zrepl/docs/configuration/prune.rst
InsanePrawn 0bbe2befce docs: prune: add prune interval visualisation
fixes #122

Co-Authored-By: Christian Schwarz <me@cschwarz.com>

Signed-off-by: InsanePrawn <insane.prawny@gmail.com>
2020-08-21 22:05:05 +02:00

264 lines
9.5 KiB
ReStructuredText

.. _prune:
Pruning Policies
================
In zrepl, *pruning* means *destroying snapshots*.
Pruning must happen on both sides of a replication or the systems would inevitably run out of disk space at some point.
Typically, the requirements to temporal resolution and maximum retention time differ per side.
For example, when using zrepl to back up a busy database server, you will want high temporal resolution (snapshots every 10 min) for the last 24h in case of administrative disasters, but cannot afford to store them for much longer because you might have high turnover volume in the database.
On the receiving side, you may have more disk space available, or need to comply with other backup retention policies.
zrepl uses a set of **keep rules** per sending and receiving side to determine which snapshots shall be kept per filesystem.
**A snapshot that is not kept by any rule is destroyed.**
The keep rules are **evaluated on the active side** (:ref:`push <job-push>` or :ref:`pull job <job-pull>`) of the replication setup, for both active and passive side, after replication completed or was determined to have failed permanently.
Example Configuration:
::
jobs:
- type: push
name: ...
connect: ...
filesystems: {
"<": true,
"tmp": false
}
snapshotting:
type: periodic
prefix: zrepl_
interval: 10m
pruning:
keep_sender:
- type: not_replicated
# make sure manually created snapshots by the administrator are kept
- type: regex
regex: "^manual_.*"
- type: grid
grid: 1x1h(keep=all) | 24x1h | 14x1d
regex: "^zrepl_.*"
keep_receiver:
- type: grid
grid: 1x1h(keep=all) | 24x1h | 35x1d | 6x30d
regex: "^zrepl_.*"
# manually created snapshots will be kept forever on receiver
- type: regex
regex: "^manual_.*"
.. DANGER::
You might have **existing snapshots** of filesystems affected by pruning which you want to keep, i.e. not be destroyed by zrepl.
Make sure to actually add the necessary ``regex`` keep rules on both sides, like with ``manual`` in the example above.
.. _prune-keep-not-replicated:
Policy ``not_replicated``
-------------------------
::
jobs:
- type: push
pruning:
keep_sender:
- type: not_replicated
...
``not_replicated`` keeps all snapshots that have not been replicated to the receiving side.
It only makes sense to specify this rule on a sender (source or push job).
The state required to evaluate this rule is stored in the :ref:`replication cursor bookmark <replication-cursor-and-last-received-hold>` on the sending side.
.. _prune-keep-retention-grid:
Policy ``grid``
---------------
::
jobs:
- type: pull
pruning:
keep_receiver:
- type: grid
regex: "^zrepl_.*"
grid: 1x1h(keep=all) | 24x1h | 35x1d | 6x30d
│ │ │
└─ 1 repetition of a one-hour interval with keep=all
│ │
└─ 24 repetitions of a one-hour interval with keep=1
└─ 6 repetitions of a 30-day interval with keep=1
...
The retention grid can be thought of as a time-based sieve that thins out snapshots as they get older.
The ``grid`` field specifies a list of adjacent time intervals.
The following procedure happens during pruning:
#. The list of snapshots is filtered by the regular expression in ``regex``.
Only snapshots names that match the regex are considered for this rule, all others are not affected.
#. The filtered list of snapshots is sorted by ``creation``.
#. The left edge of the first interval is aligned to the ``creation`` date of the youngest snapshot.
#. A list of buckets is created, one for each interval.
#. The snapshots are placed into the bucket that matches their ``creation`` date.
#. For each bucket
#. the contained snapshot list is sorted by creation.
#. snapshots from the list, oldest first, are destroyed until the specified ``keep`` count is reached.
#. all remaining snapshots on the list are kept.
The syntax to describe the list of time intervals ("buckets") is as follows:
::
Repeat x Duration (keep=all)
* The **duration** specifies the length of the interval.
* The **keep** count specifies the number of snapshots that fit into the bucket.
It can be either a positive integer or ``all`` (all snapshots are kept).
* The **repeat** count repeats the bucket definition for the specified number of times.
**Example**:
::
This grid spec produces the following list of adjacent buckets. For the sake of simplicity,
we subject all snapshots to the grid pruning policy by settings `regex: .*`.
`
grid: 1x1h(keep=all) | 2x2h | 1x3h
regex: .*
`
0h 1h 2h 3h 4h 5h 6h 7h 8h
| | | | | | | | |
|-Bucket 1-|------Bucket 2-------|-------Bucket 3------|------------Bucket 4------------|
| keep=all | keep=1 | keep=1 | keep=1 |
Let us consider the following set of snapshots @a-zA-C , taken at an interval of ~15min:
| a b c d e f g h i j k l m n o p q r s t u v w x y z A B C |
The `grid` algorithm maps them to their respective buckets:
Bucket 1: a, b, c
Bucket 2: d,e,f,g,h,i,j
Bucket 3: k,l,m,n,o,p,q,r
Bucket 4: q,r,s,t,u,v,w,x,y,z,A,B,C
It then applies the per-bucket pruning logic described above which resulting in the
following list of remaining snapshots.
| a b c d k s |
Note that it only makes sense to grow (not shorten) the interval duration for buckets
further in the past since each bucket acts like a low-pass filter for incoming snapshots
and adding a less-low-pass-filter after a low-pass one has no effect.
.. _prune-keep-last-n:
Policy ``last_n``
-----------------
::
jobs:
- type: push
pruning:
keep_receiver:
- type: last_n
count: 10
...
``last_n`` keeps the last ``count`` snapshots (last = youngest = most recent creation date).
.. _prune-keep-regex:
Policy ``regex``
----------------
::
jobs:
- type: push
pruning:
keep_receiver:
# keep all snapshots with prefix zrepl_ or manual_
- type: regex
regex: "^(zrepl|manual)_.*"
- type: push
snapshotting:
prefix: zrepl_
pruning:
keep_sender:
# keep all snapshots that were not created by zrepl
- type: regex
negate: true
regex: "^zrepl_.*"
``regex`` keeps all snapshots whose names are matched by the regular expression in ``regex``.
Like all other regular expression fields in prune policies, zrepl uses Go's `regexp.Regexp <https://golang.org/pkg/regexp/#Compile>`_ Perl-compatible regular expressions (`Syntax <https://golang.org/pkg/regexp/syntax>`_).
The optional `negate` boolean field inverts the semantics: Use it if you want to keep all snapshots that *do not* match the given regex.
.. _prune-workaround-source-side-pruning:
Source-side snapshot pruning
----------------------------
A :ref:`source jobs<job-source>` takes snapshots on the system it runs on.
The corresponding :ref:`pull job <job-pull>` on the replication target connects to the source job and replicates the snapshots.
Afterwards, the pull job coordinates pruning on both sender (the source job side) and receiver (the pull job side).
There is no built-in way to define and execute pruning on the source side independently of the pull side.
The source job will continue taking snapshots which will not be pruned until the pull side connects.
This means that **extended replication downtime will fill up the source's zpool with snapshots**.
If the above is a conceivable situation for you, consider using :ref:`push mode <job-push>`, where pruning happens on the same side where snapshots are taken.
Workaround using ``snap`` job
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
As a workaround (see GitHub :issue:`102` for development progress), a pruning-only :ref:`snap job <job-snap>` can be defined on the source side:
The snap job is in charge of snapshot creation & destruction, whereas the source job's role is reduced to just serving snapshots.
However, since, jobs are run independently, it is possible that the snap job will prune snapshots that are queued for replication / destruction by the remote pull job that connects to the source job.
Symptoms of such race conditions are spurious replication and destroy errors.
Example configuration:
::
# source side
jobs:
- type: snap
snapshotting:
type: periodic
pruning:
keep:
# source side pruning rules go here
...
- type: source
snapshotting:
type: manual
root_fs: ...
# pull side
jobs:
- type: pull
pruning:
keep_sender:
# let the source-side snap job do the pruning
- type: regex
regex: ".*"
...
keep_receiver:
# feel free to prune on the pull side as desired
...