Document max concurrent writes (TraceMachina#2169)

* Document max concurrent filesystem writes

* Update config example from JSON to JSON5

Changed code block syntax from json to json5 for better compatibility.

Author: MarcusSorealheis

web/platform/src/content/docs/docs/config/basic-configs.mdx

@@ -26,6 +26,9 @@ memory and filesystem stores instead of S3 and Redis.
       "filesystem": {
         "content_path": "/tmp/nativelink/data-worker-test/content_path-ac",
         "temp_path": "/tmp/nativelink/data-worker-test/tmp_path-ac",
+        // Limit concurrent writes to avoid disk I/O saturation.
+        // Set to 0 for unlimited.
+        "max_concurrent_writes": 8,
         "eviction_policy": {
           // 1gb.
           "max_bytes": 1000000000,
@@ -40,6 +43,9 @@ memory and filesystem stores instead of S3 and Redis.
           "filesystem": {
             "content_path": "/tmp/nativelink/data-worker-test/content_path-cas",
             "temp_path": "/tmp/nativelink/data-worker-test/tmp_path-cas",
+            // Limit concurrent writes to avoid disk I/O saturation.
+            // Set to 0 for unlimited.
+            "max_concurrent_writes": 8,
             "eviction_policy": {
               // 10gb.
               "max_bytes": 10000000000,

web/platform/src/content/docs/docs/config/production-config.mdx

@@ -454,6 +454,20 @@ Here is the final CAS Config JSON without the 99 extra shards for writing to S3.
 }
 ```
 
+## Limit Concurrent Filesystem Writes
+
+If the CAS server is under heavy disk I/O pressure, cap the number of
+simultaneous writes to the filesystem store with `max_concurrent_writes`.
+Each write streams to a temp file and calls `sync_all()`, which can saturate
+the disk and block the async runtime if too many happen at once.
+
+```json5
+"filesystem": {
+  "content_path": "/var/lib/nativelink/cas",
+  "temp_path": "/var/lib/nativelink/tmp",
+  // Set to 0 for unlimited.
+  "max_concurrent_writes": 16
+}
 ## Limit Worker Inflight Tasks
 
 If your workers are getting saturated, cap the number of concurrent tasks they