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

NetLeverage
NetLeverage

How often should you run a full backup when you also run incrementals?

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.

No source gives one right interval, and the tools' own examples differ: Veeam's documentation offers weekly and monthly full schedules, and duplicity's man page gives a full every month as its example. The interval you pick is really a choice about how long a chain of incrementals may grow before a new full starts a fresh one. A shorter interval means shorter restore chains, less dependence on one old file, and more storage and load. A longer one is cheaper until something in the chain fails.

Pick the interval from four numbers: the longest chain you will accept at restore time, what a full costs your source and your storage, how many restore points you must keep, and how often someone reads the chain back to prove it works.

What does a full backup change in an incremental chain?

An incremental backup stores what changed since the previous backup. To restore a point, the software needs the last full backup and every incremental after it, up to that point. A new full starts a new chain.

Veeam says so directly. In its forward incremental method, "by doing regular full backups, the backup chain is split into shorter series. This lowers the chances of losing the backup chain completely and makes this backup method the most reliable", at the price that it "requires more storage space than other methods". Its active and synthetic full backups "reset" the chain: later incrementals use the new full as the starting point.

Duplicity documents the dependency: "When restoring, duplicity applies patches in order, so deleting, for instance, a full backup set may make related incremental backup sets unusable."

Bareos goes a step further and enforces the interval. Its Max Full Interval directive sets "the maximum allowed age (counting from start time) of the most recent successful Full backup that is required in order to run Incremental or Differential backup jobs. If the most recent Full backup is older than this interval, Incremental and Differential backups will be upgraded to Full backups automatically." Bareos also upgrades an incremental to a full when it cannot find a valid previous full for the same job, client and FileSet.

What does a full backup cost?

The cost depends on how the full is made.

Kind of full Where the data comes from What it loads Source
Active full Read from the source again The source system and the network Veeam: "retrieves data for the whole VM from the source"
Synthetic full Built from the previous full and incrementals already in the repository The backup storage; no source read and no network Veeam: "does not retrieve VM data from the source datastore"
Full in a block-level scheme with shared chunks The disk is read in full, but only chunks the server lacks are sent The source read; less upload Proxmox Backup Server documentation
Full forced by a file-level tool A new full backup of the selected files The source and the network Duplicity full action and --full-if-older-than

Veeam's documentation describes a synthetic full as "identical" in data to a regular full, and notes that synthetic fulls on object storage need a GFS policy and that some job chains skip them. A synthetic full is built from the earlier full and incrementals, so it needs that chain to be readable when it runs. Veeam's health check page says it verifies only the latest restore point, so it does not prove that every older file in the chain is readable. A restore test of an older point is the only direct proof.

How do I turn this into an interval?

Work through these questions in order.

  1. What is the longest restore chain I can read in my window? Chain length is roughly the interval in days when you run one incremental a day. Veeam's documentation says forward incremental restore is "the most optimal in time" for points other than the earliest or latest, because the chain is divided into short series of full and incremental files, so assembling the chosen point does not take long. It gives no numbers. Time a restore of your own chain: see the related article on timing restores.
  2. Can my source take a full read? If a full read of the source in the window slows production, a synthetic full or a block-level scheme is the way to keep the interval short.
  3. Can my storage hold the extra fulls? A full cannot be removed while any later incremental you still need depends on it. Veeam notes that forward incremental sometimes "stores more restore points than specified in the retention policy settings". Leave room for that.
  4. Does anything read the chain back? Veeam's health check verifies "only the latest restore point". It does not walk every older point. A full is not a check.

Example (illustrative): fs01 has 2 TiB of data and about 40 GiB of new data each day. Retention is 28 days of daily restore points, with one incremental per day and no compression or deduplication. This is a model to compare intervals, not a prediction for your system. Veeam keeps extra restore points under its own retention rules, so its real peaks can be higher.

Full every Fulls held at peak Incrementals held at peak Space at peak Longest restore chain
7 days 5 29 11,400 GiB (11.1 TiB) full + up to 6 incrementals
14 days 3 38 7,664 GiB (7.5 TiB) full + up to 13 incrementals
28 days 2 53 6,216 GiB (6.1 TiB) full + up to 27 incrementals

The figures follow one rule: a chain is kept until its last point falls outside the 28 days. At a weekly interval that gives 5 x 2,048 + 29 x 40 = 11,400 GiB. Going from a monthly to a weekly full costs 5,184 GiB (about 5 TiB) more at peak, and shortens the worst restore chain from 28 files to 7.

A full read of 2 TiB at an assumed 200 MiB/s takes 2 x 1,024 x 1,024 / 200 = 10,486 s, about 2.9 hours. A day's incremental of 40 GiB takes 40 x 1,024 / 200 = 205 s, about 3.4 minutes. Schedule the full in a quiet window, because it reads about 51 times as much data as a day's incremental (10,486 ÷ 205).

What about schemes that never take a second full?

Some tools avoid repeated fulls by merging old increments. Veeam's forever forward incremental keeps one full file and "injects data of an incremental file into the full backup file before deleting the increment". Veeam says this can fragment the full file, so a "compact" operation is needed, which adds I/O load. Bareos Always Incremental runs incrementals and consolidates the older ones into a new backup; its documentation limits this to file-based backups. These schemes change the question from how often to take a full to how often to consolidate, and how the merged result is checked. Ask the same four questions of them.

How do I check my own setup or my provider's?

  • Bareos: read the job's Level and Max Full Interval directives. A job whose interval is unset or 0 will never be forced to a full by age, so an old full can sit under a very long chain.
  • duplicity: duplicity collection-status <url> "summarize[s] the status of the backup repository by printing the chains and sets found". Read it to see how many incrementals sit on the current full. The options --full-if-older-than <time> and --full-if-n-inc <number> force a full by age or by chain length. duplicity verify <url> <folder> tests the archive against a local copy.
  • Veeam: in the job's schedule settings, check whether synthetic or active fulls are on. The Veeam documentation says that if neither is enabled, the job produces a forever forward incremental chain. Also check whether the health check is scheduled; its default is 05:00 on the last Saturday of each month.
  • A provider: ask how many incrementals sit on one full, what forces a new full, whether it is made from the source or from stored data, and when a restore from the end of a chain was last timed.

Risk: changing a full schedule changes storage use and the timing of the largest load. Moving from monthly to weekly in the model above adds about 5 TiB at peak. To back out, restore the earlier schedule; fulls already written stay until retention removes them.

Disruptive step: commands that remove old sets, such as duplicity's remove-older-than or remove-all-but-n-full, delete backup data and, in duplicity, need --force to act rather than only list. Run them without --force first, and read what they would delete. Deleted backup sets cannot be recovered from the repository.

Risks and limits

  • These commands have not been run or lab-tested. They come from the Veeam, Bareos and duplicity documentation.
  • The storage model ignores compression, deduplication, overlapping changes between days, and each product's retention rules. Real figures can be lower, and for Veeam may be higher where extra points are kept.
  • None of these documents gives a recommended interval, so the table shows the effect of intervals and does not endorse one.
  • A shorter chain lowers dependence on one file, but each full still depends on the storage holding it. Keep a copy elsewhere.
  • Veeam documentation applies to build 13.1.2.57; Bareos to bareos-25; duplicity to 3.2.2. Earlier versions can differ.

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

Sources