[7.4.0] Implement disk cache garbage collection.

.bazelci/postsubmit.yml

@@ -18,6 +18,7 @@ tasks:
       - "//src:bazel_jdk_minimal"
       - "//src:test_repos"
       - "//src/main/java/..."
+      - "//src/tools/diskcache/..."
       - "//src/tools/execlog/..."
     test_flags:
       - "--config=ci-linux"

.bazelci/presubmit.yml

@@ -19,6 +19,7 @@ tasks:
       - "//src:bazel_jdk_minimal"
       - "//src:test_repos"
       - "//src/main/java/..."
+      - "//src/tools/diskcache/..."
       - "//src/tools/execlog/..."
     test_flags:
       - "--config=ci-linux"

site/en/remote/caching.md

@@ -292,9 +292,8 @@ This feature is unsupported on Windows.
 
 Bazel can use a directory on the file system as a remote cache. This is
 useful for sharing build artifacts when switching branches and/or working
-on multiple workspaces of the same project, such as multiple checkouts. Since
-Bazel does not garbage-collect the directory, you might want to automate a
-periodic cleanup of this directory. Enable the disk cache as follows:
+on multiple workspaces of the same project, such as multiple checkouts.
+Enable the disk cache as follows:
 
 ```posix-terminal
 build --disk_cache={{ '<var>' }}path/to/build/cache{{ '</var>' }}
@@ -305,6 +304,18 @@ You can pass a user-specific path to the `--disk_cache` flag using the `~` alias
 when enabling the disk cache for all developers of a project via the project's
 checked in `.bazelrc` file.
 
+### Garbage collection {:#disk-cache-gc}
+
+Starting with Bazel 7.4, you can use `--experimental_disk_cache_gc_max_size` and
+`--experimental_disk_cache_gc_max_age` to set a maximum size for the disk cache
+or for the age of individual cache entries. Bazel will automatically garbage
+collect the disk cache while idling between builds; the idle timer can be set
+with `--experimental_disk_cache_gc_idle_delay` (defaulting to 5 minutes).
+
+As an alternative to automatic garbage collection, we also provide a [tool](
+https://github.com/bazelbuild/bazel/tree/master/src/tools/diskcache) to run a
+garbage collection on demand.
+
 ## Known issues {:#known-issues}
 
 **Input file modification during a build**

src/BUILD

@@ -379,6 +379,7 @@ filegroup(
         "//src/tools/android:srcs",
         "//src/tools/android/java/com/google/devtools/build/android:srcs",
         "//src/tools/bzlmod:srcs",
+        "//src/tools/diskcache:srcs",
         "//src/tools/execlog:srcs",
         "//src/tools/launcher:srcs",
         "//src/tools/remote:srcs",

src/main/java/com/google/devtools/build/lib/remote/RemoteModule.java

@@ -64,6 +64,8 @@
 import com.google.devtools.build.lib.remote.circuitbreaker.CircuitBreakerFactory;
 import com.google.devtools.build.lib.remote.common.RemoteCacheClient;
 import com.google.devtools.build.lib.remote.common.RemoteExecutionClient;
+import com.google.devtools.build.lib.remote.disk.DiskCacheClient;
+import com.google.devtools.build.lib.remote.disk.DiskCacheGarbageCollectorIdleTask;
 import com.google.devtools.build.lib.remote.downloader.GrpcRemoteDownloader;
 import com.google.devtools.build.lib.remote.http.DownloadTimeoutException;
 import com.google.devtools.build.lib.remote.http.HttpException;
@@ -335,6 +337,14 @@ public void beforeCommand(CommandEnvironment env) throws AbruptExitException {
       }
     }
 
+    if (enableDiskCache) {
+      var gcIdleTask =
+          DiskCacheGarbageCollectorIdleTask.create(remoteOptions, env.getWorkingDirectory());
+      if (gcIdleTask != null) {
+        env.addIdleTask(gcIdleTask);
+      }
+    }
+
     if (!enableDiskCache && !enableHttpCache && !enableGrpcCache && !enableRemoteExecution) {
       // Quit if no remote caching or execution was enabled.
       actionContextProvider =
@@ -954,9 +964,9 @@ public void afterCommand() {
   }
 
   private static void afterCommandTask(
-      RemoteActionContextProvider actionContextProvider,
-      TempPathGenerator tempPathGenerator,
-      AsynchronousMessageOutputStream<LogEntry> rpcLogFile)
+      @Nullable RemoteActionContextProvider actionContextProvider,
+      @Nullable TempPathGenerator tempPathGenerator,
+      @Nullable AsynchronousMessageOutputStream<LogEntry> rpcLogFile)
       throws AbruptExitException {
     if (actionContextProvider != null) {
       actionContextProvider.afterCommand();

src/main/java/com/google/devtools/build/lib/remote/disk/BUILD

@@ -15,12 +15,17 @@ java_library(
     name = "disk",
     srcs = glob(["*.java"]),
     deps = [
+        "//src/main/java/com/google/devtools/build/lib/concurrent",
         "//src/main/java/com/google/devtools/build/lib/exec:spawn_runner",
         "//src/main/java/com/google/devtools/build/lib/remote:store",
         "//src/main/java/com/google/devtools/build/lib/remote/common",
         "//src/main/java/com/google/devtools/build/lib/remote/common:cache_not_found_exception",
+        "//src/main/java/com/google/devtools/build/lib/remote/options",
         "//src/main/java/com/google/devtools/build/lib/remote/util",
+        "//src/main/java/com/google/devtools/build/lib/server:idle_task",
+        "//src/main/java/com/google/devtools/build/lib/util:string",
         "//src/main/java/com/google/devtools/build/lib/vfs",
+        "//third_party:flogger",
         "//third_party:guava",
         "//third_party:jsr305",
         "//third_party/protobuf:protobuf_java",

src/main/java/com/google/devtools/build/lib/remote/disk/DiskCacheClient.java

@@ -60,8 +60,13 @@
  * when they collide.
  *
  * <p>The mtime of an entry reflects the most recent time the entry was stored *or* retrieved. This
- * property may be used to trim the disk cache to the most recently used entries. However, it's not
- * safe to trim the cache at the same time a Bazel process is accessing it.
+ * property may be used to garbage collect the disk cache by deleting the least recently accessed
+ * entries. This may be done by Bazel itself (see {@link DiskCacheGarbageCollectorIdleTask}), by
+ * another Bazel process sharing the disk cache, or by an external process. Although we could have
+ * arranged for an ongoing garbage collection to block a concurrent build, we judge it to not be
+ * worth the extra complexity; assuming that the collection policy is not overly aggressive, the
+ * likelihood of a race condition is fairly small, and an affected build is able to automatically
+ * recover by retrying.
  */
 public class DiskCacheClient implements RemoteCacheClient {