A config file that worked five seconds ago is now a 347 byte fragment of broken JSON. Your CLI tool will not start. The

A config file that worked five seconds ago is now a 347-byte fragment of broken JSON. Your CLI tool will not start. The user, who simply pressed Ctrl-C because the update was taking longer than expected, is now staring at an error stack trace they did not ask for. The two kilobytes of valid configuration that existed before the write are gone, replaced by the digital equivalent of a half-printed receipt.

This happens because writeFile is not atomic. It opens the existing path, truncates it, streams data from Node into the kernel’s page cache, and eventually closes the file descriptor. The moment the truncation happens, the old content is already gone. Everything between that truncation and the final close is a window of vulnerability. A SIGINT, a power outage, or a laptop lid slamming shut during that window leaves the filesystem holding a truncated mess. Even if Node reports the Promise as resolved, the operating system may still be buffering writes in memory. Convenience methods hide that gap, but they do not remove it.

The fix is not to write in place. The fix is to separate the act of writing from the act of publishing.

Write to a sibling, then swap

The reliable pattern has five steps. None of them are complicated, but together they move the failure window from an entire streaming write down to a single filesystem metadata operation.

First, serialize the entire payload in memory. Do this before you create any temporary file. If JSON.stringify throws because someone passed a circular object, you want that exception to bubble up before you touch the disk.

Second, write the serialized data to a temporary file located in the same directory as the target. Use a randomized name so two concurrent runs do not collide. Keeping the temp file in the same directory matters because rename is only atomic within a single filesystem. If your temp file lives on a different partition, the operating system falls back to a copy-and-delete sequence, which introduces its own failure modes and is no longer atomic.

Third, ask the kernel to flush that temporary file to physical storage. Node’s fsync, exposed here as the sync method on a filehandle, blocks until the buffers are down on the metal. This is slow, but config writes happen rarely enough that the durability is worth the milliseconds.

Fourth, rename the temporary file over the original path. On both POSIX systems and Windows, this is the commit point. Readers opening the original path will see either the complete old file or the complete new file. There is no moment in time where a reader can open the path and observe a half-written buffer.

Fifth, sync the parent directory. This catches a subtle edge case. The rename updates the directory entry, but the directory metadata itself might sit in the kernel’s page cache. Sudden power loss after a successful rename can sometimes leave the filesystem in a state where the new inode reference was never recorded durably. Syncing the directory forces that metadata update to disk and seals the transaction.

A concrete Node.js implementation

Here is what that pattern looks like in practice using only the Node.js standard library:

import { open, rename, rm } from "node:fs/promises";
import { dirname, basename, join } from "node:path";
import { randomUUID } from "node:crypto";

export async function writeJsonAtomic(path, value) {
  const directory = dirname(path);
  const temporary = join(directory, `.${basename(path)}.${randomUUID()}.tmp`);
  const body = `${JSON.stringify(value, null, 2)}\n`;
  let handle;

  try {
    handle = await open(temporary, "wx", 0o600);
    await handle.writeFile(body, "utf8");
    await handle.sync();
    await handle.close();
    handle = undefined;

    await rename(temporary, path);

    const directoryHandle = await open(directory, "r");
    try {
      await directoryHandle.sync();
    } finally {
      await directoryHandle.close();
    }
  } catch (error) {
    if (handle) await handle.close().catch(() => {});
    await rm(temporary, { force: true }).catch(() => {});
    throw error;
  }
}

A few details in this code are worth attention.

The wx flag means “write, but fail if the file already exists.” This guards against a UUID collision or an abandoned temp file from a previous crashed process. If someone has dropped a malicious file where your temp file should be, you will hear about it immediately rather than overwriting whatever is there.

The 0o600 permission mask creates the temp file with owner-read and owner-write only. Config files frequently hold secrets, API tokens, or private repository URLs. There is no reason to let other users on the system peek at the temporary file while it is being prepared.

Notice the separate sync calls on the file and then on the directory. Many developers skip the directory sync because it feels redundant. It is not. Ext4, APFS, and NTFS all handle directory updates differently, but they share a common habit of batching metadata writes for performance. If you care about surviving power loss, the directory sync is the final seal.

The cleanup in the catch block is intentionally defensive. If anything throws after the filehandle is opened, the code attempts to close the handle and remove the temporary file, swallowing any secondary errors so the original exception propagates cleanly. You do not want a permission error during cleanup to mask the real bug that caused the failure.

Where this pattern stops helping

Atomic file replacement prevents torn writes. It does not prevent lost updates. If two instances of your CLI read the same config simultaneously, both perform their edits in memory, both write fresh temp files, and both execute their renames, the second rename wins. The first process did not observe the second process’s changes. Depending on your application, this could mean a user adds a setting in one terminal and a different user removes it in another, with the final file reflecting only the last writer.

If your tool needs to support concurrent mutators, you need a coordination mechanism on top of the atomic write. An advisory lock file works for simple cases. Version vectors or a monotonic revision number inside the config itself can help detect collisions so the second writer can retry. These add complexity, and complexity is where bugs hide.

That is why the boundary matters. A single JSON blob that one process updates occasionally is a fine candidate for an atomic file write. Once you find yourself managing multiple records, enforcing schemas, or worrying about concurrent mutations, you have outgrown the filesystem. SQLite exists for exactly this reason. It gives you atomic transactions, rollback journals, and proper handling of concurrent readers and writers, all inside a single host-local file. A clever file protocol is not a database, and you should not spend maintenance budget pretending otherwise.

The real takeaway

The next time you reach for writeFile inside a CLI tool, pause. Serialization is not the hard part. Durability is. Config files are too small to stream and too important to truncate. Write the entire payload to a hidden sibling, flush it, commit it with a rename, and tell the directory about it. Your users can press Ctrl-C, yank the power cord, or close their laptop lid. When the machine comes back, the file will either contain the old world or the new one. Nothing in between.