* The ``source`` job has a whitelist of client identities that are allowed pull access.
..TIP::
The implementation of the ``sink`` job requires that the connecting client identities be a valid ZFS filesystem name components.
How Replication Works
~~~~~~~~~~~~~~~~~~~~~
One of the major design goals of the replication module is to avoid any duplication of the nontrivial logic.
As such, the code works on abstract senders and receiver **endpoints**, where typically one will be implemented by a local program object and the other is an RPC client instance.
Regardless of push- or pull-style setup, the logic executes on the active side, i.e. in the ``push`` or ``pull`` job.
The following steps take place during replication and can be monitored using the ``zrepl status`` subcommand:
* Plan the replication:
* Compare sender and receiver filesystem snapshots
* Build the **replication plan**
* Per filesystem, compute a diff between sender and receiver snapshots
* Build a list of replication steps
* If possible, use incremental sends (``zfs send -i``)
* Otherwise, use full send of most recent snapshot on sender
* Give up on filesystems that cannot be replicated without data loss
* Retry on errors that are likely temporary (i.e. network failures).
* Give up on filesystems where a permanent error was received over RPC.
* Execute the plan
* Perform replication steps in the following order:
Among all filesystems with pending replication steps, pick the filesystem whose next replication step's snapshot is the oldest.
The idea behind the execution order of replication steps is that if the sender snapshots all filesystems simultaneously at fixed intervals, the receiver will have all filesystems snapshotted at time ``T1`` before the first snapshot at ``T2 = T1 + $interval`` is replicated.
.._replication-cursor-bookmark:
The **replication cursor bookmark**``#zrepl_replication_cursor`` is kept per filesystem on the sending side of a replication setup:
It is a bookmark of the most recent successfully replicated snapshot to the receiving side.
It is is used by the :ref:`not_replicated <prune-keep-not-replicated>` keep rule to identify all snapshots that have not yet been replicated to the receiving side.
Regardless of whether that keep rule is used, the bookmark ensures that replication can always continue incrementally.
Note that there is only one cursor bookmark per filesystem, which prohibits multiple jobs to replicate the same filesystem (:ref:`see below<jobs-multiple-jobs>`).
Placeholders allow the receiving side to mirror the sender's ZFS dataset hierachy without replicating every filesystem at every intermediary dataset path component.
Consider the following example: ``S/H/J`` shall be replicated to ``R/sink/job/S/H/J``, but neither ``S/H`` nor ``S`` shall be replicated.
ZFS requires the existence of ``R/sink/job/S`` and ``R/sink/job/S/H`` in order to receive into ``R/sink/job/S/H/J``.
Thus, zrepl creates the parent filesystems as placeholders on the receiving side.
If at some point ``S/H`` and ``S`` shall be replicated, the receiving side invalidates the placeholder flag automatically.
The ``zrepl test placeholder`` command can be used to check whether a filesystem is a placeholder.
When using multiple jobs across single or multiple machines, the following rules are critical to avoid race conditions & data loss:
1. The sets of ZFS filesystems matched by the ``filesystems`` filter fields must be disjoint across all jobs configured on a machine.
2. The ZFS filesystem subtrees of jobs with ``root_fs`` must be disjoint.
3. Across all zrepl instances on all machines in the replication domain, there must be a 1:1 correspondence between active and passive jobs.
Explanations & exceptions to above rules are detailed below.
If you would like to see improvements to multi-job setups, please `open an issue on GitHub <https://github.com/zrepl/zrepl/issues/new>`_.
No Overlapping
~~~~~~~~~~~~~~
Jobs run independently of each other.
If two jobs match the same filesystem with their ``filesystems`` filter, they will operate on that filesystem independently and potentially in parallel.
For example, if job A prunes snapshots that job B is planning to replicate, the replication will fail because B asssumed the snapshot to still be present.
More subtle race conditions can occur with the :ref:`replication cursor bookmark <replication-cursor-bookmark>`, which currently only exists once per filesystem.
N push jobs to 1 sink
~~~~~~~~~~~~~~~~~~~~~
The :ref:`sink job <job-sink>` namespaces by client identity.
It is thus safe to push to one sink job with different client identities.
If the push jobs have the same client identity, the filesystems matched by the push jobs must be disjoint to avoid races.
N pull jobs from 1 source
~~~~~~~~~~~~~~~~~~~~~~~~~
Multiple pull jobs pulling from the same source have potential for race conditions during pruning:
each pull job prunes the source side independently, causing replication-prune and prune-prune races.
There is currently no way for a pull job to filter which snapshots it should attempt to replicate.
Thus, it is not possibe to just manually assert that the prune rules of all pull jobs are disjoint to avoid replication-prune and prune-prune races.
If you have the need for local replication (most likely between two local storage pools), you can use the :ref:`local transport type <transport-local>` to connect a local push job to a local sink job.
