Home
/
/
/

When using some applications and tools, including Webpack or code-server, you may encounter an error similar to the following:

Watchpack Error (watcher): Error: ENOSPC: System limit for number of file watchers reached, watch '/some/path'

This article will show you how to diagnose and troubleshoot this error, which relates to a high number of inotify watchers in use.

Background

inotify allows programs to monitor files for changes, so that they receive an event whenever a user or program modifies a file. inotify requires kernel resources (memory and processor) for each file it tracks. As a result, the Linux kernel limits the number of file watchers that each user can register. The default settings vary according to the host system distribution; on Ubuntu 20.04 LTS, the default limit is 8,192 watches per instance.

On a 64-bit system, each inotify watch that programs register will consume ~1 kB of kernel memory, which cannot be swapped to disk and is not counted against the workspace memory limit setting.

Diagnosis

If you encounter the error that's the focus of this article, the total number of watchers in use is approaching the max_user_watches setting. The following sections will show you how to verify if this is the case.

Check tunable settings

There are three kernel tuning options related to the inotify system:

  • fs.inotify.max_queued_events: The upper bound on the number of file notification events pending delivery to programs
  • fs.inotify.max_user_instances: The maximum number of inotify instances per user (programs using inotify will typically create a single instance, so this limit is unlikely to cause issues)
  • fs.inotify.max_user_watches: The maximum number of files and folders that programs can monitor for changes

To see the values for these settings that are applicable to your workspace, run:

$ sysctl fs.inotify.{max_queued_events,max_user_instances,max_user_watches}
fs.inotify.max_queued_events = 16384
fs.inotify.max_user_instances = 128
fs.inotify.max_user_watches = 8192

Because these settings are not namespace-aware, the values will be the same regardless of whether you run the commands on the host system or inside a container running on that host.

See inotify(7) for additional details regarding the inotify system.

Identify inotify consumers

To identify the programs consuming inotify watches, you can use a script that summarizes the information available in the /proc filesystem, such as inotify-consumers.

This script will show the names of programs along with the number of inotify watches registered with the kernel:

$ ./inotify-consumers
   INOTIFY
   WATCHER
    COUNT     PID USER     COMMAND
--------------------------------------
     269   254560 coder    /var/tmp/coder/code-server/lib/node /var/tmp/coder/code-server/lib/vscode/out/bootstrap-fork --type=watcherService
       5     1722 coder    /var/tmp/coder/code-server/lib/node /var/tmp/coder/code-server/lib/vscode/out/vs/server/fork
       2   254538 coder    /var/tmp/coder/code-server/lib/node /var/tmp/coder/code-server/lib/vscode/out/bootstrap-fork --type=extensionHost
       2     1507 coder    gpg-agent --homedir /home/coder/.gnupg --use-standard-socket --daemon

     278  WATCHERS TOTAL COUNT

Please note that this is a third-party script published by an individual who is not affiliated with Coder, and as such, we cannot provide a warranty or support for its usage.

To see the specific files that the tools track for changes, you can use strace to monitor invocations of the inotify_add_watch system call:

$ strace --follow-forks --trace='inotify_add_watch' inotifywait --quiet test
inotify_add_watch(3, "test", IN_ACCESS|IN_MODIFY|IN_ATTRIB|IN_CLOSE_WRITE|IN_CLOSE_NOWRITE|IN_OPEN|IN_MOVED_FROM|IN_MOVED_TO|IN_CREATE|IN_DELETE|IN_DELETE_SELF|IN_MOVE_SELF) = 1

This example shows that the inotifywait command is listening for notifications related to the test file.

Resolution

If you encounter the file watcher limit, you can do one of two things:

  1. Reduce the number of file watcher registrations
  2. Increase the maximum file watcher limit

We recommend attempting to reduce the file watcher registrations first, because increasing the number of file watches may result in high processor utilization.

Reduce watchers

Many applications include files that change rarely (e.g., third-party dependencies stored in node_modules). Your tools may watch for changes to these files and folders, consuming inotify watchers. These tools typically provide configuration settings to exclude specific files, paths, and patterns from file watching.

For example, Visual Studio Code and code-server apply the following user workspace setting by default:

"files.watcherExclude": {
  "**/.git/objects/**": true,
  "**/.git/subtree-cache/**": true,
  "**/node_modules/**": true,
  "**/.hg/store/**": true
},

Consider adding other infrequently-changed files to this list, which will cause Visual Studio Code to poll (or check periodically) for changes to those files.

For information on how to do this with other software tools, please see their documentation/user manuals.

Increase the watch limit

You can increase the kernel tunable option to increase the maximum number of inotify watches for each user. This is a global setting that applies to all users sharing the same system/Kubernetes node. To do this, modify the sysctl configuration file, or apply a DaemonSet to the Kubernetes cluster to apply that change to all nodes automatically.

For example, you can create a file called /etc/sysctl.d/watches.conf and include the following contents:

fs.inotify.max_user_watches = 65536

Alternatively, you can use the following DaemonSet with kubectl apply:

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: more-fs-watchers
  namespace: kube-system
  labels:
    app: more-fs-watchers
    k8s-app: more-fs-watchers
spec:
  selector:
    matchLabels:
      k8s-app: more-fs-watchers
  template:
    metadata:
      labels:
        name: more-fs-watchers
        k8s-app: more-fs-watchers
      annotations:
        seccomp.security.alpha.kubernetes.io/defaultProfileName: runtime/default
        apparmor.security.beta.kubernetes.io/defaultProfileName: runtime/default
    spec:
      nodeSelector:
        kubernetes.io/os: linux
      initContainers:
        - name: sysctl
          image: alpine:3
          command:
            - sysctl
            - -w
            - fs.inotify.max_user_watches=16384
          resources:
            requests:
              cpu: 10m
              memory: 1Mi
            limits:
              cpu: 100m
              memory: 5Mi
          securityContext:
            # We need to run as root in a privileged container to modify
            # /proc/sys on the host (for sysctl)
            runAsUser: 0
            privileged: true
            readOnlyRootFilesystem: true
            capabilities:
              drop:
                - ALL
      containers:
        - name: pause
          image: k8s.gcr.io/pause:3.5
          command:
            - /pause
          resources:
            requests:
              cpu: 10m
              memory: 1Mi
            limits:
              cpu: 100m
              memory: 5Mi
          securityContext:
            runAsNonRoot: true
            runAsUser: 65535
            allowPrivilegeEscalation: false
            privileged: false
            readOnlyRootFilesystem: true
            capabilities:
              drop:
                - ALL
      terminationGracePeriodSeconds: 5

This DaemonSet will ensure that the corresponding pod runs on every Linux node in the cluster. When new nodes join the cluster, such as during an autoscaling event, the DaemonSet will ensure that the pod runs on the new node as well.

You can delete the DaemonSet by running:

$ kubectl delete --namespace=kube-system daemonset more-fs-watchers
daemonset.apps "more-fs-watchers" deleted

However, note that the setting will persist until the node restarts or another program sets the fs.inotify.max_user_watches setting.

See also

If none of these steps resolve the issue, please contact us for further support.

See an opportunity to improve our docs? Make an edit.