Australian cloud for business software — built and run in Sydney since 2002

NetLeverage
NetLeverage

Synthetic full vs active full backup: what each reads, what each writes and what each proves

Technical article
Backup & Recovery

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.

An active full backup reads the whole source again and writes a brand-new full copy. A synthetic full backup builds a new full copy inside the backup store from backups already there, so the source is not read again for the data that has not changed. Both start a fresh chain. They differ in who does the work, where the load falls, and how much a new full tells you about the source.

The labels are used by several products. This explanation follows Veeam and Bareos documentation, and the details differ by product.

What does each type read and write?

Active full Synthetic full
Reads from the source All data, as for the first backup Only the changes since the last backup, taken first as a normal increment
Network use Whole dataset (after compression and de-duplication) Only that increment
Reads in the backup store None The previous full and the increments after it
Writes in the backup store A new full file A new full file (or references to existing blocks)
Load falls mainly on Source storage, network, the host doing the backup The backup store
Starts a new chain Yes Yes

Veeam says an active full "retrieves VM data from the source datastore where the VM resides, compresses and deduplicates it and writes it to the VBK file" (Full Backup Methods). A synthetic full "does not retrieve VM data from the source datastore. Instead, it synthesizes a full backup from data you already have in the backup repository", and "the created synthetic full backup file contains the same data you will have if you create an active full backup".

How is a synthetic full built, step by step?

Veeam's documentation shows a detail that the shorter descriptions leave out. On a synthetic full day the job first takes a normal incremental backup, then builds the full (How Synthetic Full Backup Works):

  1. The job runs a regular incremental, "adds a new incremental backup file to the backup chain", and reads that from production storage. This is so the synthetic full "includes the latest changes of the source VM".
  2. The repository's data mover "builds a new synthetic full backup using backup files that are already available in the backup chain, including the newly created incremental backup file".
  3. It then deletes the incremental file made at the start of the session.

So the source is not read again for unchanged data only. The changes are still read that night.

Bareos has the same idea under another name. Its VirtualFull level creates "a new Full backup ... from the last existing Full backup and the matching Differential- and Incremental-Backups", so that it is created "without transferring all the data from the client to the backup server again". It warns that "VirtualFull may require read and write access to multiple volumes" and that source and destination must be on the same Bareos storage daemon (Director configuration, Level).

Why can a synthetic full be quick or slow?

It depends on whether the store copies the data or only points at it. Veeam's Fast Clone "references existing data blocks on volumes instead of copying data blocks between files", and "data blocks are copied only when files are modified". It is used to "create synthetic full backups" on supported repositories, and it "requires that the starting and ending file offsets are aligned to cluster boundaries" (Fast Clone). Veeam's Best Practice Guide lists the I/O on the destination storage as "1 x write I/O" for an active full and "2 x I/O (1x read, 1x write)" for a synthetic full, and marks the synthetic full as benefiting from Fast Clone (Backup job methods). Without such a facility, the store reads and writes data on the scale of the full.

Example (illustrative): fs01 holds 2 TB (2,000,000 MB) and changes 4% a day. The link to backup01 is 1 Gbit/s, and we assume 70% of it is usable: 125 MB/s x 0.7 = 87.5 MB/s.

  • Active full: 2,000,000 MB / 87.5 MB/s = about 22,857 s, or about 6.3 hours of source reads and network.
  • Synthetic full, source side: the 4% increment is 80 GB, and 80,000 MB / 87.5 MB/s = about 914 s, or about 15 minutes.
  • Synthetic full, store side: if the store must write the 2 TB and sustains an assumed 300 MB/s, 2,000,000 MB / 300 MB/s = about 6,667 s, or about 1.9 hours. With block cloning the data is not rewritten, so this figure falls sharply.

The 70% efficiency, the 4% change rate and the 300 MB/s store speed are assumptions for the example, not measurements.

What does a synthetic full not prove?

A new full made from old backups carries forward whatever those backups held. That has two consequences. Veeam does not state the second; it follows from how a synthetic full is built.

  • Unchanged data is never read from the source again. Only an active full tests that the source can still be read from end to end.
  • If a block was damaged in the store, building a new full from it can copy the damage forward. Whether the product notices depends on its checks.

Veeam's own check is a health check that "performs a cyclic redundancy check (CRC) for metadata and a hash check for VM data blocks" but "always verifies only the latest restore point in the backup chain" (Health Check for Backup Files). When it finds corrupted blocks it retries by transporting "data blocks from the source datastore". On Linux immutable repositories, which "do not support repair", it says an active full is needed for that chain.

A fresh active full is therefore a way to re-establish a copy that does not depend on earlier stored blocks, at the price of a full read of production.

When does each one make sense?

Veeam's own wording gives the trade. Active fulls suit a corporate policy that "may require you to create a full backup on weekends and run incremental backup on work days". Synthetic fulls are offered for the case where "periodically running active full backups may not be an option" because they are heavy on the source and the network.

One pattern, which is a design choice rather than a rule, is a weekly synthetic full with an occasional active full to re-read the source. Whether that suits you depends on how long the source read takes against your window, and on how much you trust the stored copy. Work the first out with the window arithmetic; work the second out with a restore test.

What scheduling rules can surprise you?

Veeam documents several that decide whether the full you scheduled actually happens:

  • If synthetic and active fulls fall on the same day, "the synthetic full backup is not created" and only the active full is made.
  • A regular job on the same day as a synthetic full produces "only one backup file", the synthetic full, and no separate increment.
  • For a job chained after another job, if the first job does not run on the full day, no full is made for the chained job.
  • Synthetic fulls "cannot be created independently for backup jobs targeted at object storage"; the GFS policy must be enabled.

(Synthetic Full Backup; Active Full Backup)

Both types leave the previous full on disk "until it is automatically deleted according to the retention policy". The space need therefore peaks when old and new fulls coexist.

How can I check which one my system uses?

  1. Ask the provider which full is made, and when. The answer should name the day and the type, not just "weekly full".
  2. Look for the settings by name. In Veeam they are "Create synthetic full backups periodically on" and the active full schedule in the job's Advanced Settings.
  3. Ask whether the store supports block cloning. For Veeam this is Fast Clone, with "Align backup file data blocks" enabled automatically on supported repositories.
  4. Ask what verifies the stored data. A health check reads the stored blocks, not the source.
  5. Check free space on the day of the full. Both the old and the new full exist until retention clears the old one.

Risks and limits

  • Risk: an active full reads the whole source in one window. On busy production storage this can slow users and applications. There is no way back once it is running except cancelling the job, so schedule it outside business hours.
  • Risk: deleting a full to free space can orphan the increments that depend on it. There is no way back unless another copy exists.
  • The table and example describe Veeam and Bareos only. Other products may order the steps differently.
  • Veeam documents versions 11, 12 and 13 separately; behaviour on older versions can differ.
  • Veeam does not state that damage can be copied forward; it follows from how a synthetic full is built.
  • The example uses decimal units and round figures.

NetLeverage includes a daily full disaster-recovery backup of each hosted server at no extra charge. Talk to an engineer.

Sources