fix(base): enable duckdb disk spill + raise memory default to avoid oom on large bases

apps/server/src/core/base/query-cache/base-query-cache.integration.spec.ts

@@ -65,6 +65,9 @@ class FakeEnvService {
   getBaseQueryCacheThreads() {
     return 2;
   }
+  getBaseQueryCacheTempDirectory() {
+    return require('node:os').tmpdir() + '/docmost-duckdb-test';
+  }
   getRedisUrl() {
     return REDIS_URL;
   }
@@ -453,6 +456,9 @@ describeIntegration('BaseQueryCacheService LRU eviction', () => {
     getBaseQueryCacheThreads() {
       return 2;
     }
+    getBaseQueryCacheTempDirectory() {
+      return require('node:os').tmpdir() + '/docmost-duckdb-test';
+    }
     getRedisUrl() {
       return REDIS_URL;
     }
@@ -1101,6 +1107,9 @@ describeIntegration('BaseQueryCacheService warm-up on boot', () => {
     getBaseQueryCacheThreads() {
       return 2;
     }
+    getBaseQueryCacheTempDirectory() {
+      return require('node:os').tmpdir() + '/docmost-duckdb-test';
+    }
     getRedisUrl() {
       return REDIS_URL;
     }

apps/server/src/core/base/query-cache/collection-loader.ts

@@ -36,16 +36,34 @@ export class CollectionLoader {
     const properties = await this.basePropertyRepo.findByBaseId(baseId);
     const specs = buildColumnSpecs(properties);
 
-    const { memoryLimit, threads } = this.config.config;
+    const { memoryLimit, threads, tempDirectory } = this.config.config;
+
+    // Ensure the temp directory exists so DuckDB can spill to it.
+    // Swallow errors — if creation fails, DuckDB will fail its own sanity
+    // check and we'll log that instead of crashing here.
+    try {
+      const fs = require('node:fs');
+      fs.mkdirSync(tempDirectory, { recursive: true });
+    } catch {
+      /* swallow */
+    }
+
     const instance = await DuckDBInstance.create(':memory:', {
       memory_limit: memoryLimit,
       threads: String(threads),
+      temp_directory: tempDirectory,
     });
     const connection = await instance.connect();
 
     try {
       await this.pgExtension.configureOnConnection(connection);
 
+      // Disable insertion-order preservation during bulk load — DuckDB's docs
+      // explicitly recommend this for memory-pressure on large inserts. Our
+      // loader doesn't depend on the insertion order (we sort via indexes
+      // or keyset cursors later), so this is free memory savings.
+      await connection.run('SET preserve_insertion_order = false');
+
       // Bulk load via CREATE TABLE AS SELECT. JSONB extraction happens
       // server-side via the base_cell_* helpers; DuckDB streams typed
       // columns over COPY BINARY into its vectorized insert path.

apps/server/src/core/base/query-cache/query-cache.config.ts

@@ -9,6 +9,7 @@ export type QueryCacheConfig = {
   memoryLimit: string;
   threads: number;
   trace: boolean;
+  tempDirectory: string;
 };
 
 @Injectable()
@@ -23,6 +24,7 @@ export class QueryCacheConfigProvider {
       memoryLimit: env.getBaseQueryCacheMemoryLimit(),
       threads: env.getBaseQueryCacheThreads(),
       trace: env.getBaseQueryCacheTrace(),
+      tempDirectory: env.getBaseQueryCacheTempDirectory(),
     };
   }
 }