query_scalar(+ "SELECT jsonb_build_object(+ 'document', (SELECT to_jsonb(d) FROM documents d WHERE d.id = $1),+ 'chunks', (SELECT coalesce(jsonb_agg(to_jsonb(c) ORDER BY c.seq, c.id), '[]')+ FROM chunks c WHERE c.document_id = $1 AND c.superseded_at IS NULL),+ 'jobs', (SELECT coalesce(jsonb_agg(to_jsonb(j) ORDER BY j.id), '[]')+ FROM jobs j WHERE j.payload->>'document_id' = $1::text),+ 'alerts', (SELECT coalesce(jsonb_agg(to_jsonb(a) ORDER BY a.id), '[]')+ FROM alerts a WHERE a.kb_id = $2))",+ )+ .bind(doc)+ .bind(f.kb)+ .fetch_one(&f.pool)+ .await?)+}++async fn late_embedding_preserves_newer_revision(+ fail_old: bool,+ restore_original: bool,+) -> anyhow::Result<()> {+ let Some(f) = fixture(FakeEmbed::new(Duration::ZERO)).await? else {+ return Ok(());+ };+ let doc = f.document_to_process(4).await?;+ let original = utopia_store::documents::get(&f.pool, doc).await?;+ let versions = utopia_store::documents::versions(&f.pool, doc).await?;+ let original_version = versions.last().expect("the document has a version").version;+ let original_text = String::from_utf8(f.state.blob.get(&original.sha256).await?)?;+ let (resume, old, server) = pause_embedding(&f, doc, fail_old).await?;+ let old_chunks: Vec<Uuid> = sqlx::query_scalar(+ "SELECT id FROM chunks WHERE document_id = $1 AND superseded_at IS NULL ORDER BY seq",+ )+ .bind(doc)+ .fetch_all(&f.pool)+ .await?;+ assert!(old_chunks.len() > 1 && old_chunks.len() <= super::EMBED_BATCH);++ let new_text = "The NEW budget is 200.";+ replace_document_text(&f, doc, new_text).await?;+ if restore_original {+ // B 尚未解析,恢复 A 时会认领旧任务正在嵌入的同一批 chunk ID。+ replace_document_text(&f, doc, &original_text).await?;+ }+ super::process_document(&f.state, doc).await?;+ let newer = utopia_store::documents::get(&f.pool, doc).await?;+ assert_eq!(newer.status, "ready");+ assert_eq!(newer.error, None);+ if restore_original {+ assert_eq!(newer.sha256, original.sha256);+ let versions = utopia_store::documents::versions(&f.pool, doc).await?;+ assert!(versions.last().expect("the document has a version").version > original_version);+ let current_chunks: Vec<Uuid> = sqlx::query_scalar(+ "SELECT id FROM chunks+ WHERE document_id = $1 AND superseded_at IS NULL ORDER BY seq",+ )+ .bind(doc)+ .fetch_all(&f.pool)+ .await?;+ assert_eq!(+ current_chunks, old_chunks,+ "restoring A reuses the chunks held by the old request"+ );+ } else {+ assert_ne!(newer.chunk_count, old_chunks.len() as i32);+ }++ // 新版抽取也已完成,旧任务不能再把它改成 queued 或另排一次抽取。+ utopia_store::documents::set_graph_status(&f.pool, doc, "done").await?;+ sqlx::query("UPDATE jobs SET status = 'done' WHERE payload->>'document_id' = $1")+ .bind(doc.to_string())+ .execute(&f.pool)+ .await?;+ let completed = processing_snapshot(&f, doc).await?;+ resume.notify_one();+ let old_result = tokio::time::timeout(Duration::from_secs(10), old).await??;+ server.abort();+ let current = processing_snapshot(&f, doc).await?;+ sqlx::query("DELETE FROM jobs WHERE payload->>'document_id' = $1")+ .bind(doc.to_string())+ .execute(&f.pool)+ .await?;+ f.cleanup().await?;+ old_result?;+ assert_eq!(+ current, completed,+ "a late embedding must leave the newer revision intact"+ );+ Ok(())+}++#[tokio::test]+async fn a_late_successful_embedding_preserves_a_newer_revision() -> anyhow::Result<()> {+ late_embedding_preserves_newer_revision(false, false).await+}++#[tokio::test]+async fn a_late_failed_embedding_preserves_a_newer_revision() -> anyhow::Result<()> {+ late_embedding_preserves_newer_revision(true, false).await+}++#[tokio::test]+async fn a_late_embedding_cannot_accept_a_restored_sha_as_its_revision() -> anyhow::Result<()> {+ late_embedding_preserves_newer_revision(false, true).await+}++#[tokio::test]+async fn a_late_embedding_preserves_a_deleted_document() -> anyhow::Result<()> {+ for fail_old in [false, true] {+ let Some(f) = fixture(FakeEmbed::new(Duration::ZERO)).await? else {+ return Ok(());+ };+ let doc = f.document_to_process(4).await?;+ let (resume, old, server) = pause_embedding(&f, doc, fail_old).await?;+ utopia_store::documents::delete(&f.pool, f.kb, doc, None).await?;+ let deleted = processing_snapshot(&f, doc).await?;+ assert!(deleted["document"]["deleted_at"].is_string());+ assert_eq!(deleted["chunks"], serde_json::json!([]));+ resume.notify_one();+ let result = tokio::time::timeout(Duration::from_secs(10), old).await??;+ server.abort();+ let current = processing_snapshot(&f, doc).await?;+ f.cleanup().await?;+ result?;+ assert_eq!(+ current, deleted,+ "deleted documents reject old success and failure"+ );+ }+ Ok(())+}++#[tokio::test]+async fn an_embedding_failure_does_not_hide_a_database_write_error() -> anyhow::Result<()> {+ let Some(mut f) = fixture(FakeEmbed::new(Duration::ZERO)).await? else {+ return Ok(());+ };+ let url = utopia_store::test_db::url().expect("the fixture has a configured database");+ let doc = f.document_to_process(4).await?;+ let (resume, processing, server) = pause_embedding(&f, doc, true).await?;+ f.pool.close().await;+ resume.notify_one();+ let result = tokio::time::timeout(Duration::from_secs(10), processing).await??;+ server.abort();+ // 重新连接只为拆夹具,处理器持有的旧连接池仍然关闭。+ f.pool = sqlx::PgPool::connect(&url).await?;+ f.cleanup().await?;+ let error = result.expect_err("failure status could not be persisted");+ assert!(+ matches!(+ error.downcast_ref::<utopia_core::AppError>(),+ Some(utopia_core::AppError::Db(sqlx::Error::PoolClosed))+ ),+ "the database error must reach the caller: {error:#}"+ );+ Ok(())+}+/// #611:一个 NUL 从前让整篇文档失败——坏的只是几个字节,丢的是整篇#[tokio::test]async fn a_nul_byte_does_not_fail_the_whole_document() -> anyhow::Result<()> {diff --git a/crates/utopia-store/src/documents.rs b/crates/utopia-store/src/documents.rsindex 7072a99d..3822fbad 100644--- a/crates/utopia-store/src/documents.rs+++ b/crates/utopia-store/src/documents.rs@@ -619,6 +619,57 @@ pub async fn get(pool: &PgPool, id: Uuid) -> AppResult<Document> {.ok_or(AppError::NotFound)}+/// 一次处理的输入;版本号区分内容 A → B → A,不能只比较 SHA。+pub struct ProcessingInput {+ pub document: Document,+ pub version: i32,+}++pub async fn processing_input(pool: &PgPool, id: Uuid) -> AppResult<ProcessingInput> {+ let mut tx = pool.begin().await?;+ let document = sqlx::query_as("SELECT * FROM documents WHERE id = $1 FOR NO KEY UPDATE")+ .bind(id)+ .fetch_optional(&mut *tx)+ .await?+ .ok_or(AppError::NotFound)?;+ let version = processing_version_tx(&mut tx, id).await?;+ tx.commit().await?;+ Ok(ProcessingInput { document, version })+}++async fn processing_version_tx(tx: &mut Transaction<'_, Postgres>, id: Uuid) -> AppResult<i32> {+ Ok(sqlx::query_scalar(+ "SELECT coalesce(max(version), 0) FROM document_versions WHERE document_id = $1",+ )+ .bind(id)+ .fetch_one(&mut **tx)+ .await?)+}++/// 锁后再读版本:等待内容更新的事务提交后,要看到它刚追加的版本。+async fn lock_processing_input_tx(+ tx: &mut Transaction<'_, Postgres>,+ input: &ProcessingInput,+) -> AppResult<bool> {+ let current: Option<(String, bool)> = sqlx::query_as(+ "SELECT sha256, deleted_at IS NOT NULL FROM documents WHERE id = $1 FOR NO KEY UPDATE",+ )+ .bind(input.document.id)+ .fetch_optional(&mut **tx)+ .await?;+ if !matches!(current, Some((ref sha, false)) if sha == &input.document.sha256) {+ return Ok(false);+ }+ Ok(processing_version_tx(tx, input.document.id).await? == input.version)+}++pub async fn processing_is_current(pool: &PgPool, input: &ProcessingInput) -> AppResult<bool> {+ let mut tx = pool.begin().await?;+ let current = lock_processing_input_tx(&mut tx, input).await?;+ tx.commit().await?;+ Ok(current)+}+/// A recorded original, with only the facts the ingestion ledger guarantees.#[derive(Debug, Clone, Serialize, sqlx::FromRow)]pub struct DocumentVersion {@@ -771,6 +822,12 @@ pub async fn record_version(sha256: &str,size_bytes: i64,) -> AppResult<()> {+ let mut tx = pool.begin().await?;+ // 版本追加也要和处理回写互斥,外键的 KEY SHARE 锁不会挡住它。+ sqlx::query("SELECT id FROM documents WHERE id = $1 FOR NO KEY UPDATE")+ .bind(document_id)+ .execute(&mut *tx)+ .await?;sqlx::query("INSERT INTO document_versions (id, document_id, version, sha256, size_bytes, doc_time)VALUES ($1, $2,@@ -781,8 +838,9 @@ pub async fn record_version(.bind(document_id).bind(sha256).bind(size_bytes)- .execute(pool)+ .execute(&mut *tx).await?;+ tx.commit().await?;Ok(())}@@ -858,32 +916,51 @@ pub async fn set_status(pool: &PgPool, id: Uuid, status: &str) -> AppResult<()>Ok(())}-pub async fn set_failed(pool: &PgPool, id: Uuid, error: &str) -> AppResult<()> {+pub async fn set_status_if_current(+ pool: &PgPool,+ input: &ProcessingInput,+ status: &str,+) -> AppResult<bool> {+ let mut tx = pool.begin().await?;+ if !lock_processing_input_tx(&mut tx, input).await? {+ return Ok(false);+ }sqlx::query(- "UPDATE documents SET status = 'failed', error = $2, reader_task = NULL, updated_at = now()+ "UPDATE documents SET status = $2, error = NULL, reader_needed = NULL, updated_at = now()WHERE id = $1",)- .bind(id)- .bind(error)- .execute(pool)+ .bind(input.document.id)+ .bind(status)+ .execute(&mut *tx).await?;- Ok(())+ tx.commit().await?;+ Ok(true)}-/// 这份文件要 `reader` 那一种模型才读得出字,而那种模型没配(0040):停在 failed,-/// 记下缺的是哪一种,配上之后按它重新排队-pub async fn set_needs_reader(pool: &PgPool, id: Uuid, reader: &str, error: &str) -> AppResult<()> {+/// 当前输入失败才记下原因;缺读取模型时仍按原语义等待配上后重排。+pub async fn set_failed_if_current(+ pool: &PgPool,+ input: &ProcessingInput,+ error: &str,+ reader: Option<&str>,+) -> AppResult<bool> {+ let mut tx = pool.begin().await?;+ if !lock_processing_input_tx(&mut tx, input).await? {+ return Ok(false);+ }sqlx::query(- "UPDATE documents SET status = 'failed', error = $3, reader_needed = $2, reader_task = NULL,+ "UPDATE documents SET status = 'failed', error = $2,+ reader_needed = coalesce($3, reader_needed), reader_task = NULL,updated_at = now()WHERE id = $1",)- .bind(id)- .bind(reader)+ .bind(input.document.id).bind(error)- .execute(pool)+ .bind(reader)+ .execute(&mut *tx).await?;- Ok(())+ tx.commit().await?;+ Ok(true)}/// 配上了 `reader` 那一种模型:这个工作区里等着它的文档(活着的)重新排进处理队列。@@ -971,6 +1048,44 @@ pub async fn set_ready(pool: &PgPool, id: Uuid, text_len: i32, chunk_count: i32)Ok(())}+/// 就绪、抽取状态和入队一起提交,过期任务不能留下新版的后续工作。+pub async fn finish_processing_if_current(+ pool: &PgPool,+ input: &ProcessingInput,+ text_len: i32,+ chunk_count: i32,+ graph_status: Option<&str>,+) -> AppResult<bool> {+ let mut tx = pool.begin().await?;+ if !lock_processing_input_tx(&mut tx, input).await? {+ return Ok(false);+ }+ sqlx::query(+ "UPDATE documents SET status = 'ready', error = NULL, reader_needed = NULL, reader_task = NULL,+ text_len = $2, chunk_count = $3,+ graph_status = coalesce($4, graph_status),+ graph_error = CASE WHEN $4 IS NULL THEN graph_error ELSE NULL END,+ updated_at = now() WHERE id = $1",+ )+ .bind(input.document.id)+ .bind(text_len)+ .bind(chunk_count)+ .bind(graph_status)+ .execute(&mut *tx)+ .await?;+ if graph_status == Some("queued") {+ crate::jobs::enqueue_with_max_attempts_tx(+ &mut tx,+ "extract_document",+ serde_json::json!({ "document_id": input.document.id }),+ 3,+ )+ .await?;+ }+ tx.commit().await?;+ Ok(true)+}+/// 一次删除的产出,给审计与界面提示用。#[derive(Debug, Clone, Copy, PartialEq, Eq)]pub struct DeletionReport {@@ -1407,12 +1522,17 @@ pub async fn replace_chunks(/// 返回 None 表示输入已过期,调用方应结束本次处理。pub async fn replace_chunks_if_current(pool: &PgPool,- kb_id: Uuid,- document_id: Uuid,+ input: &ProcessingInput,pieces: &[ChunkPiece],- sha256: &str,) -> AppResult<Option<Vec<(String, String)>>> {- replace_chunks_for_snapshot(pool, kb_id, document_id, pieces, Some(sha256)).await+ replace_chunks_for_snapshot(+ pool,+ input.document.kb_id,+ input.document.id,+ pieces,+ Some(input),+ )+ .await}async fn replace_chunks_for_snapshot(@@ -1420,7 +1540,7 @@ async fn replace_chunks_for_snapshot(kb_id: Uuid,document_id: Uuid,pieces: &[ChunkPiece],- sha256: Option<&str>,+ input: Option<&ProcessingInput>,) -> AppResult<Option<Vec<(String, String)>>> {let mut tx = pool.begin().await?;// 两个任务可能同时处理同一文档。先锁父记录,即使还没有分块,@@ -1430,16 +1550,15 @@ async fn replace_chunks_for_snapshot(// 冲突,于是这个事务活着的时候,这份文档所有子表的插入都被挡住(chunks、// document_versions、memory::append),而这个事务是每个分块一个来回——四千块的// 文档要锁四秒。两者对另一个 `replace_chunks` 的互斥是一样的- let current: Option<(String, bool)> = sqlx::query_as(- "SELECT sha256, deleted_at IS NOT NULL FROM documents WHERE id = $1 FOR NO KEY UPDATE",- )- .bind(document_id)- .fetch_optional(&mut *tx)- .await?;- if let Some(expected) = sha256 {- if !matches!(current, Some((ref sha, false)) if sha == expected) {+ if let Some(input) = input {+ if !lock_processing_input_tx(&mut tx, input).await? {return Ok(None);}+ } else {+ sqlx::query("SELECT id FROM documents WHERE id = $1 FOR NO KEY UPDATE")+ .bind(document_id)+ .execute(&mut *tx)+ .await?;}let (version,): (i32,) = sqlx::query_as("SELECT COALESCE(MAX(version), 1) FROM document_versions WHERE document_id = $1",@@ -1705,7 +1824,29 @@ pub async fn chunks_pending_embedding(}pub async fn set_embeddings(pool: &PgPool, items: &[(Uuid, Vec<f32>)]) -> AppResult<()> {+ set_embeddings_for_input(pool, items, None).await?;+ Ok(())+}++pub async fn set_embeddings_if_current(+ pool: &PgPool,+ input: &ProcessingInput,+ items: &[(Uuid, Vec<f32>)],+) -> AppResult<bool> {+ set_embeddings_for_input(pool, items, Some(input)).await+}++async fn set_embeddings_for_input(+ pool: &PgPool,+ items: &[(Uuid, Vec<f32>)],+ input: Option<&ProcessingInput>,+) -> AppResult<bool> {let mut tx = pool.begin().await?;+ if let Some(input) = input {+ if !lock_processing_input_tx(&mut tx, input).await? {+ return Ok(false);+ }+ }for (id, emb) in items {sqlx::query("UPDATE chunks SET embedding = $2 WHERE id = $1").bind(id)@@ -1719,7 +1860,7 @@ pub async fn set_embeddings(pool: &PgPool, items: &[(Uuid, Vec<f32>)]) -> AppRescrate::vector_index::request(pool, crate::vector_index::Target::Chunks, first.len()).await?;}- Ok(())+ Ok(true)}/// 向量近邻检索(余弦距离)。维度写成字面量、两侧 cast、`relaxed_order`——三条