diff --git a/Dockerfile b/Dockerfile
index a592c193..b026b62d 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -36,4 +36,4 @@ RUN if echo "$BASE_IMAGE" | grep -q "node-caged"; then \
 			echo "export NODE_CAGED=false" > /app/.runtime-env; \
 		fi
 
-CMD ["sh", "-c", ". /app/.runtime-env && exec node --expose-gc /app/server/index.mjs"]
\ No newline at end of file
+CMD ["sh", "-c", ". /app/.runtime-env && exec node /app/server/index.mjs"]
\ No newline at end of file
diff --git a/README.md b/README.md
index 90307017..32120c11 100644
--- a/README.md
+++ b/README.md
@@ -28,6 +28,10 @@ volumes:
   cache-data:
 ```
 
+Multipart downloads and background merges use stream backpressure and normal V8 garbage collection.
+The server does not require Node's `--expose-gc` flag or force collection between parts.
+Aggregate memory usage still depends on workload concurrency and configured storage buffers.
+
 ## Documentation
 
 👉 <https://gha-cache-server.falcondev.io/getting-started> 👈
diff --git a/lib/storage.ts b/lib/storage.ts
index 02a4519d..9296766c 100644
--- a/lib/storage.ts
+++ b/lib/storage.ts
@@ -150,8 +150,6 @@ export class Storage {
 
     responseStream.end()
     mergerStream.end()
-
-    await globalThis.gc?.()
   }
 
   private async *streamParts(location: StorageLocation, signal?: AbortSignal) {
@@ -170,8 +168,6 @@ export class Storage {
       } finally {
         partStream.destroy()
       }
-
-      await globalThis.gc?.()
     }
   }
 
diff --git a/plugins/setup.ts b/plugins/setup.ts
index 783d73e2..b9975b34 100644
--- a/plugins/setup.ts
+++ b/plugins/setup.ts
@@ -12,11 +12,6 @@ export default defineNitroPlugin(async (nitro) => {
   if (cluster.isPrimary) {
     logger.info(`🚀 Starting GitHub Actions Cache Server (${version})`)
 
-    if (!globalThis.gc)
-      logger.warn(
-        'Garbage collection is not exposed. Start the process with `node --expose-gc` for improved memory usage under high load.',
-      )
-
     for (const signal of ['SIGTERM', 'SIGINT'] satisfies NodeJS.Signals[]) {
       process.on(signal, () => {
         const workers = Object.values(cluster.workers ?? {})
diff --git a/tests/cleanup-lifecycle.test.ts b/tests/cleanup-lifecycle.test.ts
index 73dba28c..82b26261 100644
--- a/tests/cleanup-lifecycle.test.ts
+++ b/tests/cleanup-lifecycle.test.ts
@@ -12,6 +12,89 @@ describe('cleanup lifecycle', () => {
   })
   afterAll(() => vi.unstubAllGlobals())
 
+  test('concurrent multipart restores preserve exact bytes before and after merging', async () => {
+    const db = await getDatabase()
+    const storage = await Storage.fromEnv()
+    const entries = Array.from({ length: 4 }, (_, entryIndex) => ({
+      key: randomUUID(),
+      version: 'v1',
+      scope: 'refs/heads/main',
+      repoId: '123',
+      parts: Array.from({ length: 8 }, (_, partIndex) =>
+        Buffer.from(
+          Array.from(
+            { length: 16 * 1024 + partIndex },
+            (_, offset) => (entryIndex * 31 + partIndex * 17 + offset) % 256,
+          ),
+        ),
+      ),
+    }))
+
+    try {
+      const caches = []
+      for (const entry of entries) {
+        const upload = await storage.createUpload(entry)
+        expect(upload).toBeDefined()
+        for (const [partIndex, part] of entry.parts.entries())
+          await storage.uploadPart(upload!.id, partIndex, Readable.from(part))
+        await storage.completeUpload(entry)
+        const cache = await storage.matchCacheEntry({
+          keys: [entry.key],
+          version: entry.version,
+          scopes: [entry.scope],
+          repoId: entry.repoId,
+        })
+        expect(cache).toBeDefined()
+        caches.push({ cache: cache!.match, expected: Buffer.concat(entry.parts) })
+      }
+
+      for (const phase of ['cold', 'warm']) {
+        await Promise.all(
+          caches.map(async ({ cache, expected }) => {
+            const stream = await storage.download(cache.id)
+            expect(stream, phase).toBeDefined()
+            const chunks: Buffer[] = []
+            for await (const chunk of stream!) chunks.push(Buffer.from(chunk))
+            expect(Buffer.concat(chunks), phase).toEqual(expected)
+          }),
+        )
+        await storage.waitForOngoingMerges()
+        for (const { cache } of caches) {
+          const location = await db
+            .selectFrom('storage_locations')
+            .where('id', '=', cache.locationId)
+            .select('mergedAt')
+            .executeTakeFirstOrThrow()
+          expect(location.mergedAt, phase).not.toBeNull()
+        }
+      }
+    } finally {
+      await storage.waitForOngoingMerges()
+      const locations = await db
+        .selectFrom('cache_entries')
+        .innerJoin('storage_locations', 'storage_locations.id', 'cache_entries.locationId')
+        .where(
+          'cache_entries.key',
+          'in',
+          entries.map(({ key }) => key),
+        )
+        .select(['storage_locations.id', 'storage_locations.folderName'])
+        .execute()
+      for (const { id, folderName } of locations) {
+        await db.deleteFrom('storage_locations').where('id', '=', id).execute()
+        await storage.adapter.deleteFolder(folderName)
+      }
+      await db
+        .deleteFrom('uploads')
+        .where(
+          'key',
+          'in',
+          entries.map(({ key }) => key),
+        )
+        .execute()
+    }
+  })
+
   test('evicts least-recently-used entries to 90% only after exceeding the budget', async () => {
     const db = await getDatabase()
     const storage = await Storage.fromEnv()