About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.
A file-level incremental backup decides which whole files changed and copies those files. A block-level incremental backup asks the disk, or the platform under it, which regions of the disk changed and copies only those regions. The difference is not only the amount copied. It is also how the backup finds out what changed, and what happens when that source of information is wrong or gets reset.
If you chose a product by the word "incremental", the question to ask is: incremental according to what? The answer decides how long the nightly job takes, how much it stores, and which faults make it quietly copy everything again.
Why does an incremental backup need to know what changed?
An incremental backup does two separate jobs. First it works out what to read from the source. Then it works out what to store. Many products treat these as one step, but they are different, and they can use different methods.
The slow way to find changes is to read everything and compare. Microsoft describes this plainly in its change journal documentation: scanning the entire volume is the brute force method, "often not an acceptable approach because of the decrease in system performance it would cause". Every cheaper method is a way of avoiding that full read.
How do file-level methods decide a file changed?
File-level methods work from file metadata. Four methods are described here.
The archive attribute. Windows keeps an Archive flag on each file. Microsoft's attrib page says that setting it "marks files that have changed since the last time they were backed up. The xcopy command uses archive attributes." A tool copies flagged files, then clears the flag. A later write sets it again. The flag is simple, and it has pitfalls of its own, especially when two tools share it.
Timestamps. robocopy can select files by modified time with /maxage:<n>, which Microsoft describes as excluding "files older than n days or a date based on when the files were last modified". xcopy /d copies files "changed on or after the specified date". Timestamp methods are only as honest as the clock and the software that sets file times. A restored or copied file can keep an old modified time and be missed.
Metadata comparison against the last backup. restic documents the rule it uses on Unix systems: a file at the same path is presumed unchanged if modification time, change time (ctime), size and inode number all match. It explains why it checks ctime as well as mtime: "Unix programs can freely change mtime (and some do)". On Windows the rule is simpler: path, size and modification time. The restic manual notes the cost of the shortcut: change detection is skipped for renamed files, so a renamed folder is read again in full.
The NTFS change journal. NTFS can keep an update sequence number (USN) change journal for each volume. Microsoft says that "when any change is made to a file or directory in a volume, the USN change journal for that volume is updated with a description of the change and the name of the file or directory". A backup program reads the journal instead of walking the folder tree.
What does the NTFS change journal give you, and what does it not?
The journal is a list of events, not a copy of the data. Microsoft's page on change journal records states three limits:
- It records the fact and reason for a change, not enough to reverse it.
- Repeated changes of the same kind may collapse into one record. Several writes with no close in between "result in only one change record".
- NTFS "may delete old records to conserve space". Microsoft says the indexing service then re-indexes the volume; a backup tool that needs the lost records would need its own full scan.
Microsoft also documents that each journal has an identifier, and that NTFS "restamps" the journal with a new identifier when existing records "either are or may be unusable". A backup tool that stored a position in the old journal must treat a new identifier as "start again", which means a full read. A restamped journal makes the tool read everything again.
How do block-level methods decide what changed?
Block-level methods sit below the file system. They do not know about files at all. They track which blocks of a virtual or physical disk were written.
VMware changed block tracking (CBT). The vSphere API method QueryChangedDiskAreas returns "a list of areas of a virtual disk belonging to this VM that have been modified since a well-defined point in the past". The start of that period is a changeId the backup program saved at the previous backup. Broadcom's reference describes the typical use: a backup application obtains the changeId from the disk's backing information at backup time, then uses it next time to ask for changes. The reference also states the guarantee: results "may contain false positives (i.e: flag areas of the disk as modified that are not). However, it is guaranteed that no changes will be missed." Whether CBT is on for a VM is a property, changeTrackingEnabled. Resets and the full reads that follow them have their own topic: "Incremental backups suddenly run as full after a CBT reset".
Hyper-V resilient change tracking (RCT). From Windows Server 2016 Hyper-V can back up through its WMI interface. Microsoft says this "still utilizes VSS inside the virtual machine" but not in the host, and "a combination of reference points and resilient change tracking (RCT) is used". Microsoft lists limits of the Win32 route to the changed data: it can only be accessed locally and does not support shared virtual hard disk files.
Chunk-level comparison by the backup software. Some tools find changes by reading and splitting everything, then keeping only what is new. restic splits files into variable-length blobs of 512 KiB to 8 MiB, averaging about 1 MiB. Proxmox Backup Server documents that "the actual upload of Snapshots is incremental while each Snapshot references all chunks and is thus a full backup". These tools can still read the whole source unless a faster test, such as the file metadata rule above, lets them skip unchanged files.
Which method fits which situation?
| Method | What it tracks | What it copies | Weak point |
|---|---|---|---|
| Archive attribute | A flag per file | Whole changed files | Two tools sharing the flag; files that change without setting it |
| Timestamps | Modified time | Whole changed files | Restored or copied files with old times; clock errors |
| NTFS change journal | Events per volume | Whole changed files | Old records discarded; journal restamped or re-created |
| VMware CBT | Changed disk areas since a changeId |
Changed blocks of the virtual disk | Tracking reset or disabled forces a full read |
| Hyper-V RCT | Changed disk areas since a reference point | Changed blocks of the virtual disk | Windows Server 2016 or later only; host-local access limits |
| Chunk comparison | Content hashes in the store | Only chunks new to the store | Still reads unchanged files unless a metadata test skips them |
Example (illustrative): fs01 is a Windows file server in a virtual machine. One shared 50 GiB mailbox archive file is edited every day, and about 0.5 GiB of its blocks actually change. Assume nothing else changes. A file-level incremental copies the whole 50 GiB file each night: 50 GiB for the first full copy plus 30 × 50 GiB = 1,550 GiB after 30 nights. A block-level incremental of the virtual disk copies about 0.5 GiB each night, plus a little for file system records: 50 GiB + 30 × 0.5 GiB = 65 GiB. The ratio is 1,550 ÷ 65, about 24 times. If the same server's data were thousands of small files that each change completely, the two methods would store close to the same amount, and file-level may be simpler to restore from.
What does this mean for a restore?
Block-level backups restore a disk, so they suit whole-server recovery. File-level backups restore named files without rebuilding a disk. Many products offer both from one backup by mounting the disk image. Check whether yours does; the difference between restoring an image and restoring files is a separate question with its own trade-offs.
Neither method makes a restore faster by itself. The restore still needs the full backup plus every increment that applies. What block-level tracking changes is the nightly cost, and what a reset costs when the tracking list is lost.
How can you check this on your own system?
Ask your provider or read your product's settings for these specific answers:
- Which method finds changes: file attribute, timestamp, journal, or platform block tracking?
- What does the job do when that source is unavailable or reset: read everything again, or stop with an error?
- Is change information stored per disk (block level) or per file (file level)?
On a Windows file server you can read the journal state. Microsoft documents fsutil usn queryjournal <volumepath>, with fsutil usn queryjournal c: as its example. Microsoft describes it as querying "a volume's USN data to gather information about the current change journal, its records, and its capacity". It queries and does not change anything. Save the output from the backup source and compare it with a later run. If the journal identifier appears in it and has changed, the journal was restamped or re-created.
To see what an archive-bit or date-based tool would pick up without copying anything, robocopy has /l ("files are to be listed only (and not copied, deleted, or time stamped)") and xcopy has /l. Combine with /a for archive-flagged files; Microsoft states that robocopy /a "copies only files for which the Archive attribute is set". Avoid /m, which clears the flag.
Risk: robocopy /m and xcopy /m reset the archive attribute on source files. If another backup tool relies on that flag, it will miss those files from then on. There is no way to restore the flag state except by setting it again with attrib +a, which marks everything as changed.
For VMware, the vSphere API exposes the changeTrackingEnabled property on the virtual machine configuration. Reading it requires access to the vSphere API or a tool built on it. Ask your administrator rather than guessing from the backup job's speed.
Read next
Related articles you may want next.
- TroubleshootingIncremental backups suddenly run as full after a CBT resetInvestigate full reads after CBT changes, distinguish configuration from integrity, and understand supplier-supported recovery and restore acceptance.Read article
- Backup & RecoveryWindows Server Backup and wbadmin: the limits to know before you rely on themUnderstand Windows Server Backup retention, VSS modes, system-state exclusions and bare-metal recovery limits before relying on a backup.Read article
- Backup & RecoveryWhat a Linux file backup should leave out, and how exclusions quietly lose dataWhich Linux directories a file-level backup should exclude, how CACHEDIR.TAG and one-file-system options work, and the exclusions that lose data.Read article
Risks and limits
- These commands have not been run or lab-tested. They are taken from the documentation of each vendor.
- The Microsoft pages show update dates from 2021 to 2025. The Broadcom, restic and Proxmox pages show no date; restic's manual names itself 0.19.1 and the Proxmox pages name 4.2.11-1.
- Microsoft's page on Hyper-V backup approaches is the source for RCT. It does not describe how RCT lists are stored or when they reset.
- GNU tar, Linux ext4 and ZFS also have their own incremental methods that are not covered here.
- The example's numbers are arithmetic on assumed figures. Real data does not change in neat amounts, and block-level lists include false positives.
Sources
- Change Journals — Microsoft Learn, updated 2021-01-07, accessed 17 November 2026
- Change Journal Records — Microsoft Learn, updated 2021-01-07, accessed 17 November 2026
- Using the Change Journal Identifier — Microsoft Learn, updated 2021-01-07, accessed 17 November 2026
- fsutil usn — Microsoft Learn, updated 2024-11-01, accessed 17 November 2026
- attrib — Microsoft Learn, updated 2023-09-25, accessed 17 November 2026
- xcopy — Microsoft Learn, updated 2024-05-28, accessed 17 November 2026
- robocopy — Microsoft Learn, updated 2025-03-17, accessed 17 November 2026
- Hyper-V backup approaches — Microsoft Learn, updated 2025-08-14, accessed 17 November 2026
- VirtualMachine.QueryChangedDiskAreas — Broadcom vSphere Web Services API, no page date, accessed 17 November 2026
- vim.vm.ConfigInfo — Broadcom vSphere Web Services API, no page date, accessed 17 November 2026
- Backing up (change detection) — restic 0.19.1 documentation, no page date, accessed 17 November 2026
- Design: backups and deduplication — restic 0.19.1 documentation, no page date, accessed 17 November 2026
- Technical overview — Proxmox Backup 4.2.11-1 documentation, no page date, accessed 17 November 2026
This article was originally published on .